INTERNATIONAL WORKSHOP Patient Safety - MedicalNet-BG
INTERNATIONAL WORKSHOP Patient Safety - MedicalNet-BG
INTERNATIONAL WORKSHOP Patient Safety - MedicalNet-BG
Create successful ePaper yourself
Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.
<strong>INTERNATIONAL</strong> <strong>WORKSHOP</strong><br />
<strong>Patient</strong> <strong>Safety</strong> through Intelligent<br />
Procedures in Medication (PSIP)<br />
Related to the FP7 ICT project PSIP grant 216130<br />
PROCEEDINGS<br />
Edited by<br />
Régis Beuscart, Dimitar Tcharaktchiev and Galia Angelova<br />
Sofia, Bulgaria<br />
23 June 2011
International Workshop<br />
<strong>Patient</strong> <strong>Safety</strong> through Intelligent Procedures in Medication<br />
(related to the FP7 ICT project PSIP, grant 216130)<br />
Sofia, Bulgaria<br />
23 June 2011<br />
PROCEEDINGS<br />
ISBN 978-954-452-015-1<br />
Designed and printed by INCOMA Ltd.<br />
Shoumen, Bulgaria
Organisers and Sponsors<br />
The International Workshop “<strong>Patient</strong> <strong>Safety</strong> through Intelligent<br />
Procedures in Medication” is organised by:<br />
Linguistic Modelling Department,<br />
Institute for Information and Communication Technologies (IICT),<br />
Bulgarian Academy of Sciences (BAS)<br />
and<br />
Medical Informatics Department,<br />
University Specialised Hospital for Active Treatment of Endocrinology (USHATE)<br />
“Acad. I. Penchev”,<br />
Medical University- Sofia, Bulgaria<br />
The International Workshop “<strong>Patient</strong> <strong>Safety</strong> through Intelligent<br />
Procedures in Medication” is partially supported by:<br />
The European Commission via the project PSIP, ICT FP7 grant 216130<br />
Bulgarian Academy of Sciences<br />
Medical University - Sofia, Bulgaria<br />
Foundation Europartners 2000<br />
Association PROREC – Bulgaria
Editors’ Foreword<br />
Adverse Drug Events (ADE) due to product safety problems and medication errors due to human<br />
factors (HF) are a major Public Health issue. They endanger the patients’ safety and instigate<br />
considerable extra hospital costs. The project <strong>Patient</strong> <strong>Safety</strong> through Intelligent Procedures in<br />
Medication (PSIP) develops and validates healthcare ICT applications which help reducing the<br />
risks of preventable ADEs. PSIP facilitates the systematic production of epidemiological<br />
knowledge on ADE and provides healthcare professionals and patients with relevant knowledge<br />
(guidelines, recommendations, etc.). The approach of PSIP is validated in various settings which<br />
ameliorate the entire medication cycle in a hospital environment.<br />
The integrated project PSIP is funded via the European Community’s Seventh Framework<br />
Programme (FP7) under grant agreement 216130 in the period January 2008 – July 2011. In May<br />
2010, two new partners from Bulgaria have joined the PSIP consortium: the University<br />
Specialised Hospital for Active Treatment of Endocrinology (USHATE) “Acad. I. Penchev”,<br />
Medical University - Sofia and the Institute of Information and Communication Technologies<br />
(IICT), Bulgarian Academy of Sciences (BAS). The project extension was focused on one major<br />
direction: validation and assessment of PSIP ADE rules, of suitable subset of the PSIP Clinical<br />
Decision Support System (CDSS) modules and in general, of the PSIP’s contextualisation and<br />
interoperability potential. To support the compilation of a relatively large repository of patientrelated<br />
data in USHATE, IICT-BAS has implemented complex modules for automatic extraction<br />
of diagnoses, drugs and lab data from the free text of hospital patient records in Bulgarian<br />
language. USHATE has integrated a PSIP-compliant repository, has extended its hospital<br />
information system in a suitable manner and has performed the PSIP validation.<br />
This workshop presents mostly the results, performed in the extended PSIP project and focused<br />
on the PSIP validation in USHATE. At the same time the workshop presents all components<br />
from the main PSIP that are needed to understand the tasks performed in the project extension.<br />
Special emphasis is put on the validation exercise and the human factors relevant to the practical<br />
applications of IT solutions in clinical environments.<br />
The PSIP validation in USHATE shows that the Bulgarian medical experts have positive attitude<br />
towards innovative ICT solutions in the everyday clinical practice. The workshop attendees have<br />
also asked numerous questions and made valuable comments, which shows the need of further<br />
developing the eHealth environment in Bulgaria. We hope that the lessons learnt in PSIP will<br />
have long term impact on the IT applications for processing medical data in Bulgaria and that the<br />
linguistic resources in Bulgarian language, which have been collected for the PSIP project, will<br />
be further expanded and reused in the clinical practice.<br />
We would like to thank the authors for contributing to the workshop proceedings and to the<br />
workshop presenters for the interesting lectures delivered at the oral presentations. We are<br />
grateful to the Bulgarian Academy of Sciences for hosting this workshop it its Headquarters<br />
building. Our special gratitude goes to the Foundation Europartners 2000 (Chairman Miroslav<br />
Nikolov) for the support in many technical tasks related to the workshop organisation.<br />
June 2011<br />
Régis Beuscart, Dimitar Tcharaktchiev and Galia Angelova
Workshop Programme<br />
23 June 2011, Headquarters of Bulgarian Academy of Sciences, Sofia, Str. "15 November" 1<br />
9 00 -9 15 Opening<br />
9 15 -9 45 Régis Beuscart (University Hospital Lille)<br />
Overview of the PSIP project for the detection and prevention of Adverse Drug Events<br />
9 45 -10 15 Grégoire Ficheur (University Hospital Lille)<br />
Interoperability of Medical Databases: Construction of Mapping Between Hospitals<br />
Laboratory Results Assisted by Automated Comparison of their Distributions<br />
10 15 45<br />
-10<br />
Nicolas Leroy (University Hospital Lille)<br />
Human Factors in the PSIP project<br />
10 45 -11 15 Coffee Break<br />
Vassilis Kilintzis (Aristotel University of Thessaloniki)<br />
Knowledge Engineering and Development of Clinical Decision Support Systems for<br />
11 15 -11 45 Adverse Drug Event Prevention<br />
11 45 -12 15 Suzanne Pereira (Vidal SA, France)<br />
Comparison of Two Methods for French Drug Names Extraction<br />
12 15 -12 30 Discussion<br />
12 30 -13 30 Lunch Break<br />
13 30 -13 50 Galia Angelova (Institute of Information and Communication Technologies, Bulgarian<br />
Academy of Sciences)<br />
Automatic extraction of patient-related entities from Bulgarian hospital records<br />
13 50 -14 20 Svetla Boytcheva (Institute of Information and Communication Technologies, Bulgarian<br />
Academy of Sciences)<br />
Demo: Automatic extraction of ICD-10 codes for diagnoses and ATC codes and dosages<br />
for drugs<br />
14 20 -14 40 Dimitar Tcharaktchiev (University Specialised Hospital for Active Treatment of<br />
Endocrinology /USHATE/ “Acad. I. Penchev”, Medical University – Sofia)<br />
Development of experimental repository for Bulgarian patients and a testbed for<br />
integration of PSIP CDSS modules with the USHATE hospital information system<br />
14 40 -15 00 Hristo Dimitrov (University Specialised Hospital for Active Treatment of<br />
Endocrinology /USHATE/ “Acad. I. Penchev”, Medical University – Sofia)<br />
Demo: Testbed for integration of CDSS modules and PSIP validation<br />
15 00 -15 30 Coffee Break<br />
15 30 -16 00 Elske Ammenwerth and Martin Jung (University for Health Sciences, Medical<br />
Informatics and Technology, Hall in Tirol, Austria)<br />
Expectations and barriers of physicians versus CPOE (Computerised Physician Order<br />
Entry)<br />
16 00 -16 20 Krassimira Nechkova-Atanassova (University Specialised Hospital for Active<br />
Treatment of Endocrinology /USHATE/ “Acad. I. Penchev”, Medical University – Sofia)<br />
Survey of the Attitudes toward Innovation of the Physicians at USHATE<br />
16 20 -17 00 Round table, discussion<br />
17 00 Closing
Table of Contents<br />
PSIP Overview: Detection and Prevention of Adverse Drug Events<br />
Régis BEUSCART, Emmanuel CHAZARD …………………………………………...……. 1<br />
Interoperability of Medical Databases: Construction of Mapping between Hospitals<br />
Laboratory Results Assisted by Automated Comparison of their Distributions<br />
Grégoire FICHEUR, Emmanuel CHAZARD, Aurélien SCHAFFAR, Régis BEUSCART … 4<br />
Human Factors in the PSIP Project<br />
Nicolas LEROY, Romaric MARCILLY, Marie-Catherine BEUSCART-ZEPHIR …………. 9<br />
Developing Decision Support Services for <strong>Patient</strong> Medication <strong>Safety</strong>:<br />
A Knowledge Engineering Perspective<br />
Vassilis KOUTKIAS, Vassilis KILINTZIS, Régis BEUSCART, Nicos MAGLAVERAS .. 13<br />
Comparison of Two Methods for French Drug Names Extraction<br />
Suzanne PEREIRA, Catherine LETORD, Sophie TESSIER, Solenne REIN,<br />
Stefan DARMONI, Elisabeth SERROT ……………………………………………..…….. 17<br />
Automatic extraction of patient-related entities from Bulgarian hospital records<br />
Galia ANGELOVA ……………………………………………………………………….… 21<br />
Automatic extraction of ICD-10 codes for diagnoses and ATC codes and dosages for drugs<br />
Svetla BOYTCHEVA …………………………………………………………………….… 25<br />
Development of experimental repository for Bulgarian patients and a testbed for<br />
integration of PSIP CDSS modules with the USHATE hospital information system<br />
Dimitar TCHARAKTCHIEV ………….…………………………………………………… 35<br />
Testbed for integration of CDSS modules and PSIP validation in USHATE, Medical University –<br />
Sofia<br />
Hristo DIMITROV ………….…………………………………………………….………… 39<br />
Expectations and Barriers versus cxCDSS-CPOE: A European User Survey<br />
Elske AMMENWERTH, Martin JUNG ..…………………………………………...……… 49<br />
Survey of the attitudes of the physicians at the USHATE hospital towards innovations<br />
Krassimira NECHKOVA-ATANASSOVA ..…..………………………………...………… 53
PSIP Overview: Detection and Prevention of Adverse Drug Events<br />
Régis BEUSCART 1 , Emmanuel CHAZARD 2<br />
1<br />
Lille University Hospital, UDSL-EA2694-France, Regis.BEUSCART@CHRU-LILLE.FR<br />
2<br />
Lille University Hospital, UDSL-EA2694-France, emmanuelchazard@yahoo.fr<br />
Abstract<br />
The development of Clinical Information Systems (CIS),<br />
Laboratory Management Systems (LMS) and Computer<br />
Provider Order Entry Systems (CPOE) for the<br />
computerization of hospital settings make available a large<br />
amount of data, that are now available for statistical<br />
analysis and epidemiological studies. In this paper, we<br />
describe how we utilized the data issued from more than<br />
100,000 medical records for detecting Adverse Drug<br />
Events. First we performed a Knowledge Data Discovery<br />
phase for discovering detection rules. Second, we applied<br />
these rules on real medical data bases to automatically<br />
identify potential ADE cases. Third, we implemented the<br />
rules in Clinical Decision Systems, which are now<br />
available through the Hospital Information System (HIS) or<br />
through the web. The resulting applications are currently<br />
used in routine for the prevention of ADE.<br />
Keywords<br />
Clinical Information Systems, Data Mining, Knowledge<br />
data Discovery, Adverse Drug Events, Hospital<br />
Information Systems, Clinical Decision Support Systems.<br />
1. Introduction<br />
For the last thirty years, ICT (Information and<br />
Communication Technology) applications have been<br />
developed and implemented in hospital settings.<br />
Remarkable efforts have been made to incorporate in<br />
Hospital Information Systems (HIS), specific<br />
applications to support the medication ordering,<br />
dispensing and administration functions. These<br />
applications are usually referred as Computer<br />
Provider Order Entry (CPOE) systems. As a<br />
consequence, huge amounts of data are available in<br />
Hospital Information System (HIS), Lab Management<br />
Systems (LMS) and CPOE. More often, these data are<br />
only used for the management of patients, or for<br />
medico-economic purposes.<br />
These data can also be utilized for a better knowledge<br />
of the hospitalized patients, a better comprehension of<br />
the hospitalization characteristics, the assessment of<br />
the quality of care and epidemiologic studies. In this<br />
paper, we will show how they can be used for <strong>Patient</strong><br />
<strong>Safety</strong>.<br />
<strong>Patient</strong> <strong>Safety</strong> and the identification of Adverse Drug<br />
Events (ADEs) in Healthcare have become a major<br />
public health issue. Every national survey so far<br />
confirms the trend shown in U.S.: in about 10% of the<br />
admissions, serious mistakes occur, and in 60% of the<br />
cases, medications are involved [1]. But Adverse<br />
Drug Events are not systematically declared by<br />
physicians of pharmacists to pharmaco-vigilance<br />
units. It is estimated that only 1-2% of the ADEs<br />
occurring in an hospital are reported. So, most of the<br />
usually cited numbers are estimates and nationwide<br />
extrapolations issued from quantitative or qualitative<br />
regional studies. Therefore, these numbers or<br />
estimates bear a relative uncertainty. Moreover, their<br />
reliability can be questioned.<br />
Comparing, on one hand the difficulty to obtain<br />
reliable epidemiological information on ADE [2], on<br />
the other hand the availability of a large number of<br />
data recorded in the Hospital Information Systems, it<br />
seems possible to identify healthcare situations where<br />
the patient safety is at risk. The objective is, by using<br />
Data Mining techniques, to get a better knowledge of<br />
the prevalence of Adverse Drug Events, and of their<br />
characteristics.<br />
As a consequence, this knowledge can be used for two<br />
objectives: (1) a retrospective analysis of collected<br />
medical records to estimate the prevalence and the<br />
incidence of Adverse Drug Events in different<br />
contexts within hospitals; (2) a prospective prevention<br />
of Adverse Drug Events by using a Clinical Decision<br />
System able to detect the risks associated with<br />
medications to a patient in a defined context.<br />
2. Methodology<br />
Knowledge Data Discovery.<br />
Workshop <strong>Patient</strong> <strong>Safety</strong> through Intelligent Procedures in Medication, Sofia, Bulgaria, 2011<br />
1
The study has been performed in more than 100,000<br />
anonymized medical records obtained after extraction<br />
from HIS according to a common data model.<br />
Data Mining was first performed on a first set of<br />
records (30,000) for Knowledge Data Discovery<br />
(KDD), inducing new rules of knowledge for<br />
detecting ADE. This has been done by means of nonsupervised<br />
and supervised data mining methods.<br />
Supervised methods like Decision trees (CART,<br />
CHAID), and Association Rules have been the more<br />
efficient for KDD.<br />
The knowledge about ADEs issued from the<br />
application of data mining methods has been<br />
summarized by means of ADE detection rules. An<br />
ADE detection rule is made of one or several Boolean<br />
conditions that lead to an outcome, with a given<br />
probability, such as:<br />
Cause1 & Cause2 & Cause3 Outcome<br />
That representation is widely used either for<br />
retrospective ADE detection [3]. In this work, we<br />
obtain a set of 236 rules involving 1 to 4 Causes to<br />
lead to an Outcome. The Causes can be: demographic<br />
characteristics of the patients, drug administrations or<br />
discontinuations, laboratory results, or diagnoses. The<br />
number and the kind of the conditions were not<br />
constrained by the methods but were optimized by the<br />
use of statistical procedures. These rules enable to<br />
identify 56 different types of outcomes organized in 3<br />
categories: Coagulation disorders, Ionic and Renal<br />
disorders, Miscellaneous.<br />
Validation of the Knowledge Base (PSIP KB)<br />
Ruleset<br />
236 rules describing the knowledge for ADE<br />
identification has been collected. A phase of<br />
validation, and refinement has been necessary to make<br />
the rules usable in clinical practice. This validation<br />
has been realized on the set of the remaining collected<br />
70,000 records and on some test-cases provided by<br />
clinicians. Several technical meetings are organized<br />
with experts: physicians, pharmacologists,<br />
pharmacists and statisticians. The rules are examined<br />
by the experts and validated against summaries of<br />
products characteristics and bibliography. That review<br />
used several drug-related web information portals<br />
(Pharmacorama 2009, BDAM 2009, Thériaque 2009),<br />
Pubmed referenced papers (Pubmed 2009), and<br />
French summaries of products characteristics<br />
provided by the Vidal Company.<br />
During this review, the experts may suggest different<br />
things: queries to check what are the pathological<br />
context of the cases and what are precisely the drugs<br />
involved, or tuning of the rules, to test different<br />
classes or subclasses of drugs. Sometimes, rules are<br />
manually enforced in agreement with the academic<br />
knowledge in order to test some hypothesis.<br />
The following part presents two tools implementing<br />
the PSIP rules set for ADE identification and<br />
prevention.<br />
3. Results<br />
ADE Scorecards<br />
A special software “ADE Scorecards”, can semiautomatically<br />
detect potential cases of ADEs in a<br />
large set of medical records has been developed [4].<br />
The calculation and presentation of the statistical<br />
analysis of a set of clinical records to detect potential<br />
ADEs relies on two steps: (1) a computation step,<br />
consisting in applying the ADE detection rules to the<br />
hospital stays in order to detect the ADE cases and to<br />
compute statistics about ADEs (e.g. incidence and<br />
prevalence of the identified Adverse Drug Events per<br />
hospital and per medical unit), (2) the second step, a<br />
web-based display tool, consists in displaying the<br />
statistics and the ADE cases. Thus global<br />
comparisons are also possible between medical units,<br />
hospitals, regions, and even countries.<br />
This application is now routinely used by two<br />
experimental hospitals in a protocol to improve the<br />
patient safety through the prevention or the<br />
monitoring of the potential ADEs.<br />
Contextualized Decision Support System.<br />
The Knowledge Base (PSIP KB) has also been<br />
exploited to build Clinical Decision Systems. The<br />
ADE Detection Rules have been implemented in a<br />
PSIP Knowledge Base System (KBS) to provide<br />
physicians and Healthcare Professionals with<br />
information and alerts concerning the treatment of a<br />
patient, taking into account the geographical<br />
environment, the clinical characteristics and the<br />
current status of the patient (through the lab tests<br />
results), and obviously the current medications. This<br />
PSIP KBS encapsulates signals capable of<br />
2<br />
Workshop <strong>Patient</strong> <strong>Safety</strong> through Intelligent Procedures in Medication, Sofia, Bulgaria, 2011
automatically detecting potential ADEs susceptible of<br />
occurring for a hospitalized patient. It has been<br />
implemented in existing medical application (e.g. a<br />
French Electronic Health Record, a Danish CPOE). A<br />
web prototype (independent of any medical<br />
applications) has been also developed. These CDSS<br />
are available on the PSIP website (http://psip.univlille2.fr/prototypes/public/).<br />
4. Discussion<br />
Many clinical data are now available from Clinical<br />
Information Systems including Electronic Health<br />
records, Lab Management Systems and CPOEs. By<br />
aggregating these data, and analyzing them with<br />
appropriate statistical tools, it is possible to identify<br />
retrospectively the existence, occurrence, and<br />
frequency of various events (nosocomial infections,<br />
Adverse Drug reactions, adverse drug events,<br />
medication errors, falls, surgical performances, etc.).<br />
In our study we performed Data Mining to identify<br />
cases where the common occurrence of certain<br />
clinical features, lab results and drug prescription<br />
make the probability of an Adverse drug Event<br />
higher. The performance of this method, estimated<br />
after an expert review, is about 60%. Moreover, the<br />
experts review shows that the world is complex and<br />
not black/white. In many cases, contraindicated drugs<br />
are prescribed in presence of another drug or despite a<br />
diagnosis. But experts explain that in complex<br />
medical situations, involving multiple diseases, this<br />
medical attitude is comprehensible;<br />
Given the data at hand in hospital databases, our<br />
studies demonstrate that it is possible, with a<br />
reasonable confidence, to identify Adverse Drug<br />
Events, and their frequency.<br />
It becomes also possible to provide physicians and<br />
nurses with alerts, information and recommendations<br />
for a better usage of medications during the<br />
prescription or the administration phases.<br />
By mining large sets of data, it is possible to gain<br />
automatically estimations of events with a better<br />
precision than the current manual methods.<br />
Clinical Information systems represent a vast source<br />
of data and knowledge that only starts to be exploited<br />
but must be more extensively explored through<br />
adequate statistical methods.<br />
The results of the PSIP Project have been presented in<br />
2 international workshops, the minutes of which are<br />
presented in 2 books [5, 6].<br />
6. Acknowledgements<br />
The research leading to these results has received<br />
funding from the European Community’s Seventh<br />
Framework Program (FP7/2007-2013), under Grant<br />
Agreement n° 216130 – the PSIP project.<br />
7. References<br />
[1] Sauer F, <strong>Patient</strong> and medication safety, EJHP-P 11 2005-4.<br />
[2] Murff H. J., Patel V. L., Hripcsak G., Bates D. W. Detecting<br />
adverse events for patient safety research: a review of<br />
current methodologies, J Biomed Inform; 36 (1-2) 2003, pp.<br />
131-43.<br />
[3] Koutkias V. et al. Construting clinical decision systems fro<br />
adverse drug events prevention: a knowledge-based<br />
approach, Proc. AMIA Annu Symp. pp. 63-74, 2010.<br />
[4] Chazard E., Baceanu A., Ferret L, Ficheur G. The ADE<br />
Scorecards: a tool for ADE detection in Electronic health<br />
records, In Koutkias V. et al. [Eds] <strong>Patient</strong> <strong>Safety</strong><br />
Informatics, Amsterdam, IOS Press, 2011.<br />
[5] Beuscart R, Hackl W, Nøhr C, Detection and prevention of<br />
Adverse Drug Events -information technologies and human<br />
factors. Preface, Stud Health Technol, IOS Press, 148, 2009.<br />
[6] Koutkias V., Niès J., Jensen S., Maglaveras N., Beuscart R.,<br />
<strong>Patient</strong> <strong>Safety</strong> informatics, Stud Health Technol, IOS Press,<br />
166, 2011.<br />
5. Conclusion<br />
With the continuous increase in the collection of data<br />
in Clinical Information Systems, it is now possible to<br />
use these data for performing advanced statistical<br />
studies Through Knowledge Data Discovery, new<br />
knowledge or contextualization of existing knowledge<br />
can be obtained for helping the decision process in<br />
medicine.<br />
Workshop <strong>Patient</strong> <strong>Safety</strong> through Intelligent Procedures in Medication, Sofia, Bulgaria, 2011<br />
3
Interoperability of medical databases:<br />
Construction of mapping between hospitals laboratory<br />
results assisted by automated comparison of their<br />
distributions<br />
Grégoire Ficheur, MD 1 , Emmanuel Chazard, MD, PhD 1 , Aurélien Schaffar 1 , Régis Beuscart, MD, PhD 1<br />
1<br />
Department of Medical Information and Archives, CHU Lille,<br />
UDSL EA 2694; Univ Lille Nord de France; F-59000 Lille, France,<br />
gregoire.ficheur@gmail.com<br />
Abstract<br />
In hospital information systems, laboratory results are stored<br />
using specific terminologies which differ between hospitals. The<br />
objective is to create a tool helping to build a mapping between a<br />
target terminology (reference dataset) and a new one. Using a<br />
training sample consisting of correct and incorrect correspondence<br />
between parameters of different hospitals, a match probability<br />
score is built. This model allows also to determine the theoretical<br />
conversion factor between parameters. This score is evaluated on<br />
a test sample of a new hospital: For each reference parameter, best<br />
candidates are returned and sorted in decreasing order using the<br />
probability given by the model. The correct correspondent of 9 of<br />
15 reference parameters are ranked in the top three among more<br />
than 70. All conversion factors are correct. A mapping webtool is<br />
built to present the essential information for best candidates.<br />
Using this tool, an expert has found all the correct pairs.<br />
Keywords<br />
Interoperability, medical databases, data-mining,<br />
