18.08.2013 Views

Elsevier Editorial System(tm) for Hearing Research Manuscript Draft ...

Elsevier Editorial System(tm) for Hearing Research Manuscript Draft ...

Elsevier Editorial System(tm) for Hearing Research Manuscript Draft ...

SHOW MORE
SHOW LESS

You also want an ePaper? Increase the reach of your titles

YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.

<strong>Elsevier</strong> <strong>Editorial</strong> <strong>System</strong>(<strong>tm</strong>) <strong>for</strong> <strong>Hearing</strong> <strong>Research</strong><br />

<strong>Manuscript</strong> <strong>Draft</strong><br />

<strong>Manuscript</strong> Number: HEARES-D-11-00139R1<br />

Title: Semi-automatic attenuation of cochlear implant artifacts <strong>for</strong> the evaluation of late auditory<br />

evoked potentials<br />

Article Type: <strong>Research</strong> paper<br />

Keywords: cochlear implant, AEPs, N1, artifact attenuation, test-retest reliability<br />

Corresponding Author: Prof. Stefan Debener, Ph.D.<br />

Corresponding Author's Institution: University of Oldenburg<br />

First Author: Filipa C Viola, Msc<br />

Order of Authors: Filipa C Viola, Msc; Maarten De Vos, PhD; Jemma Hine, PhD; Pascale Sandmann, PhD;<br />

Stefan Bleeck, PhD; Julie Eyles; Stefan Debener, Ph.D.<br />

Abstract: Electrical artifacts caused by the cochlear implant (CI) contaminate electroencephalographic<br />

(EEG) recordings from implanted individuals and corrupt auditory evoked potential (AEPs).<br />

Independent component analysis (ICA) is efficient in attenuating the electrical CI artifact and AEPs can<br />

be successfully reconstructed. However the manual selection of CI artifact related independent<br />

components (ICs) obtained with ICA is unsatisfactory, since it contains expert-choices and is time<br />

consuming.<br />

We developed a new procedure to evaluate temporal and topographical properties of ICs and semiautomatically<br />

select those components representing electrical CI artifact. The CI Artifact Correction<br />

(CIAC) algorithm was tested on EEG data from two different studies. The first consists of published<br />

datasets from 18 CI users listening to environmental sounds. Compared to the manual IC selection<br />

per<strong>for</strong>med by an expert the sensitivity of CIAC was 91.7% and the specificity 92.3%. After CIAC-based<br />

attenuation of CI artifacts, a high correlation between age and N1-P2 peak-to-peak amplitude was<br />

observed in the AEPs, replicating previously reported findings and further confirming the algorithm's<br />

validity.<br />

In the second study AEPs in response to pure tone and white noise stimuli from 12 CI users that had<br />

also participated in the other study were evaluated. CI artifacts were attenuated based on the IC<br />

selection per<strong>for</strong>med semi-automatically by CIAC and manually by one expert. Again, a correlation<br />

between N1 amplitude and age was found. Moreover, a high test-retest reliability <strong>for</strong> AEP N1<br />

amplitudes and latencies suggested that CIAC based attenuation reliably preserves plausible individual<br />

response characteristics.<br />

We conclude that CIAC enables the objective and efficient attenuation of the CI artifact in EEG<br />

recordings, as it provided a reasonable reconstruction of individual AEPs. The systematic pattern of<br />

individual differences in N1 amplitudes and latencies observed with different stimuli at different<br />

sessions, strongly suggests that CIAC can overcome the electrical artifact problem. Thus CIAC facilitates<br />

the use of cortical AEPs as an objective measurement of auditory rehabilitation.


