Swiss Medical Informatics - SGMI
Swiss Medical Informatics - SGMI
Swiss Medical Informatics - SGMI
You also want an ePaper? Increase the reach of your titles
YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.
Typical recommendationsare useful to prescribe themost<br />
appropriate antibiotic, according to different parameters<br />
such aspatient situation, clinical assessment, but also<br />
costs, benefits, adverse effects, and the risk of resistance<br />
development. As acase study, we started investigating<br />
guidelines for geriatrics from the University Hospitals of<br />
Geneva (HUG). Totransform such verbose documents into<br />
machine-readable data, the guidelines were transformed<br />
intodatabase tuples. The translation from French<br />
to English was performed manually, assisted by Frenchto-English<br />
translation tools (e.g. http://eagl.unige.ch/<br />
EAGLm/), and aSNOMEDcategoriser [10].For some of the<br />
queries several answers were possible –three on average<br />
–asshown in table 1, where severe diverticulitis caused<br />
by enterobacteriaceae can be treated by three different antibiotics:<br />
ceftriaxone, metronidazole and piperacillintazobactam;<br />
each of them is unambiguously associated<br />
with aunique terminological identifier.<br />
Further, two search engines corresponding to different<br />
search models were tested: easyIR (a relevance-driven<br />
searchengine well known for outperforming other search<br />
methods on MEDLINE searchtasks [11])and PubMed (the<br />
NCBI’s Boolean search instrument). In addition, different<br />
combinations of the outputs of the two engines (PubMed<br />
and easyIR) were tested to combine the strengths of both<br />
engines.<br />
All targets werenormalised using standardterminologies.<br />
Three terminologies were tested to normalise the antibiotic<br />
type of target: alist of 70 SNOMED CT terms, corresponding<br />
chiefly to the available antibiotics at the University<br />
Hospitals of Geneva, alist of MeSH terms, including<br />
synonymous terms, corresponding to the UMLS Semantic<br />
Type T195, and alist of 70 MeSH terms, including synonymous<br />
terms, corresponding mostly to the available antibiotics<br />
at the University Hospitals of Geneva, associated<br />
with WHO-ATC identifiers. The disease type of target was<br />
normalised using alist of MeSH terms corresponding to<br />
disease, corresponding tothe following UMLS Semantic<br />
Types T020, T190, T049, T019, T047, T050, T033, T037,<br />
T047, T191, T046 and T184. The pathogen type of target<br />
was normalised using asubsetofthe NEWTtaxonomy corresponding<br />
to the bacterial taxonomy.<br />
Furthermore,tofine-tune thequestion-answering module,<br />
several descriptors, in particular generic ones, needed<br />
to be removed. Thus, infectious diseases or cross-infection<br />
were removed from the descriptor list for the disease<br />
type of target. Finally, specific keywords were used tore-<br />
PROCEEDINGS ANNUAL MEETING 2009<br />
fine the search equation. Thus, we added context-specific<br />
descriptors such as geriatrics, elderly, etc. The impact of<br />
general keywords such asrecommended antibiotic, antibiotherapy,<br />
etc. was also tested.<br />
Results and discussion<br />
The evaluation of our results isdone with TrecEval, a<br />
program developed to evaluate TREC (Text Retrieval Conferences)<br />
results using NIST(US NationalInstitute of Standards<br />
and Technologies) evaluation procedures. Fine-tuning<br />
of the engine was based on the TREC Genomics<br />
competitions: see e.g.[11]. As usual in information retrieval<br />
[12] and factoid question-answering tasks, we focus on precision-oriented<br />
metrics. In particular, the so-called precision<br />
at recall 0, i.e. the precision of the top-returned answer,isused<br />
to evaluate the effectiveness of our approach.<br />
To complement this metrics, which provides the precision<br />
of the system used without user interaction, we also measure<br />
the recall of the system achieved by the top five answers.<br />
Thus, we try to estimate how useful such asystem<br />
would be when used by an expert able to validate the guideline<br />
generator’s ranked output.<br />
The 64 triplets generated manually are used as the gold<br />
standard. Each rule/query concerns a specific disease<br />
caused by aspecific pathogen and isrepresented by atuple<br />
of four columns(table 1). Diseases, pathogens and antibiotics<br />
wereentered for each entry.Optionally, conditions<br />
were also addeddepending on the entries, suchasweight<br />
or age. Evaluation is done with the focus on the precision<br />
of the first retrieved antibiotic, which corresponds to the<br />
top precision, noted P0 in the following.<br />
Three terminological targets have been tested. The best results<br />
are obtained using asubsetofthe MeSHincluding 70<br />
antibioticentities. P0 is of 0.58 for the PubMed engine and<br />
of 0.52 for the easyIR engine. Eachantibioticentity is mentioned<br />
by several terms, allowing more results to be retrieved.<br />
Thus, amoxycillinwith clavulanatepotassiumcan<br />
also be mentioned as amoxicillin-clavulanic acid or augmentin.<br />
Moreover, using alimited number of antibiotic<br />
entities avoids returninggeneral terms, suchasAnti-Bacterial<br />
Agents. Using keywords provides no significant improvement<br />
in top precision compared to the baseline system,<br />
but aslight decline in top precision.<br />
From figure 1weobserve that thetwo engines, which tend<br />
to perform very similarly on average, seem not tobehave<br />
Table 1<br />
Example of manually-generated rules: terminological identifiers are provided in parenthesis. In this example the infection can be treated by five different<br />
antibiotics. The use of ceftriaxone requires the precondition severe.<br />
Pathologies Pathogenic agents Antibiotics Other conditions<br />
Diverticulitis (D004238) Enterobacteriaceae (543) Amoxicillin-potassium clavulanate combination (D019980);<br />
ciprofloxacin (D002939);<br />
metronidazole (D008795)<br />
Diverticulitis (D004238) Enterobacteriaceae Ceftriaxone (D002443); Severe<br />
(543) metronidazole (D008795);<br />
piperacillin-tazobactam combination product (C085143)<br />
<strong>Swiss</strong> <strong>Medical</strong> <strong>Informatics</strong> 2009; n o 67<br />
37