laboratory results, statistic distribution<br />
1. Introduction<br />
The construction of inter-hospital databases offers<br />
perspective to analyze larger databases and thus gain<br />
more power in analysis. The European PSIP project<br />
(1), in which this work is included, is a recent<br />
example of construction of such database. The aim of<br />
this project is to detect and prevent the occurrence of<br />
adverse drug reactions by analyzing large datasets<br />
using the methods of data-mining (2).<br />
This kind of analysis requires syntactic<br />
interoperability (ie between databases) and several<br />
standards have emerged in this context including<br />
Health Level 7 (3). Semantic interoperability (ie<br />
between terminologies) is the second aspect of this<br />
convergence and requires the use of common<br />
terminologies. Thus, one can cite the International<br />
Classification of Diseases (4) (ICD) for diagnoses and<br />
also Anatomical Therapeutic Chemical Classification<br />
(5) (ATC) for drugs. However, the encoding of<br />
laboratory results continues to pose a specific problem<br />
in this context and semantic interoperability of<br />
laboratory results is not a reality yet.<br />
The structured description of laboratory results in<br />
databases is based on different terminologies. Logical<br />
Observation Identifiers Names and Codes (LOINC®)<br />
(6-8), SNOMED-CT (9) or International Union of<br />
Pure and Applied Chemistry (IUPAC) (10,11) can be<br />
retained, but mostly a local terminology is used with<br />
low accuracy level (ie less detailed) and only adapted<br />
for clinical use. So a same parameter measured in one<br />
system with one method is named differently between<br />
2 different hospitals and it requires a difficult manual<br />
time-consuming work to build a mapping between all<br />
used terminologies and a unique reference. The<br />
difficulties to map a local to an international<br />
terminology are detailed by Lin and al. (12) analyzing<br />
the correctness of the LOINC mappings of 3 large<br />
institutions.<br />
2. Related Work<br />
Tools can be used to build mappings between these<br />
terminologies, they are based on character string<br />
recognition in the label of parameter or in reports.<br />
“The Map to LOINC” presented by A.N. Khan (13)<br />
uses an automatic mapping tool to map local<br />
laboratory labels to LOINC. Initially, two experts<br />
manually assigned LOINC codes to the tests in a<br />
master merging laboratory tests names and synonyms.<br />
Then an automated mapping tool was developed to<br />
map local laboratory test names to LOINC using the<br />
4<br />
Workshop <strong>Patient</strong> <strong>Safety</strong> through Intelligent Procedures in Medication, Sofia, Bulgaria, 2011
mappings specified in the master file. Evaluation goes<br />
on 4,967 laboratory active tests, 67% have been<br />
mapped automatically, 19% manually and 14% have<br />
been considering as uncodable tests. Local naming<br />
choice was the principal cause of failure (including<br />
variant due to a different facility naming’s<br />
convention).<br />
K.N. Lee and al. (14,15) have built a tool based on the<br />
6 attributes of the LOINC terminology. Each LOINC<br />
code is based on an unique combination of 6 attributes<br />
so each code can be thought as having a unique set of<br />
6 relationships, one to each attribute. All this sets are<br />
stored in a knowledge base with also synonyms and<br />
rules. For a new laboratory result, a set of<br />
relationships is created based on information given by<br />
laboratory (like unit) and a tool compares the set of<br />
relationships of the new entrant to all the sets of<br />
relationships in the knowledge base. In this way,<br />
exact LOINC code can be found. This method needs<br />
manually work to create lists of synonyms and rules.<br />
Two more projects to map local radiology terms to<br />
LOINC (16) and to map clinical narrative to LOINC<br />
(17) can be cited: They don’t concern laboratory<br />
results specifically.<br />
Labels of laboratory results are not explicit in many<br />
cases: For example, for “total bilirubinemia”, the label<br />
“BT” is found in a first database, “Bilirubins” in<br />
another one, and “Bilirubine totale” in a third one.<br />
Most of labels are not explicit enough and, for this<br />
reason, the units, the bounds of normality and the<br />
distribution of parameters are used to find the correct<br />
correspondence during expert review. These steps are<br />
very useful for expert validation, so it might be useful<br />
to automatically compare the distributions of<br />
parameters to develop a new kind of mapping tool.<br />
In this context, Zollo and al. (18) defined each<br />
parameter using “name, frequency, unit, code, cooccurences”<br />
and also “mean and standard deviation”,<br />
these 2 last elements approaching the description of<br />
distribution of the parameter. This work doesn’t seem<br />
to detail how to find automatically the conversion<br />
coefficient factor for a same kind of laboratory result<br />
with different units.<br />
The objective is to create a tool helping to construct a<br />
mapping between a reference parameter (with a target<br />
terminology) and a new parameter from an incoming<br />
data using a different terminology. The target<br />
terminology has to be simple for clinical use and<br />
ready to use for data-mining. Two steps are planned:<br />
1. Construct a model giving a probability of<br />
correspondence of two parameters. The<br />
approach is not to use the labels but rather to<br />
compare the statistical distributions of<br />
parameters. This model must also identify the<br />
unit.<br />
2. Present the results and the discriminant<br />
information via a web tool that can help an<br />
expert to build this mapping.<br />
3. Material<br />
Here are presented the datasets used in this project.<br />
Table 1 contains more details about these datasets.<br />
The reference dataset with reference distributions of<br />
laboratory results is taken from Denain’s (FR)<br />
hospital center. From this reference data, a simple<br />
target terminology is manually created. This target<br />
terminology could be replaced by an existing<br />
international terminology. Among the 233 found<br />
parameters, 15 most used parameters are selected for<br />
this study, these 15 parameters are the target to match<br />
with the other datasets.<br />
Hospital<br />
Center<br />
Denain<br />
(FR)<br />
Rouen<br />
(FR)<br />
Copenhagen 1<br />
(DK)<br />
Copenhagen 2<br />
(DK)<br />
Use<br />
Name<br />
Reference set<br />
Target<br />
Learning set<br />
Newcomer 1a<br />
Learning set<br />
Newcomer 1b<br />
Validation set<br />
Newcomer 2<br />
Selected<br />
parameters<br />
15<br />
(to match)<br />
Terminology<br />
Target<br />
terminology<br />
74 Local<br />
terminology<br />
71 IUPAC<br />
74 IUPAC<br />
Table 1 - Details about datasets used in the project<br />
The two datasets used in addition to the dataset of<br />
reference for learning phase contain different<br />
terminologies. Only the parameters having a number<br />
of values above 1‰ of the total number of laboratory<br />
results are retained in the datasets of Rouen hospital<br />
center and Copenhagen hospital 1. This limit is<br />
chosen in order to have a good representation of the<br />
distribution and to remove rare parameters. The Table<br />
2 illustrates the terminologies used to represent the<br />
same laboratory results (sodium ion and glycemia)<br />
among 3 different datasets.<br />
Workshop <strong>Patient</strong> <strong>Safety</strong> through Intelligent Procedures in Medication, Sofia, Bulgaria, 2011<br />
5
Hospital<br />
center<br />
Label of<br />
parameter<br />
Kind<br />
Unit<br />
Denain NA1 Sodium ion meq/l<br />
Denain GLY1 Glycemia g/l<br />
Copenhagen<br />
1 & 2<br />
Copenhagen<br />
1 & 2<br />
NPU03429 Sodium ion mmol/l<br />
DNK35842 Glycemia mmol/l<br />
Rouen Sodium Sodium ion mmol/l<br />
Rouen Glycémie Glycemia mmol/l<br />
Table 2 - Same laboratory results in different datasets<br />
The evaluation is made with a dataset from a second<br />
hospital of Copenhagen (DK). These results are<br />
encoded according to IUPAC classification, but this<br />
setting is not used (of course) to make the<br />
correspondence between pairs of parameters, it is used<br />
only to check the quality of the found correspondence.<br />
4. Method<br />
The comparison of distributions is strongly<br />
intertwined with the finding of theoretical conversion<br />
coefficient between two parameters.<br />
For each reference parameters, theoretical conversion<br />
coefficients possible are known. For example,<br />
international unit for creatininium is “μmol/l”, but a<br />
lot of hospitals use another unit “mg/l”. The<br />
theoretical conversion coefficients are “1” if the two<br />
parameters have the same unit, “8.85” (from μmol/l to<br />
mg/l) and “1/8.85” (from mg/l to μmol/l). Figure 1<br />
illustrates this point: On the left is presented<br />
creatinemia’s distribution from two different data with<br />
two different units and on the right creatinemia from<br />
the second dataset is multiplied by the nearest<br />
theoretical conversion coefficient and distributions<br />
overlap. For the parameters counting a number of<br />
elements (for example the number of red blood cells),<br />
all the powers of 10 are possible coefficients. It is the<br />
same with some concentrations as hemoglobin that<br />
may be in g/l or g/dl.<br />
Figure 1 - Effect on distribution of conversion coefficient<br />
To determine the nearest theoretical conversion<br />
coefficient, all theoretical coefficients (known for a<br />
reference) are compared with the ratio of medians<br />
between the reference and the candidate. This nearest<br />
theoretical conversion coefficient is used to multiply<br />
the values of a new entrant parameter and thus the two<br />
parameters have the same scale. Next variables<br />
describing the distribution are evaluated after this<br />
conversion.<br />
5 robust and discriminating variables are retained to<br />
describe the distribution of a laboratory result:<br />
Kolmogorov test’s statistic<br />
10th, 65th and 85th quantiles ratio of the 2<br />
compared parameters<br />
R squared of the linear regression (with<br />
quantile new variable as dependent variable<br />
and quantile reference variable as<br />
independent variable).<br />
The 5 variables presented above are the explanatory<br />
variables. In the table containing training data, each<br />
line corresponds to the values obtained for a pair of<br />
parameters (target vs. newcomer 1a & 1b). This table<br />
also contains a column for the binary dependent<br />
variable “couple” that indicates whether the<br />
correspondence of the couple (seen on the line) is true<br />
or false. The variable “couple” is the variable to<br />
explain. A logistic regression is performed with these<br />
variables and a model is obtained. The model thus<br />
built directly returns a score describing the probability<br />
of match for a pair of tested parameters.<br />
Validity and robustness of the resulting model are<br />
studied. All analyses are performed with R software<br />
version 2.11.1 (19).<br />
This score is then evaluated on a test sample of a new<br />
hospital (newcomer 2). For each reference parameter,<br />
best candidates are returned and sorted in decreasing<br />
order using the probability given by the logistic<br />
model. All results are saved using a XML format.<br />
6<br />
Workshop <strong>Patient</strong> <strong>Safety</strong> through Intelligent Procedures in Medication, Sofia, Bulgaria, 2011
A mapping web tool (using directly XML results<br />
files) is built to present the essential information of<br />
best candidates. This tool is designed to help an<br />
expert to take a decision.<br />
The web interface shows, for each reference<br />
parameter:<br />
A classification of best corresponding candidates<br />
sorted in decreasing order using the probability given<br />
by the logistic regression. Only the corresponding<br />
candidates with a probability above a limit score are<br />
retained.<br />
Their graphic distribution: Used Graphics combine<br />
the graphic distribution of the reference parameter and<br />
the graphic distribution of the new parameter<br />
multiplied by the nearest theoretical conversion<br />
coefficient.<br />
Their names and a few describing attributes<br />
aggregating the main information required to validate<br />
the visual impression: Unit, number of values, nearest<br />
theoretical conversion coefficient and bounds of<br />
normality.<br />
5. Results<br />
For each reference, all correct parameters are returned<br />
among the short list candidates and the median<br />
number of candidates is 6. The main result is the rank<br />
given by the model to the correct couple (a target<br />
parameter and its correspondent from newcomer 2).<br />
For the correct candidate, the first rank is obtained for<br />
7 of them and 9 are in the top three. The worst rank is<br />
8, obtained for 1 couple. All conversion factors found<br />
for these parameters are correct. These results are<br />
presented in Table 3.<br />
6. Conclusion<br />
A logistic model based on the comparison of<br />
distributions of laboratory results is built. It is then<br />
used to select a short list of compatible parameters for<br />
each of the target parameters. All relevant parameters<br />
are among the top 8 candidates and all the preselected<br />
conversion factors are correct. A short list of<br />
candidates is then presented via a web tool allowing<br />
to an independent expert to find all the correct pairs in<br />
a short time.<br />
Reference data compared<br />
to newcomer 2<br />
Results given by the model for<br />
the correct candidate<br />
Parameter Rank (/74) Conversion<br />
coefficient<br />
Sodium Ion 1 1<br />
Kalemia 1 1<br />
Creatininium 1 0.113<br />
Haemoglobin 1 1.667<br />
Leukocytes 1 1000<br />
C Reactive Protein 1 1<br />
Carbamide 1 16.6<br />
Calcium Ion 2 40<br />
Creatin Kinase 2 1<br />
Aspartate transaminase 2 1<br />
Glycemia 4 0.18<br />
Erythrocytes 4 1<br />
Alanine transaminase 4 1<br />
International Normalized<br />
Ratio<br />
5 1<br />
Bilirubins 8 0.585<br />
Table 3 - Rank and unit returned by the model for the correct<br />
correspondent parameters<br />
7. Discussion<br />
The number of reference parameters is reduced and<br />
related only to common parameters so well described<br />
in the dataset. For this reason, only the parameters<br />
representing more than 1‰ of the lines of the<br />
database are analyzed. It is possible that this model<br />
has lower performance on rare parameters like<br />
troponin or thyroid hormones.<br />
Some identical parameters (with same unit) have<br />
sometimes very different distribution and are thus<br />
difficult to compare with this method. For example,<br />
INR is a systematic rendering of one of our partner<br />
hospitals (when a prothrombin time is done) while in<br />
other places, it is reserved for patients with vitamin K<br />
antagonist. Thus, the distributions are very different<br />
between these places. Another example concerns the<br />
C reactive protein which shows very different<br />
distributions between hospitals: A first hypothesis is<br />
that the practices are different; another hypothesis is<br />
that it comes from the measurement tool.<br />
Finally, this method doesn’t take into account the<br />
textual results like some results of bacteriology.<br />
Workshop <strong>Patient</strong> <strong>Safety</strong> through Intelligent Procedures in Medication, Sofia, Bulgaria, 2011<br />
7
This analysis has focused on the comparison of<br />
distributions, but doesn’t try to model the distribution<br />
of each parameter individually. This complementary<br />
approach is perhaps a way to progress.<br />
Moreover, it seems possible to use a meta-rule that<br />
would organize the allocation of the same parameter<br />
between many references. For example, the parameter<br />
A is ranked first among candidates for a reference 1<br />
and only the fifth candidate for reference 2 so it seems<br />
reasonable to exclude the parameter A from the list of<br />
candidates for reference 2 and to keep it only in the<br />
list of candidates for reference 1.<br />
The fact to report on a web tool the short list of<br />
candidates (given by the model for one reference<br />
parameter) gives a real help to map a new<br />
terminology to a target terminology, the final decision<br />
being left to an expert. One of the interests of the<br />
mapping tool is to be used with any local or<br />
international terminology.<br />
8. Acknowledgements<br />
The research tasks leading to these results have received<br />
funding from the European Community’s Seventh<br />
Framework Programme (FP7/2007-2013) under grant<br />
agreement n° 216130 PSIP (<strong>Patient</strong> <strong>Safety</strong> through<br />
Intelligent Procedures in Medication).<br />
9. References<br />
[1] www.psip-project.eu [Internet]. [cit. 2011 Feb<br />
23];Available from: http://www.psip-project.eu<br />
[2] Chazard E, Preda C, Merlin B, Ficheur G,<br />
Beuscart R. Data-mining-based detection of<br />
adverse drug events. Stud Health Technol Inform.<br />
2009;150:552-6.<br />
[3] HL7. Health Level 7 [Internet]. Available from:<br />
http://www.hl7.org<br />
[4] WHO. International Classification of Diseases<br />
[Internet]. 2009;Available from:<br />
http://www.who.int/classifications/icd/en<br />
[5] WHO. Anatomical and Therapeutical<br />
Classification. 2009;Available from:<br />
http://www.whocc.no/atcddd<br />
[6] Logical Observation Identifiers Names and Codes<br />
(LOINC®) — LOINC [Internet]. [cit. 2011 Mar<br />
17];Available from: http://loinc.org/<br />
[7] Forrey AW, McDonald CJ, DeMoor G, Huff SM,<br />
Leavelle D, Leland D, et al. Logical observation<br />
identifier names and codes (LOINC) database: a<br />
public use set of codes and names for electronic<br />
reporting of clinical laboratory test results. Clin.<br />
Chem. 1996 Jan;42(1):81-90.<br />
[8] McDonald CJ, Huff SM, Suico JG, Hill G,<br />
Leavelle D, Aller R, et al. LOINC, a universal<br />
standard for identifying laboratory observations: a<br />
5-year update. Clin. Chem. 2003 Avr;49(4):624-<br />
633.<br />
[9] IHTSDO. SNOMED-CT [Internet].<br />
2009;Available from: http://www.ihtsdo.org<br />
[10] Pontet F, Magdal Petersen U, Fuentes-Arderiu X,<br />
Nordin G, Bruunshuus I, Ihalainen J, et al.<br />
Clinical laboratory sciences data transmission: the<br />
NPU coding system. Stud Health Technol Inform.<br />
2009;150:265-269.<br />
[11] IUPAC. C-NPU. 2009;Available from:<br />
http://www.iupac.org<br />
[12] Lin MC, Vreeman DJ, McDonald CJ, Huff SM. A<br />
Characterization of Local LOINC Mapping for<br />
Laboratory Tests in Three Large Institutions.<br />
Methods Inf Med [Internet]. 2010 Aoû 20 [cit.<br />
2011 Feb 1];49(5). Available from:<br />
http://www.ncbi.nlm.nih.gov/pubmed/20725694<br />
[13] Khan AN, Griffith SP, Moore C, Russell D,<br />
Rosario AC, Bertolli J. Standardizing laboratory<br />
data by mapping to LOINC. J Am Med Inform<br />
Assoc. 2006 Juin;13(3):353-355.<br />
[14] Lee KN, Yoon J, Min WK, Lim HS, Song J, Chae<br />
SL, et al. Standardization of terminology in<br />
laboratory medicine II. J. Korean Med. Sci. 2008<br />
Aoû;23(4):711-713.<br />
[15] Lau LM, Banning PD, Monson K, Knight E,<br />
Wilson PS, Shakib SC. Mapping Department of<br />
Defense laboratory results to Logical Observation<br />
Identifiers Names and Codes (LOINC). AMIA<br />
Annu Symp Proc. 2005;:430-434.<br />
[16] Vreeman DJ, McDonald CJ. A comparison of<br />
Intelligent Mapper and document similarity scores<br />
for mapping local radiology terms to LOINC.<br />
AMIA Annu Symp Proc. 2006;:809-813.<br />
[17] Fiszman M, Shin D, Sneiderman CA, Jin H,<br />
Rindflesch TC. A Knowledge Intensive Approach<br />
to Mapping Clinical Narrative to LOINC. AMIA<br />
Annu Symp Proc. 2010;2010:227-231.<br />
[18] Zollo KA, Huff SM. Automated mapping of<br />
observation codes using extensional definitions. J<br />
Am Med Inform Assoc. 2000 Déc;7(6):586-592.<br />
[19] R_Development_Core_Team. R: A Language and<br />
Environment for Statistical Computing [Internet].<br />
2009;Available from: http://www.R-project.org<br />
8<br />
Workshop <strong>Patient</strong> <strong>Safety</strong> through Intelligent Procedures in Medication, Sofia, Bulgaria, 2011
Human Factors in the PSIP project<br />
Nicolas LEROY 1 , Romaric MARCILLY 2 , Marie-Catherine BEUSCART-ZEPHIR 3<br />
1 2 3<br />
INSERM CIC-IT, Lille ; Univ Lille Nord de France ; CHU Lille ; UDSL EA 2694 ; F-59000 Lille, France<br />
1 nicolas.leroy-2@univ-lille2.fr 2 romaric.marcilly@univ-lille2.fr 3 mcbeuscart@univ-lille2.fr<br />
Abstract<br />
The goal of the ISO 13407 Human Centered Design<br />
Process is to ensure that the development and use of<br />
interactive system take into account the user’s needs.<br />
We applied this approach to the PSIP project and<br />
analyzed the user’s tasks connected to the medication<br />
use process. As a result some recommendations for<br />
the design of the system were produced. An iterative<br />
evaluation process of the usability of the PSIP tools<br />
was performed to ensure that the systems are easy to<br />
use and meet the user needs.<br />
Keywords<br />
<strong>Patient</strong> safety, Usability, User Centered Design<br />
Process, Adverse Drug Event, Human Factor<br />
1. Introduction<br />
The first objective of the PSIP project is to improve<br />
the safety of the medication use process. Besides, the<br />
efficiency of the organization and the well-being and<br />
satisfaction of users are also addressed. These two<br />
aspects are linked to the acceptability of the system by<br />
the organization and the final users.<br />
To reach these goals, we have some specific Human<br />
Factors objectives: the satisfaction of user’s needs, a<br />
good usability of the tools, and the consistency of<br />
organizations, policies and procedures. For the<br />
satisfaction of these objectives, specific human factors<br />
methodologies were used in a user-centered design<br />
process (figure 1).<br />
Figure 1 : Human Factors objectives and methodology<br />
2. Methods<br />
2.1 User-centered design process<br />
The ISO 13407 user centered design process is a<br />
multidisciplinary process divided into four parts<br />
(figure 2).<br />
The first part refers to the specification of the context<br />
of use. At this stage, the main objectives are to<br />
understand the real activity of the users, the second<br />
part refers to the specification of the user’s needs and<br />
the third part refers to recommendations for the design<br />
of the products and the last part refers to the<br />
evaluation of the design. This is an iterative process:<br />
if the results of the evaluation are not good enough,<br />
another loop of specifications and evaluations is<br />
performed.<br />
Workshop <strong>Patient</strong> <strong>Safety</strong> through Intelligent Procedures in Medication, Sofia, Bulgaria, 2011<br />
9
Figure 2 : ISO 13407 User Centered Design Process<br />
2.2 Analysis of the context of use<br />
The analysis of the work situations in the medication<br />
use process is mandatory to support the user-centred<br />
design of the system and therefore enhance the<br />
acceptance of this system and ultimately its<br />
efficiency. The methods used for this work analysis<br />
must allow the identification and description of the<br />
relationships between (i) the obligations inherent to<br />
the professionals’ tasks and activities, (ii) the<br />
constraints imposed by the work environment and (iii)<br />
the resources that can be mobilized by the<br />
professionals. A number of well-known techniques<br />
are available to help the analyst collect, organize,<br />
formalize and analyze relevant data. To structure the<br />
techniques and methods in the PSIP project, we rely<br />
on the general framework from Jens Rasmussen & al.,<br />
the Cognitive Work Analysis (CWA) [1], which is the<br />
internationally recognized reference for the study of<br />
complex sociotechnical environments.<br />
The interview is one of the most commonly used<br />
approaches for information gathering [2]. It has to be<br />
completed with observational techniques such as<br />
direct visual observation, video-recording, participant<br />
observation, time-lapsed photography, etc. The<br />
triangulation of these techniques allows uncovering<br />
information on the work situation which cannot be<br />
acquired in any other way [3].<br />
Figure 3: illustration of the analysis of the work situation<br />
(pharmacist assistant preparing the drug pills)<br />
Figure 4: illustration of the analysis of the work situation<br />
(Physicians during the design sessions)<br />
Three observers trained in naturalistic observation in<br />
complex settings conducted observations and<br />
interviews. To generalize the results to different<br />
cultures, countries and modes of functioning, an<br />
exploratory phase was undertaken to gather general<br />