*Cover Letter<br />

PROF. DR. DEBENER, IPSY, FAK. V, CARL VON OSSIETZKY UNIVERSITÄT OLDENBURG · 26111 OLDENBURG<br />

Barbara Canlon<br />

Editor-in-Chief<br />

<strong>Hearing</strong> <strong>Research</strong><br />

Barbara.Canlon@ki.se<br />

Journal article re-submission<br />

Dear Dr. Canlon,<br />

Please find attached the revised version of our original manuscript HEARES-D-<br />

11-00139, entitled “Semi-automatic attenuation of cochlear implant artifacts <strong>for</strong><br />

the evaluation of late auditory evoked potentials” which we would like to submit<br />

<strong>for</strong> consideration <strong>for</strong> publication in <strong>Hearing</strong> <strong>Research</strong>. We would like to thank<br />

you <strong>for</strong> providing us with reviews of three experts in the field. We have revised<br />

the paper along the lines suggested by the reviewers and believe that the revision<br />

further improved the quality of the manuscript. Please find below a detailed<br />

point-by-point reply to the reviewers comments, including the implied additional<br />

analyses. For the reviewers’ and your convenience we have copied and<br />

numbered the original comments into the reply (in italics), along with our<br />

response. We have carefully considered all suggestions and revised the<br />

manuscript accordingly. All revisions are marked in red font in the revised<br />

version of the manuscript.<br />

We are looking <strong>for</strong>ward to your editorial decision.<br />

Sincerely,<br />

Stefan Debener, PhD<br />

AG Neuropsychologie<br />

Prof. Dr. Stefan Debener<br />

TELEFONDURCHWAHL<br />

(0441) 7 98 – 4271<br />

FAX<br />

(0441) 7 98 – 5522<br />

EMAIL<br />

stefan.debener@uni-oldenburg.de<br />

OLDENBURG, DEN<br />

24.10.2011<br />

POSTANSCHRIFT<br />

D-26111 Oldenburg<br />

PAKETANSCHRIFT<br />

Ammerländer Heerstraße 114 - 118<br />

D-26129 Oldenburg<br />

STANDORT<br />

Carl-von-Ossietzky –Str. 9 – 11<br />

W4-0-044


*Response to Reviewers<br />

Dear Dr. Canlon,<br />

Thank you very much <strong>for</strong> providing us with feedback to our original submission HEARES-D-<br />

11-00139 entitled "Automatic attenuation of cochlear implant artifacts <strong>for</strong> the evaluation of<br />

late auditory evoked potentials". Please find below a point-by-point reply to the reviewers<br />

comments. For your convenience we have copied and numbered the original comments into<br />

the reply (in italics), along with our response. We have carefully considered all suggestions<br />

and revised the manuscript accordingly. All revisions are marked in red font in the revised<br />

version of the manuscript.<br />

Reviewer #1:<br />

1) My only major concern relates to the use of the word 'automatic' in the title and throughout<br />

the text. From the authors description of the methods the 3 parameters (Thr_rv, Thr_deriv<br />

and Thr_corr) had to be manually adjusted to get adequate artifact attenuation <strong>for</strong> each of<br />

the two datasets collected. This need <strong>for</strong> manual adjus<strong>tm</strong>ent of parameters would lead me to<br />

refer to the method as semi-automatic and not automatic. This is an important point given<br />

that this algorithm only relates to the automatic selection of ICA components (not the ICA or<br />

EEG analysis). If three parameters need to be adjusted to correctly select the independent<br />

components that represent the artifact, then within this time frame a user could just as easily<br />

visually select the three most likely independent components. This point is not mentioned in<br />

the Discussion.<br />

Reply: The described method indeed relies on user input at some point and we agree that the<br />

title is there<strong>for</strong>e misleading. We changed the title to “Semi-automatic attenuation of cochlear<br />

implant artifacts <strong>for</strong> the evaluation of late auditory evoked potentials”. We also changed the<br />

occurrences of “automatic” to “semi-automatic” in throughout the text. The reviewer also<br />

points out that “within this time frame the user could just as easily visually select the three<br />

most likely independent components”. We would like to highlight that the parameters need to<br />

be adjusted just once <strong>for</strong> each study set, which includes multiple datasets. We have now<br />

clarified in the manuscript that CIAC was developed to find ICs that represent CI artifacts<br />

across CI users, as long as they were stimulated with the same auditory stimuli. Thus the<br />

thresholds are set <strong>for</strong> a study set and not <strong>for</strong> each single dataset. Furthermore, we disagree that<br />

the time needed to adjust parameters would be comparable to that required to visually inspect<br />