information on the work situations and their<br />
organization. It was performed in Danish and French<br />
(CHU de Lille and CH of Denain) hospitals.<br />
Observations were supported by video recording<br />
and/or handwritten time-stamped, detailed field notes.<br />
Preliminary interviews were carried out with the<br />
target professional to gather general information about<br />
the work situation (global organization, technical<br />
system, habits of work, etc.). In the Denmark<br />
Hospital, the observations were supported by videorecording.<br />
In the French hospitals, as the HF team has<br />
been working on a daily basis in the CHU de Lille and<br />
involved in a great number of IT projects for over 7<br />
years, always using the same methods, a number of<br />
10<br />
Workshop <strong>Patient</strong> <strong>Safety</strong> through Intelligent Procedures in Medication, Sofia, Bulgaria, 2011
observation notes from other observations were added<br />
to the main collection of data.<br />
Then, the observations and interviews were<br />
systematized according to the determinants of the<br />
work situations [4].<br />
2.3 Specifications and design requirements<br />
This step of the requirements determination process<br />
allowed obtaining a set of common supports which<br />
could be exchanged with the different stakeholders of<br />
the PSIP project (users, analysts/designers, managers,<br />
ergonomists). Several techniques have been used to<br />
describe and structure collected data:<br />
1. Some techniques issued from human factors<br />
domain as quantitative data, SHEL model, and<br />
taxonomy.<br />
2. Some techniques issued from software<br />
engineering and Human Computer Interaction<br />
domain as UML models<br />
3. Textual description and mock-ups<br />
2.4 Evaluation of the design<br />
Methods for usability evaluation and re-engineering<br />
have been applied to six prototypes. Given the wide<br />
variety of prototypes it has not been adequate to apply<br />
the same evaluation method to all the prototypes.<br />
Different methods from the partner’s main toolboxes<br />
have been careful selected and adapted to each<br />
particular prototype to meet the unique demand of<br />
each situation. The usability evaluations have been<br />
performed in different locations and situations ranging<br />
from simple laboratory tests involving no real users to<br />
full field studies were the prototype is used in routine<br />
clinical work [5]. An overview of the relevant method<br />
for each test approach and the type of users involved<br />
are shown in Table 1.<br />
Laboratory<br />
Field<br />
Usability<br />
inspection<br />
In lab<br />
No users<br />
Focus<br />
groups<br />
and<br />
interviews<br />
End users<br />
Usability<br />
test<br />
In lab<br />
Simulatio<br />
n (tasks)<br />
End users<br />
Full<br />
scale<br />
simulati<br />
on<br />
Simulat<br />
ed<br />
environ<br />
ment<br />
End<br />
users<br />
Field<br />
Clinical study<br />
(on site<br />
usability<br />
study)<br />
Clinical field<br />
End users<br />
Real patients<br />
cases<br />
Table 1: Usability evaluation methods used in PSIP<br />
3. Results<br />
3.1 Description of work situation<br />
The descriptions of the work situations allowed<br />
identifying a total of twelve noticeable “moments” of<br />
the work situation. Each “moment” described refers to<br />
either the medication use process or the gathering of<br />
information about the patient’s state or to<br />
communications between the professionals:<br />
1. Ward round<br />
2. Drugs’ prescription<br />
3. Control of new prescriptions by the pharmacist<br />
4. Infusions’ preparation<br />
5. Medication organizer’s preparation<br />
6. Infusions’ administration<br />
7. Drugs’ administration rounds<br />
8. Laboratory ordering and reporting cycle<br />
9. Physiological parameters’ measurement<br />
10. Nurses’ shift handover<br />
11. Briefing physician / nurse<br />
12. Transfer of a patient<br />
The actual safety procedures currently performed by<br />
the Health care professionals for the prevention of<br />
ADEs were also analyzed. At a macroscopic level,<br />
four main types of safety procedures are<br />
distinguished:<br />
1. The control of the decision of the physician and of<br />
its implementation: is the medical decision<br />
adapted to the patient’s case and how, through the<br />
prescription, will it be carried out?<br />
2. The control of the validity of the order according<br />
to the evolution of the patient’s process: is the<br />
prescription still adapted to the patient’s condition<br />
in spite of the patient’s state evolutions?<br />
3. The control of the correspondence between what<br />
the patient is getting and what he should get: is<br />
what the patient is getting similar to what has<br />
been prescribed for him?<br />
4. The compliance of the patient about his<br />
medications’ taking and its control: is the patient<br />
taking what he is supposed to take?<br />
3.2 Specifications and requirements<br />
The initial prescribing step is of critical importance:<br />
The system should be a clinician partner and support<br />
the decision making of the physician. Moreover, the<br />
system should not interrupt the physician during the<br />
Workshop <strong>Patient</strong> <strong>Safety</strong> through Intelligent Procedures in Medication, Sofia, Bulgaria, 2011<br />
11
prescription and the content and display of the alert<br />
must be adapted to the clinical context [6].<br />
For example, if a rule with a risk of hyponatremia is<br />
activated, we may have different types of clinical<br />
contexts.<br />
Example 1: If the physician prescribed a monitoring<br />
of the natremia, and the current results of the natremia<br />
are good, the context is “Risk of hyponatremia, but<br />
for the moment it’s OK”<br />
Example 2: if the physician didn’t prescribe the<br />
monitoring of the natremia, the patient may be in<br />
danger. The context is “Risk of hyponatremia, and the<br />
effect is not monitored, BE CARREFUL”<br />
Figure 5: example of clinical context for a PSIP alert<br />
The activity is also distributed for the control of the<br />
medication process. The system will have to support<br />
the existing collective continuous control and<br />
monitoring of the drug effects. As a consequence, the<br />
system should also be a team player and support the<br />
cooperation between the physician, nurses,<br />
pharmacists and patients. The system should also<br />
raise the collective awareness on ADEs occurring in<br />
the clinical setting.<br />
3.3 Evaluation of the design<br />
As the usability evaluations has been performed on all<br />
the six prototypes in the PSIP project almost the entire<br />
PSIP team (more than 45 persons) has been involved<br />
in various degrees in this process.<br />
Example of the web prototype: It is a web based tool<br />
that provides ADE information independently of the<br />
availability of CPOE systems. Through this<br />
application, healthcare professionals as physicians and<br />
pharmacists can display patient’s hospitalization data<br />
and check the medication prescriptions to verify<br />
whether they triggers a PSIP alert.<br />
The evaluation started with a usability inspection<br />
carried out with a heuristic evaluation method. The<br />
list of problems identified and rated for their severity<br />
was then presented to the designers / developers of the<br />
prototype, and several meetings allowed finding<br />
technical solutions to fix most of the problems. At the<br />
end of the re-engineering, the ergonomists<br />
documented the number of problems solved and the<br />
list of remaining problems (if any). Once the<br />
prototype considered usable enough (all major<br />
problems fixed) the re-engineering phase was<br />
validated, and usability tests were carried out with<br />
end-users.<br />
4. Discussion<br />
We applied a user centred design lifecycle to the PSIP<br />
project. The focus was on the analysis of the current<br />
work system and the design of the expected situation<br />
The expected HF benefits for the PSIP project are: (1)<br />
tools adapted to the user’s needs (2) a good<br />
acceptability by the final users (3) products easy to<br />
learn and to understand (4) The knowledge of PSIP<br />
really use by the clinicians.<br />
5. Acknowledgements<br />
The research tasks leading to these results have<br />
received funding from the European Community’s<br />
Seventh Framework Programme (FP7/2007-2013)<br />
under grant agreement n° 216130 PSIP (<strong>Patient</strong> <strong>Safety</strong><br />
through Intelligent Procedures in Medication).<br />
6. References<br />
[1] Rasmussen J, Pejtersen AM, Goodstein LP. Cognitive<br />
systems engineering. New York: Wiley; 1994.<br />
[2] Meister D. Behavioural Analysis and Measurement<br />
Methods. New York: Wiley and Sons; 1985.<br />
[3] Kirwan B, Ainsworth LK. A guide to task analysis.<br />
Philadelphia: Taylor & Francis; 2001.<br />
[4] Riccioli C. Leroy N. Pelayo S., The PSIP approach to<br />
account for human factors in adverse drug events:<br />
preliminary field studies, Stud Health Technol Inform<br />
148, 197–205, 2009.<br />
[5] Kanstrup AM. Christiansen MB. Nøhr C. Four<br />
principles for user interface design of computerised<br />
clinical decision support systems. Stud Health Technol<br />
Inform. 166:65-73, 2011.<br />
[6] Marcilly R. Leroy N. Luyckx M. Pelayo S. Riccioli C.<br />
Beuscart-Zéphir MC. Medication Related<br />
Computerized Decision Support System (CDSS):<br />
Make it a Clinicians' Partner! Stud Health Technol<br />
Inform. 166:84-94, 2011.<br />
12<br />
Workshop <strong>Patient</strong> <strong>Safety</strong> through Intelligent Procedures in Medication, Sofia, Bulgaria, 2011
Developing Decision Support Services for <strong>Patient</strong> Medication <strong>Safety</strong>:<br />
A Knowledge Engineering Perspective<br />
Vassilis KOUTKIAS 1 , Vassilis KILINTZIS 1 , Régis BEUSCART 2 , Nicos MAGLAVERAS 1<br />
1<br />
Lab of Medical Informatics, Medical School, Aristotle University, Thessaloniki, GREECE,<br />
email: {bikout, billyk, nicmag}@med.auth.gr<br />
2<br />
Lille University Hospital, EA2694, FRANCE, email: regis.beuscart@univ-lille2.fr<br />
Abstract<br />
Adverse drug events (ADEs) constitute a major public<br />
health issue, endangering patients’ safety and<br />
introducing significant additional healthcare costs.<br />
The EU-funded project PSIP (<strong>Patient</strong> <strong>Safety</strong> through<br />
Intelligent Procedures in medication) aims to develop<br />
medication-related Clinical Decision Support Systems<br />
(CDSS) for ADE prevention. In order to identify the<br />
origin of preventable ADEs, PSIP employs novel data<br />
and semantic mining techniques as well as human<br />
factor analyses, all applied on patient records and<br />
actual hospital settings available from several clinical<br />
environments across Europe. This new knowledge<br />
combined with existing evidence, i.e. drug<br />
interactions and already identified ADE cases<br />
reported in the literature, is used to populate a<br />
Knowledge Base (KB) that constitutes the basis for<br />
developing contextualized CDSS modules for ADE<br />
prevention at the point of care. In this paper, we focus<br />
on the knowledge engineering approach employed<br />
towards the construction of a flexible Knowledgebased<br />
System (KBS) that is able to encompass<br />
updates of the source knowledge and support the<br />
deployment of medication-related CDSS modules.<br />
Keywords<br />
Adverse Drug Event (ADE), ADE Prevention, Clinical<br />
Decision Support System (CDSS), Medication-related<br />
CDSS, Knowledge Engineering, Knowledge-based System,<br />
<strong>Patient</strong> Medication <strong>Safety</strong>.<br />
1. Introduction<br />
Adverse Drug Events (ADEs) due to medication<br />
errors and human factors constitute a major public<br />
health issue, endangering patients’ safety and causing<br />
significant healthcare costs [1]. Information<br />
Technology (IT) is envisioned to play an important<br />
role towards the reduction of preventable ADEs by<br />
offering relevant knowledge and decision support<br />
tools to healthcare professionals along the entire<br />
prescription - dispensation - administration -<br />
compliance (PDAC) medication cycle [2]. Although<br />
there have been several efforts towards this direction,<br />
their efficiency was impeded by the lack of reliable<br />
knowledge about ADEs, and the poor ability of such<br />
solutions to deliver contextualized knowledge focused<br />
on the problem. The European project PSIP (<strong>Patient</strong><br />
<strong>Safety</strong> through Intelligent Procedures in medication)<br />
aims at preventing medication errors and ADEs in<br />
particular by: (1) facilitating the systematic<br />
production of epidemiological knowledge on ADEs,<br />
and (2) improving the PDAC cycle in hospitals by<br />
introducing sophisticated medication-related Clinical<br />
Decision Support System (CDSS) modules at the<br />
point of care based on the above knowledge.<br />
In particular, the first sub-objective involves<br />
generating new knowledge on ADEs, in order to<br />
identify as exactly as possible (per hospital) their<br />
number, type, consequences and causes. Data and<br />
semantic mining techniques applied on structured<br />
hospital databases and data collections of free-texts<br />
(letters, reports, etc.) respectively, have been<br />
employed to identify observed ADEs along with their<br />
frequencies and probabilities, thus giving a better<br />
understanding of potential risks. The second subobjective<br />
involves the development of a knowledgebased<br />
framework integrating the knowledge<br />
discovered with existing evidence, such as drug<br />
interactions and already known ADE cases reported in<br />
the literature. The aim of this framework is the<br />
delivery of contextualized knowledge fitting the local<br />
risk parameters in the form of alerts and decision<br />
support functions to healthcare professionals.<br />
Workshop <strong>Patient</strong> <strong>Safety</strong> through Intelligent Procedures in Medication, Sofia, Bulgaria, 2011<br />
13
The proposed knowledge framework constitutes the<br />
backbone of the PSIP platform that is independent of<br />
existing Clinical Information Systems, such as CPOE<br />
(Computerized Physician Order Entry) and EHR<br />
(Electronic Health Record) systems. It is important to<br />
note that the platform provides the appropriate<br />
connectivity mechanisms enabling such systems to<br />
access, exploit and integrate the incorporated<br />
knowledge in their local context. In this paper, we<br />
briefly present the overall rationale of PSIP and focus<br />
on the knowledge engineering approach employed<br />
towards the construction of a flexible Knowledgebased<br />
System (KBS) and framework constituting the<br />
core part of the CDSS modules for ADE prevention.<br />
2. ADE Prevention: A Systematic,<br />
Knowledge-based Approach<br />
The research and development activities of PSIP have<br />
been organized in the following three stages: (1)<br />
improvement of knowledge on ADEs, (2)<br />
development of CDSS modules, and (3)<br />
contextualization of CDSS modules and integration in<br />
existing IT solutions and usage. Initially, the available<br />
patient data have been identified residing at the<br />
Hospital Information Systems (HIS) of the<br />
participating healthcare organizations. In this context,<br />
a common data model has been developed [3], in<br />
order to exhaustively define the nature, type, and<br />
possible values of parameters that are potentially<br />
relevant to ADEs, including drug-related information,<br />
diagnostic information, lab examination results,<br />
description of procedures, demographic data, etc.<br />
The subsequent knowledge discovery activities that<br />
were applied focused primarily on the identification<br />
of clinical cases that are potential ADEs. In particular,<br />
the analysis was performed in more than 115,000<br />
patient records exported from RegionH hospitals<br />
(Copenhagen, Denmark), the General Hospital of<br />
Denain (Denain, France), the General Hospital of<br />
Rouen (Rouen, France), the University and Research<br />
Hospital of Lille (Lille, France) and the USHATE<br />
hospital (Sofia, Bulgaria). Among the techniques<br />
employed for knowledge discovery on ADEs were<br />
multivariate correspondence analysis, hierarchical<br />
classifications, decision trees and association rules<br />
[4]. The outcome of these activities resulted in: (1) a<br />
set of production rules of the form “IF Condition-1 &<br />
Condition-2 & ... & Condition-N THEN Effect-X”,<br />
corresponding to potential ADE signals, and (2)<br />
characterization of hospital stays according to their<br />
probability of showing (or being associated) with an<br />
ADE.<br />
The outcomes of the knowledge discovery phase have<br />
been filtered and validated, in order to eliminate<br />
artifacts or clinically irrelevant results [5]. More<br />
specifically, the above rules were confronted with the<br />
existing clinical pharmacological knowledge available<br />
in the scientific literature and/or in specialized<br />
information repositories such as the VIDAL®<br />
knowledge source (expertise available in the PSIP<br />
Consortium). In addition, specialist physicians or<br />
clinical pharmacologists extensively reviewed these<br />
data, in order to determine whether the corresponding<br />
patient stays present or not potential ADEs. As most<br />
of the abnormal stays were also attached to one or<br />
several rules, their review by human experts also<br />
allowed for assessing the capacity of the clinicians to<br />
understand the discovered knowledge, and the<br />
contextual clinical relevance of the rule(s) attached to<br />
each reviewed stay.<br />
Following the knowledge discovery and validation<br />
phase, a knowledge engineering framework has been<br />
established to articulate and electronically encode this<br />
knowledge, so that it may be efficiently exploited and<br />
incorporated in CDSS modules.<br />
3. Knowledge Engineering<br />
Knowledge employed in PSIP can be considered as<br />
belonging in three categories [6]: a) domain<br />
knowledge, in terms of types and facts, which is<br />
generally static and structured via concepts, relations<br />
– associations, attributes, and rule types (expressions);<br />
b) task knowledge, in terms of functional<br />
decomposition, and control; in this regard, knowledge<br />
is elaborated with respect to combination of tasks to<br />
reach a goal/workflow, or oppositely, decomposition<br />
of complex tasks into separate processes; c) inference<br />
knowledge, in terms of basic reasoning steps that can<br />
be made in the domain and are applied by tasks.<br />
The elaborated knowledge model architecture is<br />
depicted in Fig. 1. Roughly, its components are<br />
discriminated into the following categories:<br />
a) Drugs: Defines all possible drugs, containing also<br />
their categories and subcategories, based on the ATC<br />
(Anatomical Therapeutic Chemical) standard<br />
classification.<br />
b) Diagnosis: Defines medical conditions to be used<br />
as input parameters for identifying possible ADEs<br />
14<br />
Workshop <strong>Patient</strong> <strong>Safety</strong> through Intelligent Procedures in Medication, Sofia, Bulgaria, 2011
ased on the ICD-10 (International Classification of<br />
Diseases) standard classification.<br />
c) Lab results: Defines the terminology for<br />
expressing lab results in the “Conditions” part of<br />
ADE rules based on the C-NPU (Committee on<br />
Nomenclature, Properties and Units) standard<br />
classification of IUPAC (International Union of Pure<br />
and Applied Chemistry).<br />
d) Effects: Ontology-based representation of the<br />
effects as entities based on attributes containing the<br />
recommendation, the level of severity, type of risk,<br />
etc.<br />
e) <strong>Patient</strong> parameters: PSIP-specific ontology-based<br />
representation defining the terminology for expressing<br />
conditions of the ADE rules.<br />
f) Context parameters: Set of context-related<br />
parameters defined to allow future contextualization<br />
for the CDSS modules.<br />
g) Procedural Logic – Clinical Protocols and<br />
Guidelines: Description of clinical procedures,<br />
protocols and guidelines related to medication that<br />
aim to enable expressing knowledge related to human<br />
factors and complex ADE rules.<br />
h) ADE rules: The core component encapsulating<br />
knowledge about potential ADEs in the form of rules<br />
associating a number of conditions to an effect.<br />
Figure 1. Components of the knowledge model.<br />
According to the above, the PSIP knowledge model is<br />
defined as a set of ontology-based structures, either<br />
specific to the problem domain or standard<br />
classifications to be used as terminology. In addition,<br />
a rule-based component is included that is defined via<br />
a set of classes and populated with rules. The<br />
ontology-based structures and the rule-based<br />
component constitute the fundamental elements to<br />
define complex procedural logic in terms of protocols<br />
and guidelines, following an electronic formalism,<br />
i.e., the guideline modeling component. The<br />
conceptual description of the knowledge model is<br />
thoroughly presented in [6].<br />
4. Implementation and Results<br />
Figure 2 depicts the overall knowledge engineering<br />
and CDSS development procedures followed. Taking<br />
into account the requirements posed by the knowledge<br />
sources considered in the project and a state-of-the-art<br />
analysis, GASTON (http://www.medecs.nl/) was<br />
selected as the KBS/CDSS platform. The core of<br />
GASTON consists of a guideline representation<br />
formalism relying on a combination of knowledge<br />
representation approaches and concepts, i.e.,<br />
primitives, problem-solving methods (PSMs), and<br />
ontologies. This formalism uses ontologies as an<br />
underlying mechanism to represent guidelines in<br />
terms of PSMs and primitives in a consistent way.<br />
All the rules generated from the knowledge discovery<br />
activities have been implemented in the respective<br />
Knowledge Base (KB), in subsequent development<br />
iterations, while standard and problem-specific<br />
terminologies have been also incorporated and<br />
defined, respectively. Aiming to avoid potential<br />
ambiguities in the description of terms/concepts<br />
contained in the discovered knowledge, an XML<br />
schema has been defined and agreed upon for<br />
knowledge exchange among the knowledge discovery<br />
team and the knowledge authors participating in PSIP.<br />
In addition, appropriate inference mechanisms were<br />
developed, in order to filter rules and eliminate over<br />
alerting by applying contextual criteria. Aiming to<br />
provide a flexible KBS that can constitute the basis<br />
for deploying CDSS modules, various knowledge<br />
management tools were also developed. In particular,<br />
the construction of the KB and the generation of the<br />
corresponding CDSS modules are performed via fully<br />
automatic procedures using the source XML files. The<br />
procedure includes syntactic verification of the source<br />
files against the defined terminology, the population<br />
of the KB with instances (data, meta-data, rules, etc.)<br />
and the generation of CDSS modules capable of<br />
providing up to date knowledge on ADE prevention<br />
that corresponds to the local context (language,<br />
statistical significance of rules, etc).<br />
Workshop <strong>Patient</strong> <strong>Safety</strong> through Intelligent Procedures in Medication, Sofia, Bulgaria, 2011<br />
15
Figure 2. Knowledge engineering and CDSS development procedures followed.<br />
5. Conclusion<br />
Identification of ADEs constitutes a challenging<br />
problem, considering the complexity of medical<br />
information that has to be processed for ADE signals<br />
discovery, as well as the validation of this new<br />
knowledge with respect to its medical/clinical gravity.<br />
PSIP employed a systematic approach towards the<br />
identification, validation, and efficient representation<br />
of ADE-related knowledge. The ultimate goal of our<br />
research involved the development of flexible,<br />
contextualized CDSS modules for preventing ADEs<br />
in the entire PDAC medication chain at the point of<br />
care. To this end, the decision support services<br />
deployed are available for exploitation to clinical IT<br />
systems via a highly interoperable, scalable, and<br />
robust platform that provides the appropriate<br />
connectivity mechanisms enabling systems such as<br />
CPOE to access and integrate the relevant knowledge<br />
and functionality in their local context. Major<br />
challenges for future development include the<br />
incorporation of more complex knowledge sources in<br />
the KB, i.e., knowledge related to human factors and<br />
tacit knowledge, as well as exhaustive computational<br />
testing of the CDSS modules in real and demanding<br />
environments.<br />
6. Acknowledgements<br />
The research leading to these results has received<br />
funding from the European Community’s Seventh<br />
Framework Programme (FP7/2007-2013) under grant<br />
agreement n° 216130 PSIP (<strong>Patient</strong> <strong>Safety</strong> through<br />
Intelligent Procedures in medication), and has been<br />
performed by the PSIP Consortium.<br />
7. References<br />
[1] L.T. Kohn, J. Corrigan, and M. S. Donaldson, “To<br />
err is human: Building a safer health system”,<br />
National Academy Press, 2000.<br />
[2] D.W. Bates et al., “Detecting adverse events using<br />
Information Technology”, J Am Med Inform<br />
Assoc., vol. 10, no. 2, 2003, pp. 115–128.<br />
[3] E. Chazard et al., “Detection of adverse drug<br />
events: Proposal of a data model”, Stud Health<br />
Technol Inform., vol. 148, pp. 63–74, 2009.<br />
[4] E. Chazard et al., “Detection of adverse drug<br />
events: Data aggregation and data mining”, Stud<br />
Health Technol Inform., vol. 148, pp. 75–84,<br />
2009.<br />
[5] N. Leroy et al., “Human factors methods to<br />
support the experts’ review of automatically<br />
detected adverse drug events”, Stud Health<br />
Technol Inform., vol. 150, pp. 542–546, 2009.<br />
[6] V. Koutkias et al., Constructing clinical decision<br />
support systems for adverse drug event<br />
prevention: a knowledge-based approach, AMIA<br />
Annu. Symp. Proc. (2010) 402–406.<br />
16<br />
Workshop <strong>Patient</strong> <strong>Safety</strong> through Intelligent Procedures in Medication, Sofia, Bulgaria, 2011
Comparison of two methods for French Drug names extraction<br />
Suzanne Pereira 1 , Catherine Letord 1 , Sophie Tessier 1 , Solenne Rein 1 , Stefan Darmoni 2 , Elisabeth Serrot 1<br />
1 VIDAL, Issy-Les-Moulineaux, suzanne.pereira@vidal.fr & elisabeth.serrot@vidal.fr<br />
2<br />
CISMeF, Rouen, stefan.darmoni@chu-rouen.fr<br />
Abstract<br />
The paper describes two methods for drug names extraction<br />
for French medical records. We compared two approaches<br />
integrated in the FMTI (French multi-terminological and<br />
multilingual indexer) tool. The article summarizes the<br />
evaluations performed.<br />
Keywords<br />
Abstracting and Indexing/methods; algorithms; Information<br />
Storage and Retrieval/methods; Evaluation Study, France;<br />
Natural Language processing; Medical records; Drug<br />
therapy.<br />
1. Introduction<br />
Representing medical record data using standardized<br />
terminologies facilitates medical practice, information<br />
retrieval and costs management [1, 2].<br />
Structured drug data is needed for drug adverse events<br />
automatic detection.<br />
In some hospitals, information on drug prescriptions<br />
are available in the computerized physician order<br />
entry (CPOE). Unfortunately, all hospitals are not<br />
equipped (e.g. CHRU Lille and CHU Rouen) in this<br />
case the documents written in free text contain most<br />
of the drug prescriptions: hospitalization reports,<br />
letters, test results, etc.<br />
An extractor can extract this information based on the<br />
text (see figure 1). We have tested this idea during the<br />
PSIP project (<strong>Patient</strong> <strong>Safety</strong> through Intelligent<br />
Procedures in medication) [3, 4], using FMTI, an<br />
automatic extracting tool enable to extract, from<br />
documents, terms of several terminologies.<br />
The paper is structured as follows: Section 1 we begin<br />
with the study of existing similar works. Section 2<br />
describes our system and the evaluation of its<br />
performances for drug data extraction. Finally we<br />
discuss the results and conclude.<br />
2. Related Work<br />
Extraction of drug treatments in the hospital records<br />
and discharge summaries is a complex task. Indeed,<br />
drug treatments can be described in the reports using<br />
the brand names, the generic names, the substances,<br />
and abbreviations in any section of the document<br />
(Background, Evolution, Conclusion, Treatment,<br />
etc.).. Moreover, unlike most of the names of<br />
diseases, drug names or substances don’t follow the<br />
French syntax. The task is made even more difficult<br />
for these names hard to spell which cause the<br />
appearance of numerous misspellings in the<br />
documents. The natural language processing must be<br />
very specific here.<br />
Few computational approaches have been exploited<br />
until now:<br />
• Approaches based on lexicons have been explored.<br />
For example, Levin et al. [5] and Sirohi et al. [6],<br />
Evans et al. [7] developed a system based on<br />
lexicons (drug names and abbreviations).<br />
• Approaches based on regular expressions (names<br />
of generic or brand names) [7].<br />
• Approaches based on phonemization algorithms to<br />
handle spelling errors [5].<br />
Figure1. Example of a French discharge summary<br />
Workshop <strong>Patient</strong> <strong>Safety</strong> through Intelligent Procedures in Medication, Sofia, Bulgaria, 2011<br />
17
and the corresponding extraction of the drug names<br />
3. FMTI<br />
3.1.1 FMTI background<br />
F-MTI© 1 integrates 26 terminologies (ICD10,<br />
SNOMED3.5, ICD9, CCAM…) and the mappings<br />
between them, in 6 languages (French, English,<br />
Danish, Portuguese, Spanish, Deutch).<br />
For the PSIP project, VIDAL integrated four new<br />
terminologies and their mappings devoted to drugs<br />
into F-MTI’s knowledge sources: the Anatomical,<br />
therapeutical and chemical Classification (ATC)<br />
(N=5 514), the substances INN (N=2 942), the brand<br />
names (N=11,922) including all the abbreviations and<br />
generical names provided by the VIDAL Company<br />
and the MeSH Supplementary Concepts for chemical<br />
substances and pharmacological action terms<br />
translated into French by the CISMeF team<br />
(N=6 505).<br />
Called via a Web API, FMTI enables to extract<br />
automatically, from documents, terms of several<br />
terminologies. The indexing is performed in less than<br />
one second for a discharge summary.<br />
This tool was already evaluated for Web document<br />
indexing using MeSH [8], discharge summaries<br />
indexing using ICD10 [9] and SNOMED3.5 [10]. For<br />
drugs, the system was evaluated for therapeutical data<br />
extraction in SPC [11].<br />
In this article, we present FMTI performances for<br />
drug names extraction from discharge summaries in<br />
2009 and 2010. The first evaluation was published in<br />
2009 [12].<br />
3.1.2 FMTI system description<br />
The basic method of Bag of Words algorithm was<br />
using only stemming for normalization purpose [8]:<br />
For each discharge letters, the document is first<br />
broken into sentences. Then each sentence is<br />
normalized (accents are removed, all words are<br />
switched to lower case and stemmed…) and stop<br />
words are removed to form a bag of words containing<br />
all the content words. The “bag” thus obtained is<br />
matched independently of the order of the words<br />
against all the different terminologies terms that have<br />
been processed in the same way. All terms formed<br />
1<br />
FMTI is the proprietary of the VIDAL Company.<br />
with at least one word of the sentence are retrieved.<br />
VIDAL team worked on a phonemization algorithm<br />
to replace stemming, improve the system and solve<br />
some mistakes found in the previous study. Indeed, F-<br />
MTI encountered difficulties to recognize brand<br />
names due to incorrect spelling of the names in the<br />
discharge summaries. Some brand names are written<br />
improperly with dash ("-") or underscore ("_") or with<br />
an incorrect space " " (e.g. di-antalvic, diffu k, di<br />
hydan, cacit D, calcidose vit D, co renitec). On the<br />
contrary, some brand names were written without<br />
dash ("-") or underscore ("_") or space (" "), as<br />
normally they should have to (e.g. chibroproscar<br />
instead of chibro-proscar; bipreterax instead of<br />
bipreterax).<br />
Brand names are particular as they don’t follow the<br />
French syntax. Band Names can have English or Latin<br />
origins and in the spelling they don’t mean anything<br />
at all. That’s why using stemming (removing suffixes<br />
from the word) is logical for terms like those from<br />
ICD10 but not for Brand names. Another alternative<br />
of using the exact word is to use a phonemization<br />
algorithm. Phonemization [13] is based of the<br />
pronunciation of a word dependant of the language.<br />
The word is transformed into the corresponding list of<br />
phonemes. This method enables to take into account<br />
some incorrect spellings like “Eupanthyl” and<br />
“Eupantil”.<br />
The algorithm was significantly changed to address<br />
the specific purpose of brand names recognition.<br />
Specific English, Latin and Roman pronunciations<br />
were taken into account (“free” equal “fri”, “chol”<br />
equal “kol”, “I” equal “un”, etc.). Some abbreviations<br />
were added (“vit” equal “vitamin”), single letters<br />
(“PH”) were kept. The problems of dash, underscore<br />
and space were solved via alternative spellings. Some<br />
common words were removed (the brand name<br />
“PAR” equal to “par”). And finally some items of the<br />
discharge summary were not taken into account: “way<br />
of life” and “exams” that can produce some noise in<br />
the indexing.<br />
4. Evaluation<br />
In this study, the objective is to determine (a) the<br />
performances of FMTI using a stemming algorithm<br />
and (b) the gain of the use of a phonemization<br />
algorithm for extracting drug names from documents<br />
narrative.<br />
18<br />
Workshop <strong>Patient</strong> <strong>Safety</strong> through Intelligent Procedures in Medication, Sofia, Bulgaria, 2011
For this purpose, 50 discharge summaries were<br />
extracted from the corpus of 4,000 discharge<br />
summaries and letters used in the PSIP project to<br />
evaluate this task.<br />
Three key metrics were used to show the<br />
performances of the automatic extraction produced by<br />
FMTI compared to the reference, the manual<br />
identification: Precision, Recall and F-measure.<br />
5. Evaluation Results and Discussion<br />
5.1 Drug names extraction using the stemming<br />
algorithm: Agreement between FMTI and experts<br />
FMTI, using the stemming algorithm, enabled to<br />
index 50 French discharge summaries. Automatic<br />
indexing was manually analyzed by a CISMeF expert<br />
pharmacist in Rouen [13].<br />
• Overall precision is P = 0.84<br />
• Overall recall is R = 0.93<br />
• F-measure is F = 0.88<br />
5.2 Drug names extraction using the phonemization<br />
algorithm: Agreement between FMTI and experts<br />
FMTI, using the phonemization algorithm, enabled to<br />
index 50 French discharge summaries. Automatic<br />
indexing was manually analyzed by a VIDAL expert<br />
pharmacist in Issy-Les-Moulineaux.<br />
• Overall precision is P = 0.97<br />
• Overall recall is R = 0.95<br />
• F-measure is F = 0.96<br />
5.3 Discussion<br />
The results showed that the F-MTI tool obtained<br />
better results when using the customized<br />
phonemization algorithm than when using the<br />
stemming algorithm: plus 0.13 points for precision<br />
and plus 0.02 points for recall.<br />
The phonemization algorithm enabled to solve some<br />
problems linked with incorrect spelling of the brand<br />
names in the discharge summaries: dash ("-"),<br />
underscore ("_") and incorrect space " " (e.g. diantalvic,<br />
diffu k, di hydan, co renitec, chibro-proscar;<br />
bipreterax instead of bipreterax etc.). But we could<br />
not resolve the problem of words written with a dash<br />
in the discharge summary and written with no dash in<br />
the terminology.<br />
It also enabled to solve some misspellings or<br />
mistyping problems (e.g. ketoderme instead of<br />
ketoderm and dextropropoxifene instead of<br />
dextropropoxyfene). Some other misspellings quite<br />
frequent (e.g. triapridal instead of tiapridal,<br />
genopevaryl instead of gynopevaril, piperacetam<br />
instead of piracetam) could not be solved by this<br />
method. VIDAL team is working on similarity<br />
measures to improve the system.<br />
Some terms are ambiguous (e.g. albumin, which is<br />
both a laboratory result and a drug name). We will<br />
add a disambiguation algorithm to avoid the resulting<br />
noise.<br />
Finally, the phonemization algorithm is dependant of<br />
the language used. A specific algorithm is needed for<br />
each language used in FMTI.<br />
6. Conclusion<br />
We developed a multi-terminological and a multi<br />
lingual extractor useful for extracting drug names<br />
from free text. The evaluation showed that FMTI<br />
using a lexicon and a robust phonemization algorithm<br />
obtained very good results: 97% of precision, 95% of<br />
recall and 96% of F-measure. This is encouraging for<br />
our project PSIP, for which we have shown that in the<br />
absence of computerized prescribing system, FMTI<br />
could help extract drug information from the<br />
electronic patient record. After some improvements<br />
we hope to see FMTI integrated in an electronic<br />
health record system.<br />
7. Acknowledgements<br />
The research tasks leading to these results have<br />
received funding from the European Community’s<br />
Seventh Framework Programme (FP7/2007-2013)<br />
under grant agreement n° 216130 PSIP (<strong>Patient</strong> <strong>Safety</strong><br />
through Intelligent Procedures in Medication).<br />
8. References<br />
[1] Campbell JR, Carpenter P, Sneiderman C, Cohn S,<br />
Chute C, Warren J. Phase II Evaluation of Clinical<br />
Coding Schemes : Completeness, Taxonomy,<br />
Mapping, Definitions and clarity. JAMIA , 1997, pp.<br />
238-251.<br />
[2] Wasserman H, Wang J. An applied evaluation of<br />
SNOMED CT as a clinical vocabulary for the<br />
computerized diagnosis and problem list. AMIA<br />
Workshop <strong>Patient</strong> <strong>Safety</strong> through Intelligent Procedures in Medication, Sofia, Bulgaria, 2011<br />
19
Annu Symp Proc 2003; 699–703.<br />
[3] <strong>Patient</strong> <strong>Safety</strong> by Intelligent Procedures in<br />
medication. See : http://www.psip-project.eu [october<br />
2010].<br />
[4] Beuscart R, McNair P, Brender J. <strong>Patient</strong> safety<br />
through intelligent procedures in medication: the<br />
PSIP project. Stud Health Technol Inform. 2009 ;<br />
148 : 6–13.<br />
[5] Levin MA, Krol M, Doshi AM, and Reich DL.<br />
Extraction and mapping of drug names from free text<br />
to a standardized nomenclature. AMIA Annu Symp<br />
Proc 2007; 438–42.<br />
[6] Sirohi E, and Peissig P. Study of effect of drug<br />
lexicons on medication extraction from electronic<br />
medical records. Pac Symp Biocomput 2005; 308–18.<br />
[7] Evans DA, Brownlow ND, Hersh WR and Campbell<br />
EM. Automating concept identification in the<br />
electronic medical record: an experiment in<br />
extracting dosage information. Proc AMIA Annu Fall<br />
Symp 1996; 388–92.<br />
[8] Pereira S, Sakji S, Névéol A, Kergoulay I, Kerdelhué<br />
G, Serrot E, Joubert M, Darmoni SJ. Abstract multiterminology<br />
indexing for the assignment of MeSH<br />
descriptors. AMIA symp. 2009; 521–525.<br />
[9] Pereira S, Névéol A, Massari P, Joubert M, Darmoni<br />
SJ. Construction of a semi-automated ICD-10 coding<br />
help system to optimize medical and economic coding.<br />
Stud Health Technol Inform 2006; 124: 845–850.<br />
[10] Pereira S, Massari P, Buemi A, Dahamna B, Serrot E,<br />
Joubert M, Darmoni SJ. F-MTI : outil d'indexation<br />
multi-terminologique : application à l'indexation<br />
automatique de la SNOMED. Risques, technologies<br />
de l'information pour les pratiques médicales :<br />
comptes rendus des treizièmes journées francophones<br />
d'informatique médicale (JFIM), Informatique et<br />
santé 2009; 17: 57–67.<br />
[11] Pereira S, Plaisantin B, Korchia M, Rozanes N, Serrot<br />
E, Joubert M, Darmoni SJ. Automatic construction of<br />
dictionnaries, application to product characteristics<br />
indexing. MIE2009, the XXII International conference<br />
of the european deferation for medical informatics<br />
2009:150;512-516.<br />
[12] Merlin B, Chazard E, Pereira S, Serrot E, Sakji S,<br />
Beuscart R, Darmoni SJ. Can F-MTI semantic-mined<br />
drug codes be used for Adverse Drug Events detection<br />
when no CPOE is available? Proc. MEDINFO 2010;<br />
1025–1029.<br />
[13] Brouard F. L’art des Soundex. See:<br />
http://sqlpro.developpez.com/cours/soundex/<br />
20<br />
Workshop <strong>Patient</strong> <strong>Safety</strong> through Intelligent Procedures in Medication, Sofia, Bulgaria, 2011
Automatic Extraction of Entities from Hospital <strong>Patient</strong> Records<br />
in Bulgarian Language<br />
Galia ANGELOVA<br />
Institute of Information and Communication Technologies (IICT), Bulgarian Academy of Sciences (BAS)<br />
Block 25A, Acad. Georgi Bontchev Str., Sofia 1113, Bulgaria<br />
e-mail: galia@lml.bas.bg<br />
Abstract<br />
This paper overviews the tasks performed by IICT-BAS in<br />
the PSIP project. It considers the three experimental<br />
extractors implemented by the IICT team: for automatic<br />
mining of (i) ICD-10 diagnoses, (ii) drugs and dosages and<br />
(iii) values of clinical tests and lab data. The evaluation of<br />
these components is presented as well. The article also<br />
discusses issues related to the contextualisation and<br />
integration of all entities which have been automatically<br />
extracted from the texts of discharge letters.<br />
Keywords<br />
Automatic extraction from free text, contextualisation,<br />
integration<br />
1. Introduction<br />
The project <strong>Patient</strong> <strong>Safety</strong> through Intelligent<br />
Procedures in Medication (PSIP) investigates novel<br />
ICT scenarios for reducing the risk of preventable<br />
Adverse Drug Events (ADEs) in hospital settings.<br />
PSIP analyses large volumes of patient-related data,<br />
integrated within a decision-making framework, to<br />
provide accurate alerting, signalling of risks and<br />
supporting healthcare professionals in their decision<br />
making. The availability of large data bases with<br />
patient information is a must for the successful transferability<br />
of PSIP to a new hospital. Therefore the<br />
PSIP validation in USHATE (University Specialised<br />
Hospital for Active Treatment of Endocrinology<br />
“Acad. Iv. Penchev”, Medical University – Sofia)<br />
required to prepare an experimental repository of<br />
patient data, where ADEs can be discovered.<br />
It is well known, however, that a variety of important<br />
medical findings are traditionally stored as free text<br />
descriptions in the patient-related documentation.<br />
Many entities are structured in the Hospital<br />
Information Systems (HIS) – for instance, the drugs<br />
prescribed to the patients are maintained via the<br />
Hospital Pharmacy and the values of lab tests are<br />
automatically entered in predefined fields whenever<br />
the tests are made in the hospital. Nevertheless the<br />
textual paragraphs in the hospital <strong>Patient</strong> Records<br />
(PRs) contain essential information that needs to be<br />
mined, extracted and structured. The task of IICT-<br />
BAS in PSIP was focused on the extraction challenge<br />
in order to compliment the entities available in the<br />
HIS of USHATE. The team has developed software<br />
modules which mine the free texts of PR discharge<br />
letters to identify entities important for PSIP:<br />
• a drug extraction component, which recognises<br />
drug names and dosages and juxtaposes the<br />
corresponding ATC code to each drug name;<br />
• a component for extraction of ICD-10 codes,<br />
which assigns ICD-10 codes to accompanying<br />
diseases enumerated in the PR text (the principal<br />
diagnose is inserted in the HIS), and<br />
• a component for extraction of values of clinical<br />
tests and lab data which delivers information<br />
about the tests made outside USHATE (these<br />
values are typed in the PR text paragraphs).<br />
This article briefly describes the extractors, presents<br />
the extraction accuracy and discusses the integration<br />
and contextualisation aspects.<br />
2. Automatic Extraction of Entities from<br />
the Free Texts of Discharge Letters<br />
The performance of Information Extraction (IE) from<br />
clinical narratives gradually improves and currently<br />
exceeds 90% sensitivity, especially for English texts<br />
[1]. There are numerous potential IE applications with<br />
far-reaching effects but the present prototypes are<br />
mostly used in research settings because of scalability<br />
and generalisability issues. Nevertheless automatic<br />
text analysis is viewed as the only means to cope with<br />
the enormously large amounts of medical information<br />
available in textual form [2]. Projects for automatic<br />
text processing are running for dozens of natural<br />
Workshop <strong>Patient</strong> <strong>Safety</strong> through Intelligent Procedures in Medication, Sofia, Bulgaria, 2011<br />
21
languages and they all deliver slowly but gradually<br />
improving components for automatic recognition of<br />
various kinds of medical entities. Semantic miners for<br />
French language have been developed in the PSIP<br />
project too; they mark the state-of-the-art results for<br />
languages other than English. The French Multi-<br />
Terminology Indexer F-MTI is applied for the<br />
automatic detection of ADEs in discharge letters. It<br />
extracts ATC codes from the free text of French<br />
discharge letters with accuracy 88% when compared<br />
to the manual extraction [3, 4].<br />
Working for Bulgarian language and performing<br />
some tasks for the first time (e.g. drug extraction), we<br />
took advantage of the relatively structured paragraphs<br />
in the hospital discharge letters. These letters have<br />
mandatory content which is fixed in the Official State<br />
Gazette as Article 190(3) of the legal Agreement<br />
between the National Health Insurance Fund and the<br />
Bulgarian Medical and Dental Associations [5]. It<br />
seems that the (somewhat conservative) practice to<br />
segment the PRs into sections exists in Bulgaria since<br />
many decades and results from the centralised health<br />
requirements that were imposed to all state hospitals<br />
during the last 60-70 years. Table 1 shows the<br />
percentage of PRs with available standard sections in<br />
a corpus of 1,300 PRs.<br />
Section title PRs containing the section<br />
Diagnoses 100%<br />
Anamnesis 100%<br />
Past diseases 88,52%<br />
Allergies risk factors 43,56%<br />
Family Medical History 52,22%<br />
<strong>Patient</strong> Status 100%<br />
Lab data, clinical tests 100%<br />
Examiners’ comments 59,95%<br />
Debate 100%<br />
Treatment 26,70%<br />
Table 1: Percentage of PRs including standard sections<br />
(which can be automatically split with accuracy 99,99%)<br />
In this way, when our ICD-10 code extractor analyses<br />
the discharge letter texts in order to find diagnoses of<br />
accompanying diseases, it mines the section<br />
Diagnoses only – i.e. paragraphs of nominal phrases<br />
which enumerate only names of illnesses or special<br />
states (e.g. 'deficiency of Vitamin D'). Similarly, the<br />
drug extractor mines the Anamnesis for accompanying<br />
drugs; the Anamneses often contain subtitles like e.g.<br />
“therapy at the admission” or other phrases which<br />
signal current therapy with high probability. The<br />
extraction of lab data values is performed over the<br />
text of the sections “Lab data, clinical tests” which<br />
contain only this kind of values. Thus the extraction<br />
algorithms operate on input texts with relatively<br />
predictable structure and content, they are more<br />
precise and the system performance for Bulgarian<br />
texts is very good. The extensive evaluation of these<br />
extractors was performed on 6200 anonymised PRs of<br />
USHATE patients, diagnosed with diabetes and other<br />
endocrine diseases. The results show very high<br />
extraction accuracy for the three extractors [6, 7]:<br />
• The drug extraction component identifies 1,537<br />
drug names in 6200 PRs with accuracy 98,42%<br />
and dosage with accuracy 93,85%,<br />
• The component for extraction of ICD-10 codes<br />
assigns correctly ICD-10 codes to 84,5% of the<br />
disease names enumerated in the PR section<br />
'Diagnoses', and<br />
• The component for extraction of values of<br />
clinical tests and lab data works with precision of<br />
98,2% for the corpus of 6200 anonymised PRs.<br />
The in-depth overview of related work shows that the<br />
performance of these miners for Bulgarian language is<br />
comparable to the state-of-the-art achievements in<br />
biomedical Information Extraction.<br />
A major difficulty in the implementation was due to<br />
the lack of electronic linguistic resources for<br />
Bulgarian language. For instance, the dictionary of<br />
drugs names in Bulgarian was compiled semiautomatically<br />
with final manual assignment of<br />
relevant ATC codes to Bulgarian drug names. The<br />
USHATE Hospital Pharmacy operates with 1,182<br />
drugs but occurrences of another 355 drug names<br />
were found in the discharge letter texts. Thus the<br />