ICs and select the components representing the CI artifact. For instance, in the case of the first<br />

study (ESS) a total of 1080 ICs (18 datasets x 60 ICs) would need to be inspected. We<br />

there<strong>for</strong>e think that our tool brings advantages since it makes the selection procedure quicker<br />

and more objective.<br />

2) Remove apostrophe from "expert'-choices". On the 2nd paragraph change algorithms' to<br />

algorithm's.<br />

Reply: Revised.<br />

3a) 3rd paragraph 'In a further study.. ' Does this refer to the 2nd data set mentioned in the<br />

2nd paragraph. If so, then say 'In the second study'.<br />

Reply: Revised.<br />

1


3b) In the 2nd last sentence the phrase 'at different time points after implantation' suggests a<br />

longitudinal study tracking changes in N1 latency and amplitude over time. My understanding<br />

is that this is not the case. You are simply referring to the test-retest reliability. Please clarify<br />

this statement as at first glance it is rather confusing.<br />

Reply: The two experiments were not designed to be part of a longitudinal study, since it was<br />

not possible to re-test all CI users, and the task and type of stimuli changed. However the two<br />

experiments in fact provide AEPs from a sub-group of CI users that were tested in our<br />

laboratory at two different time points after implantation.<br />

4) Line 47-48. Surprisingly, <strong>for</strong> a study on CI EEG Koelsch et al 2004 make no mention what<br />

so ever of CI artifacts. They do not say 'electrical artifacts only [occur] at response latencies<br />

different from cortical AEPs', as the authors suggest. Also, the verb 'occurs' appears to be<br />

missing from the sentence.<br />

Reply: Revised.<br />

5) Line 59-60. As the authors correctly point out subtraction techniques severely limit the kind<br />

of stimuli that can be applied but to say that they have only been tested on a small population<br />

is wrong. Different variations of the subtraction technique are the most commonly used<br />

artifact cancelation technique (eg ECAP <strong>for</strong>ward masking paradigm, Brown et al 1990).<br />

Reply: When we mentioned that the subtraction technique has only been tested <strong>for</strong> a small<br />

population, we meant in the context of multi-channel EEG recordings, as reported by Friesen<br />

and Picton, 2010. We have amended the text in the manuscript to avoid confusion.<br />

6) Line 96-97. Are you saying that you only adjusted the CIAC parameters on study 1 and<br />

then used these settings <strong>for</strong> study 2? Please clarify.<br />

Reply: We apologize <strong>for</strong> the confusion. The parameters were adjusted <strong>for</strong> each study and it<br />

was not our goal to define a set of default parameters. We have provided some guidelines in<br />

the manuscript about reasonable ranges <strong>for</strong> the different thresholds based in our own<br />

experience. However we believe that the user should have the flexibility to control and adjust<br />

the parameters when necessary. We have revised the text in order to describe CIAC not as an<br />

“automatic” procedure but as a more objective procedure to identify ICs representing the CI<br />

artifact.<br />

7) Lines 114-115. ERP is not defined. Also, at other places in the text ERP seems to be used to<br />

refer to a specific independent component after ICA analysis, while its normal usage is to<br />

refer to an event related potential be<strong>for</strong>e ICA takes place. Please give a clear definition and<br />

tighten up the language and usage.<br />

Reply: We agree that it is important to make a distinction between event related potential<br />

(ERP) and the time-locked average of the activation of a certain IC (IC ERP). The text has<br />

been amended.<br />

2


8) Lines 137-138. Include a figure (or panel on an existing figure) which shows the pre and<br />

post correction wave<strong>for</strong>ms.<br />

Reply: We agree and have revised Fig. 4, which now includes the in<strong>for</strong>mation requested.<br />

9) Section 2.1. Please clarify the use of the correlation step and the Thr_cor parameter. Is this<br />

because you expect more than one IC to represent the artifact, and these ICs should be highly<br />

correlated?<br />

Reply: Indeed, our experience reveals that often the scalp maps of ICs representing the CI<br />

artifact have some degree of similarity at a single subject basis. In the figure below we show<br />