present version of the drug extractor uses a dictionary<br />
of 1537 drug names. Similarly, the dictionary of<br />
measurement units and Bulgarian names of clinical<br />
tests was compiled from the raw texts in the PSIP<br />
corpus especially for this project.<br />
Another complication is related to the established<br />
medical practice to use terminology in both Bulgarian<br />
and Latin, mixing Cyrillic and Latin symbols. This<br />
requires to develop multilingual dictionaries and to<br />
support long lists of transliteration rules. It turned out<br />
that considerable amount the efforts, needed for<br />
transferring PSIP to a new language environment,<br />
have to be invested in the construction of machinereadable<br />
vocabulary in the respective language.<br />
22<br />
Workshop <strong>Patient</strong> <strong>Safety</strong> through Intelligent Procedures in Medication, Sofia, Bulgaria, 2011
3. Contextualisation and Integration<br />
Our extractors deliver three types of entities identified<br />
in the PR texts and we need a strategy how to<br />
integrate them into coherent data sets with the values,<br />
available in the USHATE HIS for each patient. The<br />
diagnoses of accompanying diseases are clearly<br />
relevant at the hospitalisation moment; similarly, all<br />
clinical tests made outside USHATE for all patients<br />
are anchored to Day 0 of the respective hospitalisation<br />
as these tests are entered as valid ones in the<br />
corresponding PR sections. Thus the main issue is to<br />
invent an algorithm for automatic recognition of<br />
'current drug treatment'. This is a nontrivial task<br />
which can be tackled only by heuristics collected on a<br />
representative PR corpus.<br />
Drug names as tokens participate in all PR sections,<br />
including Diagnoses (e.g. 'Amiodaron-induced<br />
hypothyroidism'). We have found 355 drugs in the<br />
experimental corpus of 6200 PRs which might be<br />
taken by the patients for accompanying and chronic<br />
diseases (in addition to the 1182 drugs that are in use<br />
in USHATE during the period relevant for our<br />
experiment). Table 2 shows the frequency of<br />
occurrence of these 'external' 355 drugs in the PR<br />
sections. So we need to know which drugs, brought to<br />
the hospital by the patient, are taken at the moment of<br />
hospitalisation and during the hospital stay. Table 3<br />
shows the accuracy of recognition of these 355 drug<br />
occurrences in the PR texts. Some percentage of<br />
hypothetical or future treatment discussions for these<br />
drugs is wrongly considered as actual one (6% overgeneration).<br />
PR sections<br />
Percentage<br />
In the Anamnesis under header<br />
“Accompanying treatment”<br />
6%<br />
As prescription by External medical<br />
examiner (e.g. gynecologist)<br />
26%<br />
In the Debate and Treatment 68%<br />
Table 2: Distribution of occurrences of 355 'external' drugs<br />
in the PR sections<br />
Evaluated feature<br />
Percentage<br />
Precision 88%<br />
Sensitivity 92,45%<br />
Over-generation 6%<br />
Table 3: Accuracy of extraction of medication events<br />
related to mentioning 355 'external' drugs in the PR texts<br />
Heuristic observations of the local context explicate<br />
the typical phrases signaling descriptions of drug<br />
treatments. Most often the text expression, signalling<br />
the treatment at Day 0, is the phrase “at the<br />
admission” (при постъпването). This phrase occurs<br />
with slight variations in 2122 PRs (34%). On average<br />
25% of all drugs in the Anamnesis are recognised as<br />
“Day 0 medication”. The second preferred text<br />
expression is “at the moment” (в момента... ). It<br />
occurs in 908 PRs from the test set of 6,200 PRs; in<br />
703 PRs (77% of all occurrences) the phrase refers to<br />
explicitly specified drugs in the local context. The<br />
heuristic observations are based on automatically<br />
prepared concordances of local contexts but the final<br />
inspection is manual, e.g. the abovementioned 703<br />
cases were encountered after reading and marking of<br />
the list of all 908 occurrences. Please note that these<br />
typical expressions can be used in other phrases as<br />
well, e.g. “therapy at the admission: none”, “at the<br />
moment without complains” hence the training phase<br />
of our algorithms includes learning of positive and<br />
negative examples.<br />
The three extractors, implemented by IICT-BAS,<br />
were run on ,200 anonymised PRs. They have<br />
delivered numerous data items to the PSIP repository<br />
which has been constructed in USHATE (see Table<br />
4). Many values of clinical tests and lab data were<br />
available in the USHATE HIS, so the automatic<br />
extraction contributed mostly values of hormonal tests<br />
(which are often made outside the hospital). These<br />
entities were integrated with the HIS data to constitute<br />
the PSIP repository; and the repository was used for<br />
discovery of USHATE-specific ADEs. In all cases of<br />
overlapping descriptions the HIS data are preferred as<br />
more exact and reliable. Actually the extractors,<br />
developed for the Bulgarian PR texts, provided<br />
interoperability between the USHATE PRs and the<br />
PSIP data formats: once structured information is<br />
extracted from the free texts, it can be recorded in<br />
various databases using ATC and ICD codes. Further<br />
details are available in [8, 9].<br />
Number<br />
Entity type of Entities Precision Sensitivity<br />
extracted from<br />
the PR texts<br />
Diagnoses 26,826 97.30% 74.69%<br />
Drugs 160,892 97.28% 99.59%<br />
Values of<br />
lab tests<br />
114,441 99.99% 99.04%<br />
Table 4: Entities extracted automatically from free PR texts<br />
Workshop <strong>Patient</strong> <strong>Safety</strong> through Intelligent Procedures in Medication, Sofia, Bulgaria, 2011<br />
23
The excellent performance of the drug extractor<br />
enabled its on-line integration in the USHATE<br />
validation testbed. When storing an Anamnesis in the<br />
hospital information system, the drug extractor<br />
automatically displays the list of drugs taken at the<br />
moment of hospitalisation (see Figure 5 in [10]). The<br />
validating doctors were quite positive about the<br />
experimental integration of this extractor as an on-line<br />
analyser because it provides structured data in a<br />
convenient format that can be further used for<br />
prescriptions. However, we note that the extractor<br />
uses a pre-fixed dictionary of drug names; if a quite<br />
new drug name appears it will not be recognised.<br />
Thus the integration of IE modules requires constant<br />
updates of the system dictionaries.<br />
4. Conclusion<br />
Here we have presented a research effort in automatic<br />
extraction of structured information from hospital<br />
PRs, performed in order to integrate a data base for<br />
experimental discovery of ADEs. The repository for<br />
PSIP validation in USHATE is relatively small but<br />
sophisticated as it comprises the HIS data as well as<br />
the entities delivered by the automatic extractors for<br />
each patient. Our prototypes demonstrate very high<br />
extraction accuracy which is partly implied by the<br />
established structure of the discharge letters in<br />
Bulgarian hospitals. The false or negative results<br />
(including overgeneration) are an inevitable aspect of<br />
the IE performance. In our case the small percentage<br />
of false positive entities is statistically insignificant<br />
and practically negligible because PSIP has its own<br />
thresholds for discovery of ADEs and issuing alerts<br />
(we note that manual human processing of clinical<br />
texts would also include some percentage of errors).<br />
In this way our prototypes support the statement that<br />
the automatic information extraction is a valuable<br />
component in eHealth projects dealing with focused<br />
analysis of patient data.<br />
5. Acknowledgements<br />
The research tasks leading to these results have<br />
received funding from the European Community’s<br />
Seventh Framework Programme (FP7/2007-2013)<br />
under grant agreement n° 216130 PSIP (<strong>Patient</strong> <strong>Safety</strong><br />
through Intelligent Procedures in Medication).<br />
6. References<br />
[1] Meystre, S., G. Savova, K. Kipper-Schuler, and J.F.<br />
Hurdle. Extracting Information from Textual<br />
Documents in the Electronic Health Record: A Review<br />
of Recent Research. IMIA Yearbook of Medical<br />
Informatics, 2008, pp. 138–154.<br />
[2] Demner-Fushman, D., W. Chapman and C. McDonald.<br />
What can natural language processing do for clinical<br />
decision support? Journal of Biomedical Informatics,<br />
Volume 42, Issue 5, October 2009, pp. 760-772.<br />
[3] Merlin B., E. Chazard, S. Pereira, E. Serrot, S. Sakji, R.<br />
Beuscart, and S. Darmoni. Can F-MTI semantic-mined<br />
drug codes be used for Adverse Drug Events detection<br />
when no CPOE is available? Stud. Health Technol.<br />
Inform. 2010, 160 (pt 1), pp. 1025-1029.<br />
[4] Pereira, S., C. Letord, S. Tessier, S. Rein, S. Darmoni,<br />
and E. Serrot. Comparison of two methods for French<br />
Drug names extraction. In R. Beuscart, D.<br />
Tcharaktchiev, G. Angelova (Eds) Proc. Int. Workshop:<br />
<strong>Patient</strong> <strong>Safety</strong> through Intelligent Procedures in<br />
Medication, INCOMA, 2011, pp. 17-20.<br />
[5] National Framework Contract between the National<br />
Health Insurance Fund, the Bulgarian Medical<br />
Association and the Bulgarian Dental Association,<br />
Official State Gazette №106/30.12.2005, updates<br />
№68/22.08.2006 and №101/15.12.2006, Sofia,<br />
http://dv.parliament.bg/.<br />
[6] Tcharaktchiev, D., G. Angelova, S. Boytcheva, Z.<br />
Angelov, and S. Zacharieva. Completion of Structured<br />
<strong>Patient</strong> Descriptions by Semantic Mining. In Koutkias<br />
V. et al. (Eds), <strong>Patient</strong> <strong>Safety</strong> Informatics, Stud. Health<br />
Technol. Inform. 2011 Vol. 166, IOS Press, 2011, pp.<br />
260-269.<br />
[7] Boytcheva, S. Shallow Medication Extraction from<br />
Hospital <strong>Patient</strong> Records. In Koutkias V. et al. (Eds),<br />
<strong>Patient</strong> <strong>Safety</strong> Informatics, Stud. Health Techn. Inform.<br />
2011 Vol. 166, IOS Press, 2011, pp. 119-128.<br />
[8] Boytcheva, S., D. Tcharaktchiev and G. Angelova.<br />
Contenxtualisation in Automatic Extraction of Drugs<br />
from Hospital <strong>Patient</strong> Records. To appear in Proc. of<br />
MIE-2011, the 23th Int. Conf. of EFMI, Norway, 28-31<br />
August 2011, published by IOS Press.<br />
[9] Boytcheva, S., G. Angelova, Z. Angelov, D.<br />
Tcharaktchiev, and H. Dimitrov. Integrating <strong>Patient</strong>-<br />
Related Facts using Hospital Information System Data<br />
and Automatic Analysis of Free Text. To appear in the<br />
Proc. of SALUS (Special track on electronic<br />
healthcare), International Conference MURPBES:<br />
Multidisciplinary Research and Practice for Business,<br />
Enterprise and Health Information Systems, Vienna,<br />
Austria, 22-26 August 2011, Springer LNCS.<br />
[10] Dimitrov, H. Testbed for integration of CDSS modules<br />
and PSIP validation in USHATE MU Sofia. In R.<br />
Beuscart, D. Tcharaktchiev, G. Angelova (Eds) Proc.<br />
Int. Workshop: <strong>Patient</strong> <strong>Safety</strong> through Intelligent<br />
Procedures in Medication, INCOMA, 2011, pp. 39-48.<br />
24<br />
Workshop <strong>Patient</strong> <strong>Safety</strong> through Intelligent Procedures in Medication, Sofia, Bulgaria, 2011
Automatic extraction of<br />
CD Codes for Diagnoses and<br />
ATC Codes and Dosages for Drugs<br />
Svetla BOYTCHEVA<br />
Institute of Information and Communication Technologies (IICT),<br />
Bulgarian Academy of Sciences and<br />
State University of Library Studies and Information Technologies,<br />
Sofia, Bulgaria<br />
svetla.boytcheva@gmail.com<br />
Abstract<br />
This paper presents approach for automatic<br />
extraction of diagnoses and medication information<br />
from discharge letters. The proposed algorithms are<br />
designed for processing free text document in<br />
Bulgarian language. Two algorithms were designed<br />
and implemented: an algorithm for mapping of<br />
International Classification of Diseases 10 th<br />
revision (ICD10) to diagnoses extracted from<br />
patient records (PRs) and an algorithm for<br />
automatic extraction of drug events and association<br />
to them of codes from Anatomical Therapeutic<br />
Chemical (ATC) Classification System. The system<br />
presented in here was developed and applied in the<br />
PSIP project for the preparation of an experimental<br />
repository for PSIP validation from University<br />
Specialized Hospital for Active Treatment of<br />
Endocrinology “Acad. I. Penchev” (USHATE) at<br />
Medical University – Sofia.<br />
Keywords<br />
Medical Information Extraction, Semantic mining<br />
1. Introduction<br />
We deal with hospital <strong>Patient</strong> records (PRs) which are<br />
anonymized by the hospital information system of the<br />
University Specialized Hospital for Active Treatment<br />
of Endocrinology “Acad. I. Penchev” (USHATE) at<br />
Medical University – Sofia.<br />
The system presented in here was developed and<br />
applied in the PSIP project for the preparation of an<br />
experimental USHATE’s repository for PSIP<br />
validation [1].<br />
PRs in all Bulgarian hospitals have mandatory<br />
structure, which is published in the Official State<br />
Gazette within the legal Agreement between the<br />
Bulgarian Medical Association and the National<br />
Health Insurance Fund [2].<br />
In the hospital PRs, medical terminology is<br />
recorded in both Bulgarian and/or Latin language.<br />
There is no preferred language for the terminology so<br />
the two forms are used like synonyms.<br />
2. Drug Events Information Extraction<br />
We have developed automatic procedures [3] for<br />
analysis of free texts in hospital patient records in<br />
order to extract information about: drug names;<br />
dosages; modes; frequency and treatment duration and<br />
to assign the corresponding ATC code [4] to each<br />
medication event.<br />
Actually the system enables extraction of drug<br />
related information about drugs which are mentioned<br />
in the PR texts as accompanying medications but are<br />
not prescribed by the Hospital Pharmacy.<br />
The list of registered drugs in Bulgaria is provided<br />
by the Bulgarian Drug Agency [5] and it contains<br />
about 4,000 drug names and their ATC codes.<br />
Currently the Hospital Pharmacy operates with 1,537<br />
medications because USHATE is specialized mostly<br />
for treatment of diabetic patients.<br />
Our aim in the PSIP project is to extract<br />
information about drugs, taken by the patient, which<br />
are not prescribed via the Hospital Pharmacy.<br />
There are two versions of the implemented<br />
algorithms for information extraction of drug events<br />
from discharge letters in Bulgarian language:<br />
Retrospective – This version was used in the<br />
initial preprocessing stage, where major task was<br />
Workshop <strong>Patient</strong> <strong>Safety</strong> through Intelligent Procedures in Medication, Sofia, Bulgaria, 2011<br />
25
information collection from full text discharge letters<br />
included in USHATE archive (total 6200 PRs ). The<br />
implemented system usually uses automatic model for<br />
analysis of all PRs in experimental repository (Fig.<br />
10). For this task were extracted drug events from all<br />
sections of text personal records. In order to<br />
determine which drugs were used for treatment during<br />
hospitalization we maintained information from:<br />
Anamnesis – including history of disease<br />
and information for previous and current<br />
treatment;<br />
Medical Examiners comments – including<br />
information for treatment of accompanying<br />
diseases;<br />
Debate – including information for the<br />
provided treatment during hospitalization.<br />
The extracted information concerning drug events<br />
from these three sections is processed as follows:<br />
If some drug is mentioned in the Anamnesis<br />
section, but not discussed further in Debate<br />
section, this means that treatment with it was<br />
stopped and we are not concerning this drug as<br />
“current treatment”.<br />
If some drug is mentioned in the Anamnesis<br />
section and discussed further in debate section,<br />
this means this drug should be marked with<br />
“current treatment” marker. For such drugs we use<br />
last mentioned information about their dosage,<br />
because it can happened at hospital admission<br />
daily scheme to be changed. In such cases this new<br />
information is placed in Debate section. If the<br />
current treatment continues with the same scheme<br />
<br />
as admission we use information about dosage<br />
from Anamnesis section (if any)<br />
All drugs mentioned in Medical Examiners<br />
comments section are marked with “current<br />
treatment” marker, because they are recognized by<br />
the system with high confidence as present events<br />
(and not future prescriptions), except cases with<br />
'stop' and 'replace' phrases.<br />
Prospective This is the integrated version<br />
with Hospital information system at USHATE. There<br />
is also available manual mode of the system mainly<br />
used for testing and experiments (Fig. 1). For this<br />
task the algorithm provides runtime information<br />
extracted from Anamnesis section only, because the<br />
whole discharge letter is not available during the<br />
hospital admission. If anamnesis contains section<br />
“Current treatment”, system extracts information<br />
available only in this section. In other cases we<br />
process temporal information concerning history of<br />
disease and treatment in Anamnesis. All text is<br />
separated by temporal events markers and segments<br />
between them are concerned as episodes describing<br />
diagnoses, treatment, symptoms and complain. The<br />
negation events are taken in consideration as well.<br />
Drugs in Anamnesis, only if they are listed under the<br />
headers 'medication at the moment of hospitalization'<br />
or 'accompanying treatment'. Some phrases like<br />
'started treatment with' can be also interpreted as hints<br />
for 'current medication' but only if they are not<br />
followed by phrases including 'replaced by' which<br />
signal past events.<br />
<br />
Fig. 1 Manual mode of the system<br />
Allows processing of a single PR stored as<br />
text file. The text of PR is opened in section<br />
(1). After opening the text file, PR is<br />
automatically separated on sections and the<br />
text from each section is presented in<br />
separated tab (2): (i) personal data; (ii)<br />
diagnoses Diag; (iii) anamnesis; (iv) patient<br />
statusStatus; (v) lab dataabs; (vi) medical<br />
examiners comments Consult; (vii)<br />
discussion Debate; (viii) treatment; and<br />
(ix) recommendations. The first section<br />
“personal data” is skipped, because we<br />
process anonymized PRs. The third section<br />
“anamnesis” is spited on two tabs – Current<br />
Treatment (if any) and Anamnesis. The<br />
information from the last two sections is<br />
merged and presented in the Treatment tab.<br />
26<br />
Workshop <strong>Patient</strong> <strong>Safety</strong> through Intelligent Procedures in Medication, Sofia, Bulgaria, 2011
Fig. 2 Drug events manual analysis<br />
PR section can be processed separately by<br />
the system. For the main task “identification<br />
of current treatment at hospital admission”<br />
we process information from “Current<br />
Treatment” section, if available or in other<br />
case from “Anamnesis”. After choosing<br />
“Analyze” function from menu bar the<br />
selected section (1) is processed and<br />
automatically is generated list with<br />
recognized drug names within the text. In<br />
this example from “Current Treatment”<br />
section are recognized three drugs<br />
“достинекс” (Dostinex), „лтиоксин” (<br />
Thyroxin) and „дилтиазем” (Diltiazem). In<br />
this case all events are marked as present.<br />
<br />
<br />
<br />
<br />
<br />
<br />
Fig. 3 Drug events Current Treatment<br />
Drugs included in the generated list can be<br />
processed separately by using “Find” button<br />
and selecting drug name from the list or as<br />
bunch using “Find All” button.<br />
In the current example is selected<br />
“достинекс” (Dostinex) from the list (2).<br />
The system identifies in the PR text (1) the<br />
scope of the corresponding drug event and<br />
presents it in section (3). Then using rules<br />
and regular expressions is identified<br />
information about drug name, dosages,<br />
modes, frequency and treatment duration<br />
(4). The system assigns the corresponding<br />
ATC code to this medication event (5) and<br />
stores information in CSV format (6).<br />
<br />
<br />
<br />
<br />
Fig. 4 Drug events disambiguation<br />
In this example in check box list (5) are<br />
presented four possible drug packs for „л<br />
тиоксин” (Thyroxin) with 0.5 mg, 1 mg<br />
and 50 mg. Because in the PR text (1) the<br />
recognized dosage is 50 micrograms the<br />
closest dosage of 0.5 milligrams is selected<br />
from the list (5) and the ATC code of<br />
“лтиоксин 0.5 м” (Thyroxin 0.5 mg)<br />
is associated to this drug event.<br />
<br />
<br />
Workshop <strong>Patient</strong> <strong>Safety</strong> through Intelligent Procedures in Medication, Sofia, Bulgaria, 2011<br />
27
Fig. 5 Drug events daily dosage<br />
In this example in check box list (5) is<br />
presented only one possible drug packs for<br />
„дилтиазем” (Diltiazem). Because there is<br />
no other options its ATC code is<br />
automatically is associated to this drug<br />
event. In the PR text (1) the recognized<br />
dosage is 2x90 mg per day and in section (4)<br />
is calculated the daily dosage 180 mg.<br />
<br />
<br />
Fig. Drug events negation arers<br />
In some cases even there is available<br />
“Current treatment” section in PR it<br />
contains negation marker.<br />
In this example “няма” (No) causes<br />
association of empty drug events list for<br />
current treatment at hospital admission.<br />
Fig. Drug events nanesis<br />
In this case “Current treatment” section is<br />
not available and the system processes<br />
temporal events in “Anamnesis” section. In<br />
this example as current treatment is<br />
recognized the sentence “по повод на<br />
стенокардна симптоматика е хоспитализиран<br />
в кардиологична клиника, където е започнато<br />
лечение с оликард и изоптин, което<br />
продължава и до момента.” (....due to anginal<br />
symptoms was hospitalized in the cardiology<br />
clinic, where started treatment with with<br />
Olicard (Olicard 40 retard) and Isoptin,<br />
and which continues to date ...) Because for<br />
Olicard and Isoptin is not presented<br />
information for dosage we use N/A marker<br />
and further use DDD (default daily dosage).<br />
28<br />
Workshop <strong>Patient</strong> <strong>Safety</strong> through Intelligent Procedures in Medication, Sofia, Bulgaria, 2011
In this case “Current treatment” section is<br />
not available and the system processes<br />
temporal events in “Anamnesis” section. In<br />
this example as current treatment is<br />
recognized information in the sentence “от<br />
началото на заболяването на<br />
интензифицирана схема с новорапид и<br />
лантус, дозировки при постъпването<br />
новорапид 6+6+6E, лантус 18E в 22ч.”<br />
(Since the beginning of the decease with<br />
intensified regimen of treatment with Lantus and<br />
NovoRapid, dosages at hospital admission<br />
NovoRapid 6+6+6E , Lantus 18e at 10 pm.)<br />
For the scheme 6+6+6E for NovoRapid is<br />
calculated daily dosage 18E.<br />
<br />
From example in Fig.8 for Lantus is<br />
recognized daily dosage 18E. If some drug name<br />
is mentioned more than one time in the text, than<br />
using rules for temporal events analyses we<br />
choose the dosage for the closest to present time<br />
marker.<br />
<br />
In automatic mode can be processed all PRs<br />
stored in the selected folder. The result file<br />
can be generated in CSV format, XML<br />
format or Excel format. Extracted<br />
information contains PR ID, ATC, drug<br />
name and drug pack, daily dosage, mode<br />
and scope of the recognized drug event from<br />
the narrative text.<br />
Workshop <strong>Patient</strong> <strong>Safety</strong> through Intelligent Procedures in Medication, Sofia, Bulgaria, 2011<br />
29
3. Diagnoses Information Extraction<br />
The Bulgarian hospitals are reimbursed by the<br />
National Insurance Fund via the “clinical pathways”<br />
scheme. When a patient is hospitalized, they often<br />
select from the Hospital Information System (HIS)<br />
menu one diagnosis which is sufficient for the<br />
association of the desired clinical pathway to the<br />
respective patient. Thus most of complementary<br />
diseases diagnosed by the USHATE medical experts<br />
are recorded in the personal history as free text. They<br />
are entered in the Discharge letter section Diagnoses<br />
as free text and some of them are only mentioned in<br />
the section Discussion of the discharge letter.<br />
This developed an approach for automatic<br />
mapping of International Classification of Diseases<br />
1th revision (ICD1) [6] to diagnoses extracted<br />
from discharge letters. The proposed algorithms are<br />
designed for processing free text document in<br />
Bulgarian language.<br />
Diseases are often described in the medical patient<br />
records as free text using terminology, phrases and<br />
paraphrases which differ significantly from the ICD<br />
disease description. than those used in ICD1<br />
classification. In this way the task of diseases<br />
recognition (which practically means e.g. assigning<br />
standardized ICD codes to diseases’ names) is an<br />
important natural language processing (NLP)<br />
challenge [8].<br />
To solve this task we are using the following<br />
resources provided in Bulgarian language [6]:<br />
ICD1 in Excel format<br />
Index of diseases and pathological states and<br />
their modifications (Fig. 11).<br />
Fig. 11 Index of diagnoses, pathological states and their<br />
modifications<br />
The other obstacle is mixture of Latin and<br />
Bulgarian terminology used in free text diagnoses<br />
presentation. For some of Latin terms is used<br />
transliteration in Cyrillic.<br />
This component works in three steps [1,6]: (i)<br />
shallow text analysis by regular expressions and<br />
patterns matching, (ii) searching disease names in the<br />
terminology resource bank medical terminology<br />
dictionary, list of abbreviations rules and Latin –<br />
Cyrillic transliteration rules, and (iii) application of<br />
terminology binding rules manually added by experts.<br />
The regular expressions, applied at step (i) for shallow<br />
syntactic analysis, encode grammatical patterns of<br />
text phrases which describe medication events in the<br />
particular training corpus (of endocrinology patients<br />
treated at USHATE). These expressions are extracted<br />
semiautomatically from the training texts by machine<br />
learning techniques.<br />
<br />
<br />
Fig. 12 Diagnoses manual analysis<br />
Allows processing of a single PR stored<br />
as text file. The text of PR is opened in<br />
section (1). After opening the text file, PR<br />
is automatically separated on sections and<br />
the text from diagnoses section is<br />
displayed in section (2). After choosing<br />
“Analyze” function from menu bar the<br />
extracted text in section (2) is processed<br />
and automatically is generated list (2)<br />
with recognized diagnoses within the text.<br />
<br />
30<br />
Workshop <strong>Patient</strong> <strong>Safety</strong> through Intelligent Procedures in Medication, Sofia, Bulgaria, 2011
Fig. 1 1 aignn<br />
Diagnoses included in the generated list<br />
(2) can be processed separately by using<br />
“Find” button and selecting diagnose from<br />
the list or as bunch using “Find All”<br />
button. After selection of diagnose from<br />
list (2) to be processed its name is<br />
automatically excluded from list (2) and<br />
displayed in section (3). In the current<br />
example the selected diagnose “захарен<br />
диабет тип 1” (Diabetes mellitus type 1)<br />
is displayed in sections (3). The system<br />
identifies possible ICD10 codes<br />
assignments and displays them in list (4) –<br />
E10 Инсулинозависим захарен диабет<br />
(Insulindependent diabetes mellitus)<br />
<br />
<br />
<br />
<br />
<br />
Fig. 1 1 aignn<br />
The data for processed diagnoses from list<br />
(2) are displayed in list (5) for further<br />
storage in CVS format text file.<br />
It is possible the system to identify more<br />
than one possible codes for assignment, in<br />
this case different options are displayed in<br />
list (4) in decreasing order of ranking. The<br />
most appropriate association is ranked<br />
first. There are two options – automatic<br />
assignment of ICD10 code and manual<br />
assignment. If there are more than one<br />
options listed in (4), user can choose by<br />
checkbox the more appropriate code. In<br />
automatic mode the first ICD10 code<br />
with highest rank is associated.<br />
<br />
<br />
<br />
<br />
<br />
Fig. 15 Ranking<br />
In diagnose “овариална поликистоза”<br />
(Polycystic ovarian syndrome) for Latin<br />
term “овариална” (ovarian) (яйчници –<br />
in ulgarian) in 3 chars categories ICD10<br />
are recognized N83, Q50, C56, D27, E28.<br />
We exclude C56 and D27 branches,<br />
because none of “neoplasm”, “carcinoma<br />
in situ” and “melanoma” terms were<br />
presented. For further specification we<br />
search in 4 characters categories for the<br />
second term “поликистоза” (Polycystic)<br />
and we set highest rank to E28.2, which<br />
contains it, for the other codes we<br />
recognize Q50.3, N83.2 and E28.9 that<br />
refer to unspecified disorders.<br />
Workshop <strong>Patient</strong> <strong>Safety</strong> through Intelligent Procedures in Medication, Sofia, Bulgaria, 2011<br />
31
Fig. 16 Latin terminology processing<br />
In this example the diagnose “струма<br />
нодоза гр 1б/еутироидес”( Struma Nodosa<br />
Euthyroides in Latin) (Euthyroid nodular<br />
goiter in English) is presented using latin<br />
terminology with transliteration. The first<br />
term “струма” corresponds to “гуша”<br />
(goiter) in Bulgarian language. In ICD10<br />
3 characters categories it corresponds to<br />
the cluster E00E0.<br />
<br />
<br />
<br />
<br />
<br />
<br />
<br />
Fig. 17 Latin terminology processing<br />
In this example the diagnose<br />
“феохромицитома” (pheochromocytoma)<br />
is presented using latin terminology with<br />
transliteration. This term corresponds to<br />
“Доброкачествено новообразувание на<br />
надбъбречна жлеза” (neoplasm of<br />
Adrenal gland) in Bulgarian language. In<br />
ICD10 chars categories it corresponds<br />
to D35.0. The next diagnose “киста<br />
оварии декстра” (киста на яйчника – in<br />
Bulgarian, cyst of ovary – in English) () is<br />
processed similarly to example on Fig. 15<br />
with assigned code N83.0 Фоликуларна<br />
киста на яйчника (N83.0 Follicular cyst<br />
of ovary). “декстра” in Latin means (дясна<br />
– in Bulgarian, Right – in English) is not<br />
considered in classification in this case<br />
Fig. 18 Diagnoses automatic analyses<br />
In automatic mode can be processed all<br />
PRs stored in the selected folder. The<br />
result file can be generated in CSV<br />
format, XML format or Excel format.<br />
Extracted information contains PR ID,<br />
diagnose name from the narrative text in<br />
PR, ICD10 code, and diagnose name<br />
according to ICD10 classification.<br />
32<br />
Workshop <strong>Patient</strong> <strong>Safety</strong> through Intelligent Procedures in Medication, Sofia, Bulgaria, 2011
4. Evaluation Results and Discussion<br />
In NLP the performance accuracy of text extraction<br />
procedures usually is measured by the precision<br />
(percentage of correctly extracted entities as a subset<br />
of all extracted entities), recall (percentage correctly<br />
extracted entities as a subset of all entities available in<br />
the corpus) and their harmonic mean<br />
easre: F=2*Precision*Recall/(Precision+Recall).<br />
The experiments were made with a training corpus<br />
containing 1,300 PRs and the evaluation results are<br />
obtained using a test corpus, containing 6,200 PRs. In<br />
the test corpus there are 5,859 PRs with prescribed<br />
drugs during the hospitalization. The remaining 341<br />
PRs concern patients hospitalized for clinical<br />
examinations only; these 341 PRs are excluded from<br />
the evaluation.<br />
Evaluation results (Table 1 and Table 2) shows<br />
high percentage of success in drug name, drug dosage<br />
and diagnoses recognition in PRs texts.<br />
Table 1 Extraction sensitivity according to the IE<br />
performance measures<br />
Precision Recall core<br />
Drug ame 97.28% 99.59% 98.42%<br />
Dose 92.25% 95.51% 93.85%<br />
Diagnoses 97.3 % 74.68* 84.5%<br />
Table 2 Extraction sensitivity according to the IE<br />
performance measures<br />
Drug ame<br />
Extracted entities from the PRs text<br />
16 82<br />
Diagnoses 26 826<br />
The major reasons for incorrect recognition of<br />
medications are: misspelling errors, unrecognized<br />
drug events for allergies, incorrect detection of<br />
negation scope, drug events occurrence in other<br />
context and etc.<br />
The incorrect assignment of ICD10 codes for<br />
diagnoses is mainly due to: misspelling errors,<br />
unrecognized abbreviations, incorrect transliteration<br />
of Latin terminology and description of specific<br />
pathological states which is hard to classify according<br />
to ICD10 even for humans.<br />
Comparing with systems performing similar task<br />
in the PSIP project like French MultiTerminology<br />
Indexer (FMTI) [9], which indexes documentation in<br />
several health terminologies, we note that despite all<br />
complications in processing USHATE discharge<br />
letters the performance of our system is satisfactory.<br />
FMTI is applied for automatic detection of Adverse<br />
Drug Events in discharge letters. The extraction of<br />
ATC codes from the free text of French discharge<br />
letters is performed with fmeasure 88% when<br />
compared to the manual extraction; however,<br />
compared to the CPE content, the fmeasure is 49%.<br />
The main reason for better performance of our<br />
systems is that the discharge letters in French seem to<br />
have no predefined structure, which is available in<br />
Bulgaria that significantly helps to recognize events.<br />
During the Third i2b2 Shared Task and Workshop<br />
“Challenges in Natural Language Processing for<br />
Clinical Data: Medication Extraction Challenge” [10]<br />
several semi and unsupervised systems for medical<br />
information extraction were presented. They report<br />
result for machine learning base algorithms [11] high<br />
fmeasure (91.40% for medication and 94.91% for<br />
dosage). Statistical hybrid methods that combine<br />
machine learning and rulebased modules [12] report<br />
fmeasure for medication 89.9% and for dosage<br />
93.6%. The other approaches for medication<br />
information recognition are mainly based on<br />
techniues like information extraction [13], rulebased<br />
[14], event driven [15] and semantic mining [16] and<br />
report similar performance.<br />
For the second task assignment of ICD10 codes<br />
to diagnoses the results are comparable with recent<br />
systems as MIDAS (Medical Diagnosis Assistant) [17],<br />
SynDiATe [18] with about 76% Fmeasure based on<br />
combination between text parsing and semantic<br />
information derivation from a Bayesian network, the<br />
system reported in [19] uses a hybrid approach<br />
combining examplebased classification and a simple<br />
but robust classification algorithm (naive Bayes) with<br />
high performance over 22 millions PRs: fmeasure<br />
98,2%; for about 48% of the medical records at Mayo<br />
clinic, another 34% of the records are classified with<br />
fmeasure 93,1%, and the remaining 18% of the<br />
records are classified with fmeasure of 58,5%. The<br />
article [20] compares three machine learning methods<br />
on radiological reports and points out that the best f<br />
measure is 77%.<br />
Workshop <strong>Patient</strong> <strong>Safety</strong> through Intelligent Procedures in Medication, Sofia, Bulgaria, 2011<br />
33
5. Conclusion and Further Work<br />
The paper presents software modules for PSIP+<br />
project which supports the automatic extraction of<br />
medication information and diagnoses from PR texts<br />
The Semantic mining modules are strictly oriented<br />
to Bulgarian language. The plans for their further<br />
development and application are connected primarily<br />
to Bulgarian local context. Future enhancements are<br />
planned for extension of the name and dosage<br />
recognition rules, to cope with certain specific<br />
exceptions and section filtering rules. The preliminary<br />
correction of spell errors and other kinds of typos will<br />
also increase the IE accuracy.<br />
For diagnoses recognition task we plan<br />
improvement of rules for more precise code<br />
assignments.<br />
6. Acknowledgements<br />
The research tasks leading to these results have<br />
received funding from the European Community’s<br />
Seventh Framework Programme (FP7/20072013)<br />
under grant agreement n° 216130 PSIP (<strong>Patient</strong> <strong>Safety</strong><br />
through Intelligent Procedures in Medication).<br />
7. References<br />
[1] Boytcheva, S., D. Tcharaktchiev and G. Angelova.<br />
Contenxtualisation in Automatic Extraction of Drugs from Hospital<br />
<strong>Patient</strong> Records. To appear in the Proc. of MIE2011, the 23th Int.<br />
Conf. of the European Federation for Medical Informatics, Norway,<br />
2831 August 2011, published by IOS Press.<br />
[2] National Framework Contract between the National Health<br />
Insurance Fund, the Bulgarian Medical Association and the<br />
Bulgarian Dental Association, Official State Gazette<br />
№106/30.12.2005, updates №68/22.08.2006 and №101/15.12.2006,<br />
Sofia, Bulgaria, http://dv.parliament.bg/.<br />
[3] Boytcheva, S. Shallow Medication Extraction from Hospital <strong>Patient</strong><br />
Records. Stud. Health Technol. Inform. 2011, 166: pp. 260269: pp.<br />
119128.<br />
[4] ATC drugs classification, http://atc.thedrugsinfo.com/, last visited<br />
on 01/06/2011<br />
[5] Bulgarian Drug Agency, Drug List at<br />
http://www.bda.bg/images/stories/documents/register/Mp.htm<br />
[6] National Center of Health Information,<br />
http://www.nchi.government.bg/download.html<br />
[7] Boytcheva, S. Assignment of ICD10 Codes to Diagnoses in<br />
Hospital <strong>Patient</strong> Records in Bulgarian. In: Alfred, R., G. Angelova<br />
and H. Pfeiffer (Eds.). Proc. of the Int. Workshop “Extraction of<br />
Structured Information from Texts in the Biomedical Domain”<br />
(ESITBioMed 2010), ICCS2010, Malaysia, Published by MIMOS<br />
BERHAD, ISBN 9789834137137, July 2010, pp. 5666.<br />
[8] DemnerFushman, D., W. Chapman and C. McDonald. What can<br />
natural language processing do for clinical decision support? Journal<br />
of Biomedical Informatics, Volume 42, Issue 5, October 2009,<br />
(2009), 760772.<br />
[9] Merlin B., E. Chazard, S. Pereira, E. Serrot, S. Sakji, R. Beuscart,<br />
and S. Darmoni. Can FMTI semanticmined drug codes be used for<br />
Adverse Drug Events detection when no CPOE is available? In<br />
Studies in Health Technology and Informatics, Proceedings of the<br />
13th World Congress on Medical Informatics, Cape Town, South<br />
Africa, olume 160, Number pt 1, (2010), 10251029.<br />
[10] Third i2b2 SharedTask and Workshop Challenges in Natural<br />
Language Processing for Clinical Data: Medication Extraction<br />
Challenge”, https://www.i2b2.org/NLP/Medication/.<br />
[11] Patrick, J. and M. Li. A Cascade Approach to Extracting Medication<br />
Events. In: Proc. Australian Language Technology Workshop<br />
(ALTA), (2009), 99103.<br />
[12] Halgrim, S., F. Xia, I. Solti, E. Cadag, and Ö. Uzuner, Extracting<br />
medication information from discharge summaries, In Louhi '10<br />
Proceedings of the AACL HLT 2010 Second Louhi Workshop on<br />
Text and Data Mining of Health Documents, (2010), 6167.<br />
[13] Meystre, S. M., G. K. Savova, K. C. KipperSchuler, and J. F.<br />
Hurdle, Extracting Information from Textual Documents in the<br />
Electronic Health Record: A Review of Recent Research, IMIA<br />
Yearbook 2008: Access to Health Information, ol. 1 (2008), 128<br />
144.<br />
[14] Evans, D. A., N. D. Brownlowt, W. R. Hersh, and E. M. Campbell.<br />
Automating Concept Identification in the Electronic Medical<br />
Record: An Experiment in Extracting Dosage Information. AMIA<br />
1996 Symposium Proceedings, (1996), 388392.<br />
[15] Miura, Y., E. Aramaki, T. Ohkuma, M. Tonoike, D. Sugihara, H.<br />
Masuichi, and K. Ohe, AdverseEffect Relations Extraction from<br />
Massive Clinical Records, Proceedings of the Second Workshop on<br />
LP Challenges in the Information Explosion Era (LPIX 2010),<br />
Coling 2010 Beijing, 2010, 7583.<br />
[16] Chazard E., C. Preda et al., Delierable D2.3 Results of data <br />
semantic mining, PSIP Project, URL: https://www.psip<br />
project.eu/2010<br />
[17] SotelsekMargalef, A. and J. illenaRomn. MIDAS: An<br />
InformationExtraction Approach to Medical Text Classification<br />
(MIDAS: Un enfoque de extracción de información para la<br />
clasificación de texto médico), Procesamiento del lenguaje Natural<br />
n. 41, 2008, pp. 97104.<br />
[18] Hahn, U., K. Schnattinger and K. Markó. Wissensbasiertes Text<br />
Mining mit SynDiKATe (Knowledge based text mining with<br />
SynDiKATe). Künstliche Intelligenz, vol. 2, 2002.<br />
[19] Pakhomov, S., J. Buntrock and C. G. Chute. Automating the<br />
assignment of diagnosis codes to patient encounters, Journal of<br />
American Medical Informatics Association, 13, 2006, pp. 51652.<br />
[20] Coffman, A. and N. Wharton. Clinical Natural Language Processing:<br />
AutoAssigning ICD 9 Codes. Overview of the Computational<br />
Medicine Center’s 2007 Medical Natural Language Processing<br />
Challenge. Available online at<br />
http://courses.ischool.berkeley.edu/i256/f09/Final%20Projects%20w<br />
riteups/coffmanwhartonprojectfinal.pdf<br />
34<br />
Workshop <strong>Patient</strong> <strong>Safety</strong> through Intelligent Procedures in Medication, Sofia, Bulgaria, 2011
Development of experimental repository for Bulgarian patients and a testbed for<br />
integration of PSIP CDSS modules with the USHATE hospital information system<br />
Dimitar TCHARAKTCHIEV<br />
University Specialised Hospital for Active Treatment of Endocrinology (USHATE) “Acad. Ivan<br />
Pentchev”, Medical Univeristy – Sofia, Bulgaria, dimitardt@gmail.com<br />
Abstract<br />
This paper presents the set up of an experimental repository<br />
for patients treated in the University Specialised Hospital<br />
for Active Treatment of Endocrinology (USHATE). The<br />
Bulgarian repository is aligned with the structure of the<br />
Global repository developed in the PSIP project. The<br />
testbed for integration of PSIP Clinical Decision Support<br />
System (CDSS) modules with the research version of<br />
USHATE’s Hospital Information System is also discussed.<br />
Keywords: CDSS integration, HIS, EHR, EMR, CPOE.<br />
1. Introduction<br />
Huge amounts of data are available in Hospital<br />
Information Systems (HIS), Laboratory Information<br />
Management Systems (LIMS) and Computer Provider<br />
Order Entry or Computerized Physician Order Entry<br />
(CPOE) systems [1]. The Hospital Information<br />
System of USHATE designed in 1987 includes<br />
several modules: registration of demographic and<br />
medico-administrative data (diagnoses, procedures,<br />
clinical pathways, Diagnosis-Related Groups (DRGs),<br />
data concerning the admission, transfer and<br />
discharge), structured vital signs and clinical findings,<br />
management of the laboratory results ordering and<br />
reporting, radiology, pharmacy, dietology, clinical<br />
documentation management etc. The discharge<br />
summaries are also recorded in the HIS. The patient is<br />
the centre of the integration of all clinical and<br />
administrative data. Documentation and messages<br />
conforming to the rules of UN/EDIFACT are created.<br />
The standard EN/ISO EN13606 on Electronic<br />
Healthcare Record (EHR) communication and<br />
archetype paradigm are implemented [2]. Roger’s<br />
definition for Minimum Basic Data Set (MBDS) – the<br />
core of information with most commonly available set<br />
of items and most extensive range of usages – is<br />
adopted [2].<br />
Our laboratory modules provide different quality<br />
assurance functions during the data entry – the results<br />
being audited as they are being placed into the<br />
databases. The user can be alerted to any unusual<br />
results and take the proper corrective action before the<br />
result leave the bench. The results provided on-line by<br />
the laboratory analyzers undergo the same checks as<br />
manually entered data. The quality assurance include<br />
several futures: zones of analyses of lab data (normal,<br />
pathologic, intermediate), life threatening value<br />
checks (produce signals, alerts and questions<br />
concerning the acceptance of the result), analyses of<br />
all laboratory data in accordance with age, sex,<br />
method, patient condition, influence of some drugs or<br />
tests, routine statistical analysis of lab data and<br />
control results. This complex structure of the<br />
USHATE HIS provides a solid base for its inclusion<br />
in the PSIP project.<br />
2. Development of an experimental<br />
repository for Bulgarian patients<br />
The experimental repository for patients treated in the<br />
University Specialised Hospital for Active Treatment<br />
of Endocrinology is aligned with the structure of the<br />
Global PSIP repository, where are stored more than<br />
90 000 complete Electronic Medical Records (EMR),<br />
extracted from all 6 participating hospitals (from<br />
France, Denmark and Bulgaria) [1].<br />
Simplified representation of the PSIP data schema is<br />
presented at the figure 1.<br />
The extracted information includes demographic data<br />
(age and sex), data describing the hospital stay – dates<br />
of the admission, transfer and discharge, the length of<br />
stay, diagnoses coded using ICD-10, Acts (ICD-9 CM<br />
codes), Diagnosis Related Groups (DRGs), Major<br />
Diagnostic Categories (MDCs), patient medication<br />
codes using ATC classification, laboratory tests, and<br />
text files of anamnesis, patient status and discharge<br />
letters. All the data are anonymised before the transfer<br />
to the scientific databases.<br />
Most of the values are automatically transferred from<br />
the USHATE Hospital Information System (HIS) to<br />
the scientific repository. A significant amount of data<br />
Workshop <strong>Patient</strong> <strong>Safety</strong> through Intelligent Procedures in Medication, Sofia, Bulgaria, 2011<br />
35
Figure 1. The PSIP data schema of the experimental repository<br />
is added by three extractors processing a free text of<br />
discharge letters and providing more information for<br />
diagnoses, lab tests and medication [3]. The<br />
importance of the automatic text analysis for accurate<br />
filling of experimental repository can be demonstrated<br />
by the following examples: the discharge letter texts<br />
contain on average 5,26 drugs administrated per<br />
patient, while in the Hospitalization <strong>Patient</strong> Record<br />
are registered only 1,9 drugs prescribed via the<br />
hospital pharmacy; for 6 200 hospital patient records<br />
including discharge letters the extractor of ICD-10<br />
diagnoses finds in the texts 22 667 diagnoses, while<br />
there are only 9321 diagnoses registered in the<br />
USHATE Hospital Information system.<br />
The stored data in the scientific repository in<br />
USHATE are integrated in the Global PSIP<br />
Repository through middleware services.<br />
3. Testbed for CDSS integration<br />
The Testbed for integration of PSIP Clinical Decision<br />
Support System (CDSS) modules with the research<br />
version of USHATE Hospital Information System<br />
gives the possibility for real-time presentation of the<br />
generated alerts to the medical practitioners. It alerts<br />
the medical practitioners about ADE’s during the<br />
prescription process and whenever a patient file is<br />
open. Manual ADE’s checks are also possible [4]. On<br />
figure 2 are presented the relations between some<br />
modules of the USHATE HIS (research version),<br />
Drug Information Extraction module (developed by<br />
the IICT – BAS) and PSIP ADEs Web service. The<br />
connection of Bulgarian prototype to the CDSS<br />
Knowledge base is realised using PSIP Connectivity<br />
Platform and is shown on figure 3. The applied<br />
messages are based on Simple Object Access Protocol<br />
and eXtensible Markup Language over Hypertext<br />
Transfer Protocol (SOAP and XML over HTTP).<br />
The process of integration of alerts in the prescription<br />
process is presented on figure 4.<br />
4. Results and Discussion<br />
The complex HIS structure, including all the data<br />
defined in the PSIP “Common Data Model” [1], gives<br />
the possibility of USHATE to provide to the Global<br />
PSIP repository 6800 anonymised Electronic Medical<br />
Records. The natural language processing (NLP)<br />
considerably enrich recorded in the experimental<br />
repository medical data. This enables PSIP partners to<br />
36<br />
Workshop <strong>Patient</strong> <strong>Safety</strong> through Intelligent Procedures in Medication, Sofia, Bulgaria, 2011
Figure 2. UML representation of the relations between some modules of the USHATE HIS (research version),<br />
Drug Information Extraction Module (developed by the IICT – BAS) and PSIP ADEs Web service<br />
Figure 3. The connection of Bulgarian prototype to<br />
the CDSS Knowledge base<br />
adequately calculate the frequency of ADEs in<br />
USHATE and to generate contextualised rules for<br />
issuing alerts to the medical practitioners.<br />
The physicians expressed their positive opinion<br />