ICs representing the CI artifact from two CI users (CI_10 and CI_15), as selected by CIAC<br />

<strong>for</strong> each of the study sets (TNS and ESS, respectively). It can be seen that ICs from CI_10 are<br />

similar both within and between ICA decompositions. This is found <strong>for</strong> the majority of CI<br />

users. However there are also cases, such as CI_15, where the similarity is reduced. Note that<br />

a correlation approach as proposed in CORRMAP (Viola et al, 2009) would not provide<br />

satisfactory results across subjects and could even fail at the single subject level <strong>for</strong> some<br />

particular cases. The reasons <strong>for</strong> these differences are not well understood and to overcome<br />

this problem we developed a tool that evaluates both spatial and temporal patterns.<br />

10) Line 169. Mention that the CI subjects listened to the stimuli with their CIs on the<br />

everyday speech processing strategy, if indeed this was the case. You could also give a range<br />

3


of the stimulation rates used, which is not mentioned in Viola et al 2011.<br />

Reply: CI users were using their implant on their standard everyday settings. This in<strong>for</strong>mation<br />

has been added to the manuscript. The stimulation rates have been added to the table<br />

requested by reviewer #2 (please see #18).<br />

11) Line 192. How many repetitions <strong>for</strong> one averaged EPR?<br />

Reply: For the ESS the number of trials (M ± SD) was 153.8 ± 2.7. For the TNS was 1562.8 ±<br />

8.8. This in<strong>for</strong>mation was added to the manuscript.<br />

12) Line 218. 'had normal or corrected to normal vision'. Why is this relevant <strong>for</strong> this task?<br />

Reply: The sentence has been deleted.<br />

13) Fig. 1 - The light grey is very difficult to see. The black on the topographic maps is also<br />

difficult to see. Please also show a representative wave<strong>for</strong>m (from one or more electrodes)<br />

be<strong>for</strong>e IC analysis.<br />

Reply: Fig. 1 has been revised. Fig. 4 has been changed to include the in<strong>for</strong>mation requested.<br />

14) Fig. 3 -Please mark onset audio in all columns.<br />

Reply: Fig. 3 has been revised accordingly.<br />

Reviewer #2:<br />

15a) A concern is the expected robustness of this procedure. The authors do a great job<br />

discussing some of the limitations related to artifact identification (e.g., variance from device<br />

type, etc.). However, in some cases, these issues could benefit from more detailed discussion<br />

related to the algorithm itself. For example, would one expect to need different threshold<br />

values depending on the device type, or stimulation strategy?<br />

Reply: CIAC was developed to find ICs representing CI artifact across CI users, as long as<br />

the auditory stimulation delivered across individuals is the same. There<strong>for</strong>e we intentionally<br />

designed a tool that does not require adjusting thresholds at a single subject level. We would<br />

like to highlight that the CI users included in the study ESS used devices from different<br />

manufactures, as well as different processors and speech coding strategies. The figure below<br />

shows ICs selected by CIAC <strong>for</strong> 14 out of the 18 CI users. As can be seen, it is not possible to<br />

associate a specific pattern to a specific device or processor. Even when all parameters are<br />

identical, IC scalp maps differ. However note that all ICs have in common the fact that the<br />

residual variances are never smaller than 25%. Moreover, as shown in Fig.1, included in the<br />

manuscript, another common aspect is that the temporal derivative of the activation of these<br />

ICs contains spikes associated to onset/offset of the CI artifact. Thus we used this<br />

combination of temporal and spatial properties when we developed CIAC and we are<br />

4


confident that the same thresholds should allow good results <strong>for</strong> all datasets included in a<br />

study set.<br />

15b) Is there a limit to the number of recording electrodes needed to accurately identify the<br />

artifact?The final paragraph of the manuscript introduces the open source EEGLAB tool,<br />

with some hope that outside use will help to fine-tune and validate the procedure. In this<br />

light, it would be helpful to include a broader discussion of the issues that remain, and the<br />

types of validation that would be helpful <strong>for</strong> those that do wish to contribute. Tying all of this<br />

together would, I think, make <strong>for</strong> a much stronger case toward its use.<br />