concerning the alerts generated by the PSIP CDSS<br />
Knowledge Base and presented by the Bulgarian<br />
prototype [5]. No additional input is needed as the<br />
system operates on data which are automatically<br />
collected. The doctors support the further<br />
development of Bulgarian prototype.<br />
5. Conclusion<br />
The repository compliant with PSIP model is<br />
developed and set up in USHATE and includes<br />
clinical data about Bulgarian patients. The developed<br />
testbed enables to validate and assess the PSIP<br />
approach in Bulgarian clinical setting.<br />
The evaluation results show that the extraction<br />
accuracy of the applied NLP techniques is similar to<br />
the state-of-the-art achievements of leading research<br />
groups in the field [3]. The functionalities of<br />
Bulgarian prototype are comparable with other<br />
integrated test environments developed by innovative<br />
high-tech companies as Medasys and IBM [6].<br />
6. Acknowledgements<br />
The research tasks leading to these results have<br />
received funding from the European Community’s<br />
Seventh Framework Programme (FP7/2007-2013)<br />
under grant agreement n° 216130 PSIP (<strong>Patient</strong> <strong>Safety</strong><br />
through Intelligent Procedures in Medication).<br />
Workshop <strong>Patient</strong> <strong>Safety</strong> through Intelligent Procedures in Medication, Sofia, Bulgaria, 2011<br />
37
Figure 4. The integration of generated alerts in the prescription process<br />
The USHATE scientists are grateful to all PSIP experts<br />
who contributed for the smooth integration of the Bulgarian<br />
achievements into the elaborated PSIP framework and<br />
especially to the teams of Régis Beuscart, Nicos<br />
Maglaveras, Jean-François Forget, Elske Ammenwerth and<br />
Galia Angelova.<br />
7. References<br />
[1] Beuscart R. PSIP: An Overview of the Results and<br />
Clinical Implications. In Koutkias V. et al. (Eds),<br />
<strong>Patient</strong> <strong>Safety</strong> Informatics, Stud. Health Technol.<br />
Inform. 2011 Vol. 166, IOS Press, 2011, pp. 3-12.<br />
[2] Tcharaktchev D., A. Dimitrov, G. Dimov et al.<br />
MEDICA – 9 years of development and use of a<br />
Clinical Information System in the University Hospital<br />
of Endocrinology and Gerontology – Sofia. In J.<br />
Brender et al. (Eds), Medical Informatics Europe ’96,<br />
IOS Press, 1996, pp. 458–462.<br />
[3] Tcharaktchiev, D., G. Angelova, S. Boytcheva, Z.<br />
Angelov, and S. Zacharieva. Completion of Structured<br />
<strong>Patient</strong> Descriptions by Semantic Mining. In Koutkias<br />
V. et al. (Eds), <strong>Patient</strong> <strong>Safety</strong> Informatics, Stud. Health<br />
Technol. Inform. 2011 Vol. 166, 2011, pp. 260-269.<br />
[4] Dimitrov H. Testbed for integration of CDSS modules<br />
and PSIP validation in University Specialized Hospital<br />
for Active Treatment of Endocrinology, Medical<br />
University – Sofia. In R. Beuscart, D. Tcharaktchiev,<br />
G. Angelova (Eds) Proc. Int. Workshop: <strong>Patient</strong> <strong>Safety</strong><br />
through Intelligent Procedures in Medication,<br />
INCOMA, 2011, pp. 39-48.<br />
[5] Nechkova-Atanassova, K. Survey of the attitudes of the<br />
physicians at USHATE hospital towards innovations.<br />
In R. Beuscart, D. Tcharaktchiev, G. Angelova (Eds)<br />
Proc. Int. Workshop: <strong>Patient</strong> <strong>Safety</strong> through Intelligent<br />
Procedures in Medication, INCOMA, 2011, pp. 53-56.<br />
[6] Bernonville S., J. Nies, H. Pedersen et al. Three<br />
Different Cases of Exploiting Decision Support<br />
Services for Adverse Dreg Event Prevention. In<br />
Koutkias V. et al. (Eds), <strong>Patient</strong> <strong>Safety</strong> Informatics,<br />
Stud. Health Technol. Inform. 2011 Vol. 166, IOS<br />
Press, 2011, pp. 180-188.<br />
38<br />
Workshop <strong>Patient</strong> <strong>Safety</strong> through Intelligent Procedures in Medication, Sofia, Bulgaria, 2011
Testbed for integration of CDSS modules and PSIP validation in University<br />
Specialized Hospital for Active Treatment of Endocrinology, Medical University -<br />
Sofia<br />
Hristo DIMITROV<br />
University Specialised Hospital for Active Treatment of Endocrinology (USHATE) “Acad. Iv. Pentchev”<br />
Medical Univeristy – Sofia, Bulgaria<br />
h_dimitrov@yahoo.com<br />
Abstract<br />
The paper presents the development of a testbed for<br />
integration of Clinical Decision Support System<br />
(CDSS) modules with the research version of<br />
USHATE Hospital Information System (HIS). These<br />
CDSS modules enable the validation of the PSIP<br />
approach in USHATE. The interface which introduces<br />
the CDSS alerts to the physicians in the validation<br />
process is described.<br />
Keywords<br />
CDSS integration, PSIP validation<br />
1. Introduction<br />
The HIS of the USHATE is organized in modules.<br />
The first modules, operational in the end of 1987,<br />
represent the core of necessary functions and make<br />
possible registration of administrative data, clinical<br />
data, management of the laboratory results ordering<br />
and reporting. Later pharmacy module was set up [1].<br />
The new version of the system from 2007 is aligned<br />
with the CEN/ISO EN13606 EHR standard and<br />
includes the application of published archetypes [2].<br />
The system presented here was developed and applied<br />
in the PSIP project for the preparation of a testbed in<br />
USHATE to support PSIP results validation.<br />
Our aim is to demonstrate the possibility to integrate<br />
PSIP CDSS module in the process of drug<br />
prescription with Computerized Physician Order<br />
Entry (CPOE) subsystem, part of the USHATE HIS.<br />
In this way we can obtain an improvement of the<br />
quality of clinical activity and patient safety.<br />
2. CDSS modules integration<br />
Initially the USHATE experts have considered three<br />
different ways for integration of PSIP CDSS modules:<br />
• Real time Adverse Drug Events (ADE’s)<br />
checks during the prescription process - the<br />
major benefit of this method of integration is that<br />
it allows the physicians to be aware of the events<br />
that may occur even before the actual<br />
prescription.<br />
• Automatic and Manual ADE’s checks -<br />
automatic checks are performed in the USHATE<br />
HIS whenever a patient file is opened. The<br />
background check allows the physicians to<br />
continue their work without having to wait for<br />
the response of the PSIP CDSS module. The<br />
automatic check is done only once when a<br />
patient file is opened. After that additional<br />
checks may be performed but they are manual.<br />
• Backend process executing ADE’s checks - the<br />
idea of this integration is to embed ADE’s<br />
checks into USHATE drug prescription module.<br />
Different events will trigger ADE’s checks and<br />
they will be done in background, independent<br />
from the user interaction with the system. This<br />
approach requires availability of additional<br />
information in the USHATE drug prescription<br />
module including diagnoses, acts, laboratory test<br />
results, etc. In this way the CDSS ADE’s alerts<br />
can be saved into the system and only updated<br />
when new events occurs (but this required major<br />
changes in the CPOE - HIS architecture, to<br />
Workshop <strong>Patient</strong> <strong>Safety</strong> through Intelligent Procedures in Medication, Sofia, Bulgaria, 2011<br />
39
provide the necessary additional information to<br />
the drug prescription module).<br />
USHATE implemented two of the presented models<br />
of integration – during the drugs prescription process<br />
and in the moment of patient file opening. Manual<br />
check is also possible. Requests are done in parallel<br />
with normal execution so they do not disturb the user<br />
interaction with the system.<br />
All the examples below illustrate the USHATE testbed environment. No real patients are shown. The<br />
presented cases are simulated and are only used for the purpose of PSIP integration and testbed validation.<br />
The full demonstration video can be watched at: http://www.medicalnet-bg.org/psip/<br />
After we have successfully logged in into USHATE HIS, we can see the list of hospitalized patients which are<br />
currently in the hospital using the button marked as “4” on Fig. 1. The window that appears reflects the structure<br />
of the USHATE clinics with their rooms and beds. The line (rectangle) marked as “1” shows the “Pituitary<br />
clinic” that we are browsing at the moment and the rectangle marked as “2” shows the room “01” and its beds.<br />
All this information is represented as a tree view. On Fig. 1 we have selected the patient marked as “3” which is<br />
currently situated on bed 3 in room 01 in clinic 003 (Pituitary pathology). After double clicking the patient we<br />
can open his file and see the details of the hospitalization.<br />
Figure 1. Presentation of hospitalized patients in the clinics of USHATE<br />
40<br />
Workshop <strong>Patient</strong> <strong>Safety</strong> through Intelligent Procedures in Medication, Sofia, Bulgaria, 2011
Fig. 2 shows patient clinical details during the current stay. When the window appears it automatically<br />
triggers an ADE’s check based on the available information till the current moment. If some alerts are found,<br />
yellow exclamation mark appears in the upper right corner of the window. We can see it marked as “6” on Fig. 2.<br />
The medical professional can click on the icon and review the alerts. Since the window on Fig. 2 is the main<br />
interface for patient data manipulation, it contains other vital information as well:<br />
• <strong>Patient</strong> names, gender and age (in Rectangle “1”);<br />
• Principal diagnosis for the current stay coded in ICD 10 (in Rectangle “2”);<br />
• Diagnostic and treatment plan which is used to enter different diagnoses (preliminary, complications,<br />
comorbidities and differential) during the stay (in Rectangle “3”);<br />
• Laboratory tests and their results (in Rectangle “4”);<br />
• Performed acts coded in ICD 9 (in Rectangle “5”);<br />
• All the paper documents required during the stay like: status, anamnesis, discussion, discharge letter,<br />
etc. (in Rectangle “7”);<br />
• A dropdown menu where the PSIP components (scorecards, web prototypes) are integrated and the<br />
drugs prescription is performed (in Rectangle “8”).<br />
Figure 2. <strong>Patient</strong> details during the current stay in the hospital, with links to the integrated PSIP modules<br />
Workshop <strong>Patient</strong> <strong>Safety</strong> through Intelligent Procedures in Medication, Sofia, Bulgaria, 2011<br />
41
If we click on the icon marked as “6” on Fig. 2, the window shown on Fig. 3 appears. The screen contains<br />
found alerts and their descriptions. Rectangle “1” there points to the current medication including drug ATC<br />
codes, markers denoting the first day of application (0, 0, 2) and drug names in Bulgarian. Rectangle “2”<br />
represents a short description of the alerts which is oriented to the physician. Rectangle “3” presents the full<br />
information about the alerts including:<br />
• the short description for the physician,<br />
• drug that causes the event,<br />
• a long description for the physician,<br />
• confidence value,<br />
• source of the alert and<br />
This information is passed from the PSIP CDSS web service to the testbed in USHATE. The received<br />
information is formatted in HTML, thus it can be easily represented in different forms. For every alert a row<br />
appears in the zone marked as “2” and full information concerning the alert is presented in the zone marked as<br />
“3”.<br />
The ADE found in the present example states: “Кортикостероидите могат да увеличат панкресните<br />
ензими (Corticosteroids can increase the enzymes of pancreas). This event is triggered by Дехидрокортизон<br />
(Dehydrocortison) drug”.<br />
Figure 3. Presentation of automatically triggered ADE’s checks and found alerts<br />
42<br />
Workshop <strong>Patient</strong> <strong>Safety</strong> through Intelligent Procedures in Medication, Sofia, Bulgaria, 2011
Fig. 4 shows the integration of PSIP services in the USHATE hospital information system. Four different<br />
services are included in the dropdown menu described below:<br />
• An interface providing manual check for ADE’s based on the available information regarding the<br />
patient till the current moment (Rectangle “1” on Fig. 4);<br />
• The PSIP Scorecards link directly points to the PSIP web site including the contextualized<br />
information for participated hospitals and their departments (Rectangle “2”);<br />
• The PSIP web prototype link in Bulgarian language directly points to the web site for drug check for<br />
professional usage (Rectangle “3”);<br />
• The PSIP Prototypes portal link directly points to the web site presenting the PSIP major prototypes<br />
including the PSIP <strong>Patient</strong> Component (Rectangle “4”).<br />
Figure 4. Integration of PSIP service in the USHATE Hospital Information System<br />
Workshop <strong>Patient</strong> <strong>Safety</strong> through Intelligent Procedures in Medication, Sofia, Bulgaria, 2011<br />
43
To fully describe the patient drug treatment, Fig. 5 presents the process of automatic extraction of drugs from<br />
the free text of the anamnesis. Rectangle “1” points to the part of the anamnesis that states the current treatment:<br />
Терапия при постъпването: Дехидрокортизон 5 мг – 1 т/дневно; Л-Тироксин – 75 мкг/дневно;<br />
(Therapy in the moment of hospitalization: Dehydrocortison 5 mg – 1 tablet per day; L-Thyroxin – 75 mg per<br />
day).<br />
Drug extraction is performed automatically when the anamnesis is modified and saved or manually when the<br />
menu marked by “5” is clicked. When the medication extraction [3] is activated, it opens a new window if it finds<br />
drugs that have been already prescribed in ambulatory care. This window contains a table with rows for each<br />
found drug. The information that is presented within the table on Fig. 5 is: drug ATC code, drug name in<br />
Bulgarian, dosage of the application, units and the context in which the drug was found in the text. In our example<br />
two drugs were successfully extracted and marked with rectangle “2”. The first column of the table contains a<br />
checkbox which can be used to discard the drug and not to apply it further during the hospitalization. Button “3”<br />
is used to validate and apply the selected drugs; button “4” is used to discard the whole operation.<br />
Figure 5. Automatic extraction of drug therapy at the moment of hospitalization from the free text of the<br />
anamnesis<br />
44<br />
Workshop <strong>Patient</strong> <strong>Safety</strong> through Intelligent Procedures in Medication, Sofia, Bulgaria, 2011
It is also possible to perform ADE check when new drugs are prescribed. This is shown on Fig. 6. The<br />
process starts with clicking on button “1”: Изписване на медикамент (New drug prescription). After that a new<br />
window appears where the drug for the prescription has to be selected. In our example Ацетизал (Acetysal) is<br />
chosen. When the user successfully selects a drug, an automatic ADE check is performed. If some alerts are found<br />
the same yellow exclamation mark (rectangle “6”) appears in the upper right corner. Again the user can click on<br />
that icon to review the alerts. Other fields marked as “2”, “3”, “4”, “5”, “7”, and “8” are used to complete the<br />
prescription process:<br />
• Start date and hour of the application (Rectangle “2”);<br />
• Form of the application based on predefined schemas (in this case 2 times daily, Rectangle “3”);<br />
• Textual description of the admission schema parameters (in this case one tablet 2 times daily in 08:00<br />
and 20:00, Rectangle “4”);<br />
• Number of days and last hour of the reception (the number of the days is defined in the schema and<br />
only visualized on Fig. 6 as a reference, Rectangle “5”);<br />
• Button “Apply the prescription” (Rectangle “7”);<br />
• Button “Discard the prescription” (Rectangle “8”).<br />
Figure 6. ADE check during the drugs prescription process<br />
Workshop <strong>Patient</strong> <strong>Safety</strong> through Intelligent Procedures in Medication, Sofia, Bulgaria, 2011<br />
45
Fig. 7 shows the alerts found during the prescription shown on Fig. 6. As we can see in our example<br />
(Rectangle “1”) new alerts have appeared due to the prescription of Ацетизал (Acetysal). Rectangle “2” shows<br />
the short descriptions of the new ADE alerts and the full description is in the HTML component below.<br />
Figure 7. Detected ADEs during the prescription of Acetysal<br />
46<br />
Workshop <strong>Patient</strong> <strong>Safety</strong> through Intelligent Procedures in Medication, Sofia, Bulgaria, 2011
Fig. 8 shows the alerts generated by the PSIP Web Prototype in Bulgarian when the same clinical and prescription data<br />
as of a patient presented on fig. 5 and fig. 6 are used.<br />
Figure 8. Detected ADEs during the prescription of Acetysal presented in PSIP Web Prototype in Bulgarian<br />
Workshop <strong>Patient</strong> <strong>Safety</strong> through Intelligent Procedures in Medication, Sofia, Bulgaria, 2011<br />
47
3. Evaluation Results and Discussion<br />
The Testbed for integration of CDSS modules<br />
implemented in USHATE gives the possibility to alert<br />
the medical experts about ADE’s during the<br />
prescription process and whenever a patient file is<br />
open. Manual ADE’s checks are also possible. These<br />
functionalities are comparable with other integrated<br />
test environments (Medasys Prototype, IBM<br />
Prototype and PSIP Web based application) [4].<br />
In USHATE, like as other participating hospitals in<br />
the PSIP Project, the medical professionals are in<br />
favor of automatic alerts, part of a CPOE system [5].<br />
It is important to mention that the physicians working<br />
at the USHATE are open to innovations concerning<br />
the patient safety [6].<br />
4. Conclusion and Further Work<br />
Successful integration was performed and CDSS<br />
module was introduced to be used by the physicians<br />
in USHATE.<br />
In some cases physicians continue the drug therapy<br />
regarding the CDSS alerts. Because they cannot<br />
discard some of them (stop showing the alerts), alerts<br />
triggered by new drug prescription can become<br />
indistinguishable from the old ones. Eventually this<br />
will lead to over alertness and physicians will<br />
decrease their attention to the subsequent warning<br />
messages. To avoid this USHATE drug prescription<br />
module should locally store CDSS alerts and allow<br />
the physicians hide some of them (stop showing<br />
them). This approach requires CDSS module<br />
integration as a backend process as discussed in<br />
section 2 of the current report.<br />
5. Acknowledgements<br />
The research tasks leading to these results have<br />
received funding from the European Community’s<br />
Seventh Framework Programme (FP7/2007-2013)<br />
under grant agreement n° 216130 - The PSIP project<br />
(<strong>Patient</strong> <strong>Safety</strong> through Intelligent Procedures in<br />
Medication).<br />
6. References<br />
[1] Tcharaktchev D., A. Dimitrov, G. Dimov et al.<br />
MEDICA – 9 years of development and use of a<br />
Clinical Information System in the University Hospital<br />
of Endocrinology and Gerontology – Sofia. In J.<br />
Brender et al. (Eds), Medical Informatics Europe ’96,<br />
IOS Press, 1996, pp. 458–462.<br />
[2] Tcharaktchev D. Development of experimental<br />
repository for Bulgarian patients and a testbed for<br />
integration of PSIP CDSS modules within the USHATE<br />
hospital information system. In R. Beuscart, D.<br />
Tcharaktchiev, G. Angelova (Eds). Proc. International<br />
Workshop: <strong>Patient</strong> <strong>Safety</strong> through Intelligent<br />
Procedures in Medication (PSIP), INCOMA,<br />
Shoumen, 2011, pp. 35-38.<br />
[3] Boytcheva S. Shallow Medication Extraction from<br />
Hospital <strong>Patient</strong> Records. In Koutkias V. et al. (Eds)<br />
<strong>Patient</strong> <strong>Safety</strong> Informatics, IOS Press, 2011, pp. 119-<br />
128.<br />
[4] Bernonville S., J. Nies, H. Pedersen et al. Three<br />
Different Cases of Exploiting Decision Support<br />
Services for Adverse Dreg Event Prevention. In<br />
Koutkias V. et al. (Eds) <strong>Patient</strong> <strong>Safety</strong> Informatics,<br />
IOS Press, 2011, pp. 180-188.<br />
[5] Ammenwerth E. Expectations and Barriers versus<br />
cxCDSS-CPOE: A European User Survey.<br />
International. In R. Beuscart, D. Tcharaktchiev, G.<br />
Angelova (Eds) Proc. International Workshop: <strong>Patient</strong><br />
<strong>Safety</strong> through Intelligent Procedures in Medication<br />
(PSIP), INCOMA, Shoumen, 2011, pp. 49-52.<br />
[6] Nechkova-Atanassova K. Survey of the attitudes of the<br />
physicians at USHATE hospital towards innovations.<br />
In R. Beuscart, D. Tcharaktchiev, G. Angelova (Eds)<br />
Proc. International Workshop: <strong>Patient</strong> <strong>Safety</strong> through<br />
Intelligent Procedures in Medication (PSIP),<br />
INCOMA, Shoumen, 2011, pp. 53-56.<br />
48<br />
Workshop <strong>Patient</strong> <strong>Safety</strong> through Intelligent Procedures in Medication, Sofia, Bulgaria, 2011
Expectations and Barriers versus cxCDSS-CPOE:<br />
A European User Survey<br />
Elske AMMENWERTH 1 , Martin JUNG 2<br />
1<br />
University for Health Sciences, Medical Informatics and Technology,<br />
EWZ 1, 6060 Hall in Tirol, Austria, elske.ammenwerth@umit.at<br />
2<br />
University for Health Sciences, Medical Informatics and Technology,<br />
EWZ 1, 6060 Hall in Tirol, Austria, martin.jung@umit.at<br />
Abstract<br />
User acceptance can be crucial for the success of a health<br />
IT project. We analysed the attitudes of physicians in<br />
various European hospitals with regard to automatic<br />
alerting as part of CPOE systems. Overall, 275 physicians<br />
from five countries participated. Most surveyed doctors are<br />
in favour of automatic alerts as part of a CPOE system, but<br />
fear alert overload. Consequently, they prefer an adaption<br />
of alert presentation with regard to the clinical context. The<br />
level of CPOE implementation in the hospitals seems to<br />
play a minor role in the attitude of the physicians.<br />
Keywords<br />
CPOE, automatic alerting, user survey, attitudes,<br />
quantitative study<br />
1. Introduction<br />
The introduction of CPOE systems, and especially the<br />
introduction of automatic alerts within CPOE<br />
systems, is a huge organizational change, with impact<br />
on the clinical workflow and patient safety [1-2]. User<br />
acceptance is often crucial for the adoption of CPOE<br />
systems and for the success of the overall<br />
implementation project [3-4].<br />
The objective of this study is to determine the attitude<br />
of physicians towards decision-support within CPOE<br />
systems in the PSIP partner hospitals.<br />
2. Related Work<br />
<strong>Patient</strong> safety is a serious public health issue [5-6].<br />
The Council of Europe defines patient safety as “the<br />
identification, analysis and management of patientrelated<br />
risks and incidents, in order to make patient<br />
care safer and minimize harm to patients” [7] (p.8).<br />
Estimates from the World Health Organization<br />
(WHO) show that one out of ten patients is harmed<br />
while receiving hospital care [6].<br />
Medication-related events are among the most<br />
common adverse events [8]. These Adverse Drug<br />
Events (ADE) are defined as “any injury occurring<br />
during the patient’s drug therapy and resulting either<br />
from appropriate care or from unsuitable or<br />
suboptimal care” [7, p. 1]. ADEs resulting from a<br />
medication error are considered to be preventable<br />
ADEs [7, 9]. A review article by von Laue [10] in<br />
2003 found ADE rates to be ranging from 0.7% to<br />
6.5% per admission of hospitalized patients.<br />
The use of computerized physician order entry<br />
(CPOE) systems can reduce medication errors and<br />
ADEs [11-12]. CPOE systems can be equipped with<br />
different levels of clinical decision support [13].<br />
However, drug safety alerts generated by CPOE<br />
systems often show low specificity due to too many<br />
false-positive warnings [14].<br />
Overriding drug safety alerts in CPOE systems is very<br />
common and occurs in 49–96% of cases [15].<br />
Constant over-alerting may cause alert fatigue. Alert<br />
fatigue is described “as the mental state that is the<br />
result of too many alerts consuming time and mental<br />
energy, which can cause important alerts to be<br />
ignored along with clinically unimportant ones”<br />
[15, p. 139].<br />
Several studies have provided insight into the<br />
attitudes of prescribers towards drug safety alerts in<br />
CPOE systems [4, 16-17], but nothing is known about<br />
the acceptance of CPOE and automatic alerting in the<br />
PSIP partner hosiptals.<br />
Workshop <strong>Patient</strong> <strong>Safety</strong> through Intelligent Procedures in Medication, Sofia, Bulgaria, 2011<br />
49
3. Methods<br />
In 2010, a one-group cross-sectional quantitative<br />
questionnaire survey was performed in the following<br />
hospitals:<br />
• Three Hospitals of Region Hovedstaden,<br />
Danemark<br />
• Hospital of Denain, France<br />
• University Hospital of Novara, Italy<br />
• University Hospital of Rouen, France<br />
• University Specialized Hospital of Active<br />
Treatment of Endocrinology Sofia, Bulgaria<br />
• Academic Medical Centre of Amsterdam, the<br />
Netherlands<br />
All participating hospitals, with the exception of<br />
Rouen, already had a CPOE system in use, each with<br />
a different level of decision-support.<br />
The standardized questionnaire was developed based<br />
on the published questionnaires by Magnus [17], Hor<br />
[4] and Ko [16], using a four-point Likert scale. In<br />
addition, two open questions were added: What would<br />
you expect as the biggest possible benefits of<br />
automatic alerting? And what as the biggest possible<br />
problems?<br />
The questionnaires were analyzed with descriptive<br />
statistics. Furthermore, acceptance scores were<br />
calculated based on a factor analysis. The open<br />
questions were analyzed by means of a quantitative<br />
content analysis according to Mayring [18] .<br />
4. Results<br />
Overall, 275 physicians completed the questionnaire.<br />
Table 1 shows the distribution of returned<br />
questionnaires.<br />
Hospital<br />
Distributed<br />
questionnaires<br />
Returned<br />
valid<br />
questionnaires<br />
Return<br />
rate<br />
Amsterdam 998 78 7.8%<br />
Copenhagen 207 94 45.4%<br />
Denain 60 26 43.3%<br />
Novara 8 5 62.5%<br />
Rouen 100 41 41%<br />
Sofia 53 31 58.5%<br />
Overall 1,426 275 19.3%<br />
Table 1. Participation and return rate of the survey.<br />
The overall acceptance of automatic alerts was<br />
medium to high in all hospitals. The median<br />
acceptance scores, calculated by the sum of 8<br />
questions, was between around 24 and around 26 in<br />
all hospitals (the acceptance score could range from 8<br />
= minimum to 32 = maximum acceptance).<br />
Fig. 1 – Fig. 4. present selected results of individual<br />
questions. They show that most physicians support<br />
the idea of automatic alerting, and do not feel limited<br />
in their prescribing freedom. However, over half of<br />
the participants feel that there may be too many alerts.<br />
Consequently, the majority says that alerts should be<br />
filtered according to the clinical context.<br />
In the free text comments, the problem of alert<br />
overload was also mentioned by nearly half of the<br />
participants.<br />
Figure 1. Answers to the question: “I find automatic<br />
alerts a useful tool in electronic prescribing” ( “agree”<br />
and “partly agree” aggregated).<br />
Figure 2. Answers to the question: “I believe that<br />
CPOE systems limit my freedom in prescribing” (<br />
“agree” and “partly agree” aggregated).<br />
50<br />
Workshop <strong>Patient</strong> <strong>Safety</strong> through Intelligent Procedures in Medication, Sofia, Bulgaria, 2011
Figure 3. Answers to the question: “These CPOE<br />
systems generate too many alerts that are irrelevant<br />
for the patient” ( “agree” and “partly agree”<br />
aggregated).<br />
Figure 4. Answers to the question: “Alerts should be<br />
filtered according to the context of the clinical<br />
situation” ( “agree” and “partly agree” aggregated).<br />
5. Discussion<br />
The majority of physicians from all study hospitals<br />
was in favour of automatic alerting as part of CPOE<br />
systems. Interestingly, the overall opinions did not<br />
differ much between the hospitals, disregarding the<br />
differences in organaization, used technology and<br />
clinical workflow. This can be seen as an indication<br />
that the basic attitudes are quite independent from the<br />
organizational and technical context.<br />
The survey was carried out in six hospitals within the<br />
European Union; therefore the results reflect an<br />
international focus. We included hospitals with<br />
different availablity of CPOE and automatic alerting.<br />
The translation of the questionnaire was done by nonprofessional<br />
translators that were familiar with the<br />
field.<br />
6. Conclusion<br />
Most surveyed doctors are in favour of automatic<br />
alerts as part of a CPOE system, but fear alert<br />
overload. Consequently, they prefer a differentiation<br />
according to the importance of the alerts and an<br />
adaption to the clinical context. The level of CPOE<br />
implementation in the hospitals seems to play a minor<br />
role in the attitude of the physicians<br />
In all hospitals, high acceptance scores were found.<br />
This supports the idea of introducing further ways of<br />
decision-support as part of the prescription process.<br />
No major difference in the acceptances scores<br />
regarding countries or level of CPOE implementation<br />
could be spotted, which we see as an indication for a<br />
good transferability of the PSIP solutions to all<br />
participating hospitals. However, the possible<br />
problem of alert overload is a serious issue that should<br />
be addressed.<br />
7. Acknowledgements<br />
The research tasks leading to these results have<br />
received funding from the European Community’s<br />
Seventh Framework Programme (FP7/2007-2013)<br />
under grant agreement n° 216130 PSIP (<strong>Patient</strong> <strong>Safety</strong><br />
through Intelligent Procedures in Medication).<br />
We thank all colleagues in the participating hospitals<br />
that supported this survey, namely: Jens Barholdy,<br />
Carlo Cacciabue, Mauro Campanini, Stefan Darmoni,<br />
Jean Doucet, Laurie Ferret, Giuseppina Ferrotti,<br />
Monique Jaspers, Maurice Langemeijer, Kitta<br />
Lawton, Philippe Lecocq, Michel Luycks, Philippe<br />
Massari, Krassimira Neshkova, Flemming Steen<br />
Nielsen, Lars Nygård, Costanza Riccioli, Dimitar<br />
Tcharaktchiev, Jesper Vilandt<br />
8. References<br />
[1] Ash J.S., Sittig D.F., Seshadri V., Dykstra R.H., Carpenter<br />
J.D., Stavri P.Z. Adding insight: a qualitative cross-site study of<br />
physician order entry. Int J Med Inform. 2005 Aug;74(7-8):623-8.<br />
[2] Aarts J., Van Der Sijs H. CPOE, Alerts and Workflow:<br />
Taking Stock of Ten Years Research at Erasmus MC. Stud Health<br />
Technol Inform. 2009;148:165-9.<br />
[3] Campbell E.M., Sittig D.F., Ash J.S., Guappone K.P.,<br />
Dykstra R.H. Types of unintended consequences related to<br />
computerized provider order entry. J Am Med Inform Assoc.<br />
2006 Sep-Oct;13(5):547-56.<br />
[4] Hor C.P., O'Donnell J.M., Murphy A.W., O'Brien T.,<br />
Kropmans T.J. General practitioners' attitudes and preparedness<br />
towards Clinical Decision Support in e-Prescribing (CDS-eP)<br />
Workshop <strong>Patient</strong> <strong>Safety</strong> through Intelligent Procedures in Medication, Sofia, Bulgaria, 2011<br />
51
adoption in the West of Ireland: a cross sectional study. BMC<br />
Med Inform Decis Mak. 2010;10:2.<br />
[5] Byers J., White S., editors. <strong>Patient</strong> safety: principles and<br />
practice. New York: Springer; 2004.<br />
[6] WHO. 10 facts on patient safety. Geneva: World Health<br />
Organization;; 2010; Available from:<br />
http://www.who.int/features/factfiles/patient_safety/patient_safety<br />
_facts/en/index.html.<br />
[7] Council of Europe. Committee of Experts on Management of<br />
<strong>Safety</strong> and Quality in Health Care (SP-SQS) - Expert Group on<br />
Safe Medication Practices: Glossary of terms related to patient<br />
and medication safety. . . 2005; Available from:<br />
http://www.who.int/patientsafety/highlights/COE_patient_and_me<br />
dication_safety_gl.pdf. Last accessed: April 1st, 2008.<br />
[8] EC. Communication from the Commission to the European<br />
Parliament and the Council on patient safety, including the<br />
prevention and control of healthcare-associated infections.<br />
Brussels: European Commission; 2008; Available from:<br />
http://ec.europa.eu/health/ph_systems/docs/patient_com2008_en.p<br />
df.<br />
[9] Institute of Medicine, editor. Committee on Identifying and<br />
Preventing Medication Errors. Aspden P, Wolcott JA, Bootman<br />
JL, Cronenwett LR, editors. Preventing medication errors.<br />
Washington, DC: National Academies Press; 2007.<br />
[10] von Laue N.C., Schwappach D.L., Koeck C.M. The<br />
epidemiology of preventable adverse drug events: a review of the<br />
literature. Wien Klin Wochenschr. 2003 Jul 15;115(12):407-15.<br />
[11] Ammenwerth E., Schnell-Inderst P., Machan C., Siebert U.<br />
The Effect of Electronic Prescribing on Medication Errors and<br />
Adverse Drug Events: A Systematic Review J Am Med Inform<br />
Assoc. 2008;15(5):585-600.<br />
[12] Hug B.L., Witkowski D.J., Sox C.M., Keohane C.A., Seger<br />
D.L., Yoon C., et al. Adverse drug event rates in six community<br />
hospitals and the potential impact of computerized physician<br />
order entry for prevention. J Gen Intern Med. 2010 Jan;25(1):31-<br />
8.<br />
[13] Kuperman G., Bobb A., Payne T., Avery A., Gandhi T.,<br />
Burns G., et al. Medication-related clinical decision support in<br />
computerized provider order entry systems: a review. J Am Med<br />
Inform Assoc 2007;14(1):29-40.<br />
[14] Khajouei R., Jaspers M.W. The impact of CPOE medication<br />
systems' design aspects on usability, workflow and medication<br />
orders: a systematic review. Methods Inf Med. 2010;49(1):3-19.<br />
[15] van der Sijs H., Aarts J., Vulto A., Berg M. Overriding of<br />
drug safety alerts in computerized physician order entry. J Am<br />
Med Inform Assoc. 2006 Mar-Apr;13(2):138-47.<br />
[16] Ko Y., Abarca J., Malone D.C., Dare D.C., Geraets D.,<br />
Houranieh A., et al. Practitioners' views on computerized drugdrug<br />
interaction alerts in the VA system. J Am Med Inform Assoc.<br />
2007 Jan-Feb;14(1):56-64.<br />
[17] Magnus D., Rodgers S., Avery A.J. GPs' views on<br />
computerized drug interaction alerts: questionnaire survey. J Clin<br />
Pharm Ther. 2002 Oct;27(5):377-82.<br />
[18] Mayring M. Qualitative Inhaltsanalyse: Grundlagen und<br />
Techniken: Utb; 2007.<br />
52<br />
Workshop <strong>Patient</strong> <strong>Safety</strong> through Intelligent Procedures in Medication, Sofia, Bulgaria, 2011
Survey of the attitudes of the physicians at the USHATE hospital<br />
towards innovations<br />
Krassimira NECHKOVA-ATANASSOVA 1 , 2<br />
1<br />
University Specialized Hospital for Active Treatment of Endocrinology “Acad. Ivan Pentchev” (USHATE),<br />
Medical University Sofia, 2 Zdrave Str., Sofia 1431, Bulgaria,<br />
2<br />
Institute for Population and Human Studies, Department of Psychology, Bulgarian Academy of Sciences,<br />
Block 6, Acad. Georgi Bontchev Str., Sofia 1113, Bulgaria<br />
e-mail: krassimiran@yahoo.es<br />
Abstract<br />
This paper presents a survey of the physicians’ attitudes at<br />
the University Specialized Hospital for Active Treatment of<br />
Endocrinology (USHATE), Medical University - Sofia<br />
towards the introduction of automatic medication safety<br />
alerts. The great majority of the participants was in favor of<br />
the PSIP Web prototype in Bulgarian, detecting several<br />
Adverse Drug Events (ADEs) concerning the prescribed<br />
medicaments. Integrating the Clinical Decision Support<br />
System (CDSS) alerts in the process of drug prescription<br />
using the Computerized Physician Order Entry (CPOE)<br />
subsystem, part of the USHATE HIS, the extra time for<br />
data registering is not needed and almost all medical<br />
professionals expressed their satisfaction and support.<br />
Keywords<br />
Innovations, attitudes, automatic medication safety alerts<br />
1. Introduction<br />
One of the goals of the project “<strong>Patient</strong> <strong>Safety</strong> through<br />
Intelligent Procedures in Medication” /PSIP/ is to<br />
introduce a CDSS for automatic alerts concerning the<br />
adverse drug events, based on the observation of large<br />
electronic patient information archives. Because this<br />
is an innovation, one of the tasks of our team was to<br />
investigate the attitudes of the medical staff towards<br />
innovations.<br />
Attitudes are a major factor in introducing and<br />
accepting innovations. The attitude is a relatively<br />
stable habitual internal stance, or the person’s<br />
predisposition to react in a particular way as a<br />
precondition of action and experience. In general, it is<br />
a mediator between the stimulus and the reaction. It is<br />
the way of being in a particular situation.<br />
The Concise Oxford Dictionary of Current English [1]<br />
defines innovation as “introducing novelty” and<br />
“making changes”. In the literature there are three<br />
main paradigms of interpreting innovations as defined<br />
by Linn and Zaltman in their concise review of the<br />
existing theoretical approaches towards defining<br />
innovations [2]:<br />
1. The term “innovation” is used as a synonym<br />
of invention, of constructing something new<br />
through a creative process;<br />
2. Innovation is a process by which an already<br />
existing creation is accepted by the person<br />
and becomes a part of her/his knowledge and<br />
behavior;<br />
3. Innovation is viewed as a description of<br />
something material or ideal which has been<br />
invented and is considered new independent<br />
from its acceptance or non-acceptance and its<br />
dissemination.<br />
One of the most influential theories for studying<br />
innovations is the theory of cognitive styles of M.<br />
Kirton [3, 4]. It focuses on the strategies preferred by<br />
individuals in situations of change. This theory is a<br />
basis for a well-known empirical method. According<br />
to Kirton innovative processes start at an individual<br />
level but often turn to changes on the level of the<br />
social system. Innovation as a personal characteristic<br />
should be considered on a continuum between<br />
adaptation and innovation; accordingly, individuals<br />
can be regarded as adaptors and innovators. Between<br />
the two poles – the highly adaptive and the highly<br />
innovative are situated the different cognitive styles in<br />
posing and solving problems and the creativity styles.<br />
Adaptors prefer the already established methods of<br />
problem solving, and the innovators – the radically<br />
new ones.<br />
According to empirical studies described in the<br />
literature there are representatives of the both<br />
personality types in every organization. This means<br />
Workshop <strong>Patient</strong> <strong>Safety</strong> through Intelligent Procedures in Medication, Sofia, Bulgaria, 2011<br />
53
that one part of the group readily and with a great<br />
interest accepts innovations and starts using them as<br />
soon as possible, whereas the other part is slow in<br />
accepting changes and needs more time to adapt to<br />
innovations.<br />
The goal of our study was to investigate the attitudes<br />
and the perceptions of the medical staff at the<br />
USHATE before and after the introduction of the new<br />
system for automatic alerts in prescribing.<br />
2. Method of Study<br />
We used two questionnaires provided by the PSIP<br />
partners E. Ammenswerth and her colleagues from<br />
UMIT (University of Health Sciences, Medical<br />
Informatics and Technology in Hall, Tirol, Austria)<br />
[5]. Part of the survey was conducted by Martin Jung<br />
and Werner Hackl from UMIT. The survey conducted<br />
at the USHATE was divided into two phases. The<br />
first phase began in January and ended in March, and<br />
the second phase began in May and ended in the<br />
beginning of June. The goal of the phase one was to<br />
study the physicians’ attitudes towards the automatic<br />
medication safety alerts and their opinions towards<br />
systems facilitating the decision making process. The<br />
second part was conducted when the physicians<br />
already had some experience with the new system.<br />
The preparation of the surveys in USHATE included<br />
the following activities:<br />
• translation and adaptation of the questionnaires,<br />
• approbation of the questionnaires,<br />
• defining the target groups,<br />
• conducting the research, and<br />
• processing and analyzing the results.<br />
We decided to include in the target groups all the<br />
physicians working at the hospital clinics and<br />
laboratories. The entire number of the USHATE<br />
personnel is 169, including 53 physicians. All the<br />
physicians were invited to fill in the questionnaires.<br />
So in the first survey the target group consisted of 53<br />
physicians. Some 33 responded; one part of the<br />
remaining physicians was on leave, and another part<br />
refused to collaborate due to a heavy workload. From<br />
the 33 questionnaires received 31 were filled out<br />
properly and could be processed; the remaining two<br />
were invalid. The second target group consisted of 19<br />
physicians who already had worked with the system.<br />
To obtain our data we applied two questionnaires. The<br />
administration of the second questionnaire was<br />
accompanied by an interview with the participant.<br />
The survey proceeded in the following way: the<br />
psychologist handed the questionnaire to each of the<br />
participants in person, the participant gave answers to<br />
questions arising, and then the questionnaires were<br />
collected.<br />
For analysis of the results descriptive statistic<br />
methods were used.<br />
3. Results and Discussion<br />
On Figure 1 the results of the first item are shown. It<br />
was the statement “I find automatic alerts a useful<br />
tool in prescribing”. Some 67,7% of the participants<br />
responded that they agreed with the implementation<br />
of an automatic medication safety system; 29% were<br />
hesitating but rather agree, and 3,2 % disagree. This<br />
results correspond to previous research showing on<br />
innovators (those who would accept the innovations<br />
immediately) as well as adaptors (who prefer the<br />
traditional way of working) (see Section 1, also [4]).<br />
Figure 1. Assessment of the usefulness of the automatic<br />
alerts<br />
The regression analysis was applied to find out to<br />
what degree the preferences about the implementation<br />
of the automatic medication safety alert system were<br />
influenced by gender. The results are shown at the bar<br />
chart on Figure 2.<br />
95,2 % of the women are innovation oriented – 76,2%<br />
agree that the automatic alerts are a useful tool for<br />
prescribing, 19,0% partly agree; however 4,8% were<br />
against. 100 % of the males are innovation oriented -<br />
50% of the males agree, and exactly the same<br />
percentage 50% partly agrees.<br />
54<br />
Workshop <strong>Patient</strong> <strong>Safety</strong> through Intelligent Procedures in Medication, Sofia, Bulgaria, 2011
90<br />
80<br />
70<br />
60<br />
50<br />
40<br />
30<br />
20<br />
10<br />
0<br />
I find automatic alerts a useful tool in prescribing<br />
partly disagree partly agree agree<br />
Figure 2. Gender influence on preferences<br />
Male<br />
Female<br />
The descriptive statistics (see bar chart on Figure 3)<br />
shows the impact of age on answering the question<br />
about the usefulness of the ADE-scorecards as a tool<br />
to learn about ADEs occurring at the respective<br />
department. The “ADE Scorecards” is the application<br />
designed to describe the statistical results and to<br />
improve the awareness of the healthcare professionals<br />
on the ADEs occurring in their medical unit or their<br />
hospital [6].<br />
column 3 are presented the partly agreeing medical<br />
practitioners and in column 2 – the partly disagreeing<br />
practitioners.<br />
Altogether the preliminary expectations and the<br />
attitudes of the physicians at the USHATE hospital<br />
towards introducing new automatic safety alert<br />
systems are highly positive. The percentage of<br />
innovators is significantly higher then the percentage<br />
of adaptors.<br />
Some 94,7% of the participants agree with the<br />
statement „I would recommend using ADE-scorecards<br />
to my colleagues.”, and only 5,3% agree only partially<br />
and show reservations towards the system. This<br />
shows that also the physicians who have reservations<br />
are in favor of the new system (see Figure 4).<br />
Figure 4. Positive attitudes towards novel solutions<br />
Figure 3. Impact of age on the preferences<br />
It is obvious that the physicians with a longer<br />
professional experience and age tend to value the<br />
introduction of the ADE-scorecards higher than their<br />
younger colleagues. In the column 4 are presented the<br />
medical professionals agreeing that the ADEscorecards<br />
are a useful tool to learn about ADEs, in<br />
This questionnaire contains also two open questions.<br />
The first question: “What are the biggest benefits of<br />
ADE scorecards” was answered by 57,9 % of the<br />
participants. The biggest benefits are the easy access<br />
to information and reducing the side effects, followed<br />
by better safety and security for the patients and the<br />
improvement of clinical work. 42,1% of the<br />
participants have not answered this question (Figure<br />
5).<br />
The second open question was “What are the biggest<br />
problems of ADE scorecards?” The problem indicated<br />
most often is the “limited scope”, i.e. the limited<br />
number of cases at hand to gather the necessary<br />
information.<br />
Workshop <strong>Patient</strong> <strong>Safety</strong> through Intelligent Procedures in Medication, Sofia, Bulgaria, 2011<br />
55
Figure 5. Assessment of score cards benefits<br />
Concerning the use of PSIP Web prototype some<br />
doctors recommended to reduce “the amount of time<br />
necessary to put in the information”.<br />
These results are shown in Table 1.<br />
Valid<br />
Missing<br />
The limited<br />
scope<br />
The<br />
amount of<br />
time<br />
necessary<br />
to put in<br />
the data<br />
Frequ-<br />
ency<br />
Valid<br />
Percent<br />
Percent<br />
Cumulative<br />
%<br />
2 10,5 50,0 50,0<br />
2 10,5 50,0 100,0<br />
Total 4 21,1 100,0<br />
15 78,9<br />
Total 19 100,0<br />
Table 1: Problems mentioned by the medical practitioners.<br />
4. Conclusions<br />
The results obtained from this survey testify in a<br />
convincing way that the physicians at the USHATE<br />
hospital find the automatic safety alert system useful<br />
and would use it in their daily work.<br />
The benefits expected from implementing the system<br />
are:<br />
• more information concerning the adverse drug<br />
events of the prescribed therapy,<br />
• reducing the errors due to inadvertence,<br />
• quicker access to information, and<br />
• improved patient safety.<br />
The physicians who already use the system have made<br />
the following recommendations: first, to increase the<br />
number of cases for analysis, and secondly, to reduce<br />
the amount of time necessary to upload the data in the<br />
PSIP Web prototype. Integrating the CDSS alerts in<br />
the process of drug prescription using the CPOE<br />
system, the extra time for data handling is not needed<br />
and the medical professionals were satisfied.<br />
The results show that the physicians working at the<br />
USHATE hospital are open to innovations concerning<br />
the patient safety. In our view this is a great advantage<br />
that could be used for the benefit of the patients.<br />
5. Acknowledgements<br />
The research tasks leading to these results have<br />
received funding from the European Community’s<br />
Seventh Framework Programme (FP7/2007-2013)<br />
under grant agreement n°216130 PSIP project (<strong>Patient</strong><br />
<strong>Safety</strong> through Intelligent Procedures in Medication).<br />
We thank all medical professionals from the<br />
USHATE, participating in this survey, and Martin<br />
Jung and Werner Hackl from UMIT conducted part of<br />
the interviews, and all the team of E. Ammenwerth<br />
(UMIT) providing the standardized questionnaire.<br />
6. References<br />
[1] The Concise Oxford Dictionary of Current English,<br />
1965. Fifth Edition.<br />
[2] Lin N. and Zaltman G., 1973. Dimensions of<br />
innovations. – In: Processes and Phenomena of<br />
SocialChange. New York: Wiley&Sons.<br />
[3] Kirton M., 1976. Adaptors and Innovators: A<br />
description and Measure. Journal of Applied<br />
Psychology, Vol. 61, 5, pp. 622-629.<br />
[4] Kirton M., 1978. Have adaptors and innovators equal<br />
levels of creativity? Psychological Reports, 42, pp.<br />
695-698.<br />
[5] Ammenwerth, E. and M. Jung. Expectations and<br />
Barriers versus cxCDSS-CPOE: A European User<br />
Survey. In R. Beuscart, D. Tcharaktchiev, G. Angelova<br />
(Eds) Proc. International Workshop: <strong>Patient</strong> <strong>Safety</strong><br />
through Intelligent Procedures in Medication (PSIP),<br />
INCOMA, Shoumen, 2011, pp. 49-52.<br />
[6] Beuscart R. PSIP: An Overview of the Results and<br />
Clinical Implications. In Koutkias V. et al. (Eds)<br />
<strong>Patient</strong> <strong>Safety</strong> Informatics, Amsterdam, IOS Press,<br />
2011, pp. 3-12.<br />
56<br />
Workshop <strong>Patient</strong> <strong>Safety</strong> through Intelligent Procedures in Medication, Sofia, Bulgaria, 2011