Reply: Un<strong>for</strong>tunately we cannot predict if there is a limit number of recording electrodes<br />

needed to accurately identify the artifact since the number of EEG sources is unknown, and<br />

systematically testing this is, in our opinion, not possible with the available data (as an<br />

artificial sub-sampling to less channels would confound with the degree of equidistance and<br />

with the amount of spatial sampling of the head sphere). Our experience reveals that using 68<br />

channel recordings we obtain reliable ICA decompositions where the CI artifact can be<br />

separated from other sources. Nevertheless we agree with the reviewer and we have extended<br />

the discussion. We have highlighted the importance of further validating CIAC with other<br />

EEG montages and other types of stimuli.<br />

5


16) Of minor concern (and slightly related to the above question of robustness), is the clinical<br />

applicability of this tool? Realizing, of course, that EEGLAB is a research tool, and that ICA<br />

assumes at least as many sensors as underlying sources (very rare in a clinical setting), how<br />

might this be extended <strong>for</strong> use other than in an experimental / laboratory setting? Throughout<br />

the paper, especially nearing the end of the discussion, clinical application is frequently<br />

mentioned. The authors may consider great care in generalizing such a tool at the clinical<br />

level, until some of the robustness issues are addressed.<br />

Reply: We agree that in clinical settings the measurement of AEPs takes place in conditions<br />

that differ from the multi-channel EEG recordings used to validate CIAC. However we<br />

believe that moving towards a more objective method to attenuate CI artifacts can contribute<br />

to an increase in the number and in the quality of AEP studies in research settings. It is<br />

expected that outputs from these studies would contribute to developments also at the clinical<br />

level. Moreover it is also expected that the use of a tool such as CIAC will contribute to a<br />

better understanding of the characteristics of the CI artifact and in the long run could even be<br />

possible to develop a truly automatic method. In addition, we would like to point out that<br />

high-density EEG recordings are already possible in clinical settings. The geodesic sensor net<br />

(offered by Electrical Geodesics, www.egi.com) <strong>for</strong> instance allows a very short cap<br />

preparation time (few minutes <strong>for</strong> up to 256-channels!), and the system we currently favor<br />

allows, with some experience, <strong>for</strong> a quite reasonable cap application time (approx. 15 minutes<br />

<strong>for</strong> 64-channnel recordings). Even if further research (see our response to issue #15b)<br />

revealed that high-density recordings are necessary to remove the artifact proper we would<br />

not think that this prevents clinical application.<br />

17) Page 5, Line 123: ".3) a time window of interest <strong>for</strong> the AEP response". It seems that<br />

this largely constricts the robustness of the procedure. One important question related to<br />

AEPs in CI users is the variable latency of the AEP response. If the user attempts to<br />

compensate <strong>for</strong> unknown latencies by, <strong>for</strong> example, using a large window of interest, how<br />

does this change the effectiveness of IC identification? What about the case of the<br />

experimentalist that might be using a data mining approach with unknown prior in<strong>for</strong>mation<br />

about the responses or response differences due to experimental conditions? Can this<br />

procedure still be of use to such a user? For example, what happens if the entire poststimulus<br />

window is used as the window of interest? Are there other approaches to the IC<br />

estimation that can help to overcome such a limitation?<br />

Reply: We agree that the individual differences in the latency of the AEP response in CI users<br />

can be large. However we expect users to be interested in broader ranges of auditory activity<br />

that can last (or peak at) 100-200 ms or even longer, as <strong>for</strong> instance when exploring the N1-P2<br />

complex or the MMN. Thus we recommend that the users select a window of interest larger<br />

than 100 ms in order to account <strong>for</strong> any possible jitter in the responses. A conflict can only<br />

occur if the edge of the time window of interest falls in the time window representing the<br />

offset of the CI artifact. In these cases, the window of interest should be shortened, with the<br />

risk of losing in<strong>for</strong>mation about later responses, as it proved to be the case <strong>for</strong> the TNS where<br />

it was not possible to analyze the P2 response. In the case of language studies where the focus<br />

is on AEP components such as N400 or P600, it is recommend that the stimuli used are longer<br />

than 600 ms. The minimum duration of the stimuli will depend if the user also wants to<br />

investigate early responses such as the N1-P2 complex (please see below #23)). The main<br />

6


eason to require a window of interest as an input parameter is to prevent the selection of ICs<br />

that could result from a poor un-mixing. These ICs may represent brain activity of interest<br />

contaminated by residual CI artifact. As our goal is the attenuation of the CI artifact without<br />

affecting the AEPs, we wanted to make sure that these ICs are not selected. The only way to<br />

overcome this overlapping problem is by developing better ICA algorithms that would<br />

improve the un-mixing from CI artifact and brain sources.<br />

18) Pages 6-7, Lines 161-162, "Details about the clinical profile of the CI users have been<br />

described elsewhere." Even so, it would be helpful to have some more details included in the<br />

current manuscript, especially details related to the implant type, stimulation strategy, etc.<br />

Because this is mostly a paper describing technical selection of CI related artifacts, at least<br />

CI stimulation details should be included, so that the reader has easier access to the<br />

in<strong>for</strong>mation. A brief table would likely suffice <strong>for</strong> such in<strong>for</strong>mation.<br />

Reply: A table with the requested in<strong>for</strong>mation has been included into the revised manuscript.<br />

19) Page 7, Line 180, & Line 182: ".filtered offline from 1-40 Hz"; ".down-sampled to 500<br />

Hz". How does filtering and/or down-sampling the data affect the ICA routine? And, how<br />

does affect the topography and shape of the artifact ICs? Do different parameters here give<br />

different results? Would a researcher interested in 80 Hz activity be more (or, less)<br />

susceptible to mis-characterized artifacts?<br />

Reply: It is known that high-pass filtering has beneficial effects on the ICA decomposition, as<br />

ICA assumes stationarity. We commonly use a 1 Hz high-pass filter in order to remove drifts<br />

from the data and improve the ICA decomposition (<strong>for</strong> instance, see Debener, S. et al, (2010).<br />

Using ICA <strong>for</strong> the Analysis of Multi-Channel EEG Data. In M. Ullsperger & S. Debener<br />

(Eds.), Simultaneous EEG and fMRI (pp. 121-135). New York: Ox<strong>for</strong>d University Press).<br />

Regarding the low-pass filter we choose a cut-off that is used traditionally in ERP research,<br />

since normally the responses of interest are below this frequency range. However the filter<br />

settings seem to not directly influence the maximum number of CI artifact-related ICs found<br />

in the data, as well as their characteristics. For instance, in a study from Debener et al (2008,<br />

Psychophysiology) the filter settings used were from 1-80 Hz, ICs related to the CI artifact<br />

could be identified and the AEPs could be reconstructed.<br />

20) Page 7, Line 184: "Extended infomax ICA." Would this procedure be limited to only this<br />

ICA routine? One concern here, especially if one is clinically minded, is that in many cases<br />

ICA can be both computationally and time intensive. Would other ICA routines (e.g.,<br />

FastICA, JADER, etc.) be less useful? More useful? In other words, would these estimations<br />

be limited by the type of ICA routine that is chosen?<br />

Reply: It is true that some ICA routines are both computationally and time intensive. We use<br />

extended-infomax because it is our experience that it provides reliable ICA decompositions<br />

<strong>for</strong> high-density EEG recordings, independently of the presence of the CI artifact. Although<br />

there is a lack of validation studies comparing ICA algorithms, guidance can be found in the<br />

EEGLAB website<br />

(cf. http://sccn.ucsd.edu/wiki/Chapter_09:_Decomposing_Data_Using_ICA). According to<br />

7


the EEGLAB developers: “Applied to simulated, relatively low dimensional data sets <strong>for</strong><br />

which all the assumptions of ICA are exactly fulfilled, all three algorithms (infomax, jader,<br />

and fastica) return near-equivalent components. We are satisfied that Infomax ICA<br />

(runica/binica) gives stable decompositions with up to hundreds of channels (assuming<br />

enough training data are given, see below), and there<strong>for</strong>e we can recommend its use,<br />

particularly in its faster binary <strong>for</strong>m (binica()). Note about jader: this algorithm uses 4thorder<br />

moments (whereas Infomax uses (implicitly) a combination of higher-order moments)<br />

but the storage required <strong>for</strong> all the 4th-order moments become impractical <strong>for</strong> datasets with<br />

more than ~50 channels. Note about fastica: Using default parameters, this algorithm quickly<br />

computes individual components (one by one). However, the order of the components it finds<br />

cannot be known in advance, and per<strong>for</strong>ming a complete decomposition is not necessarily<br />

faster than Infomax.”<br />

21) Pages 7-8, Lines 188-189: ".equivalent current dipole modeling."<br />

It would be nice to have a better explanation of this choice of modeling. It seems that one<br />

advantage of searching <strong>for</strong> CI stimulus artifact is that we should know exactly where the<br />

source is (i.e., the cochlea, and the from the location of the return electrode). Why not use<br />

this in<strong>for</strong>mation to generate the <strong>for</strong>ward models? Does the ECD modeling return an accurate<br />

estimate of this type of source? Also, the authors earlier mention that the activity from<br />

artifact ICs is "less dipolar and, partly reveal in<strong>for</strong>mation about the location of the internal<br />

components of the CI device, as can be seen in Figure 1, top row." (page 5, line 109). To<br />

what extent are they "less dipolar"? Also, the dipole location in Figure 1 is not very clear in<br />

terms of 3D localization, and is difficult to really interpret in terms of location.<br />

It seems that from an engineering view, these activities would represent highly dipolar point<br />

charges, with in<strong>for</strong>mation flow toward the return electrode. Is this not the case? If not, what<br />

kind of activity is the implant generating? Could this less dipolar-ness be due to volume<br />

conduction effects (the cochlea is very deep in the skull). Further, if such a large activation<br />

is not precisely modeled from its known source, could this provide in<strong>for</strong>mation about the<br />

utility or limits of ECD modeling more generally?<br />

Reply: We agree that our explanation might be misleading. By biophysics, coherent activity<br />

across a small patch of cortex will have a near-dipolar projection pattern on the scalp. To<br />

estimate the location of the equivalent dipole <strong>for</strong> an IC scalp map, there<strong>for</strong>e, we can apply<br />

standard inverse source modeling methods to the IC map, as implemented in DIPFIT plugin.<br />

It is not known which parts of the CI device contribute most to the artifact seen in the EEG<br />

recordings. Other authors have pointed out that is likely that the RF transmitter is one of the<br />

main sources (Debener, et al., 2008; Henkin, et al., 2008; Martin, 2007) of the artifact but<br />

other parts may also contribute. We are not aware of any study where the activity of the<br />

implant has been modeled with an ECD at the sensor level. Here we model the IC scalp maps<br />

with an ECD with the main purpose of excluding the ICs likely to represent brain activity<br />

from the pool that will be evaluated. Another criterion is then applied to the remaining ICs to<br />

further investigate if these ICs represent CI artifacts. In Figure-II we show IC scalp maps and<br />

the respective residual variances obtained after dipole fitting. The fact that all ICs<br />

representing CI artifacts have residual variances larger than 30% indicates that the likelihood<br />

that the generators of such scalp patterns are cortical sources is low. Whether this is related to<br />

the curved positioning of (simultaneously active) electrodes in the cochlea we do not know<br />

<strong>for</strong> certain (but this could be tested easily). Over the past few years we have analysed several<br />

dozen EEGs, recorded from CI users in different labs, with the ICA method and did not notice<br />

a single case where the artifactual ICs had a clear dipolar pattern and unambiguous source<br />

8


location in<strong>for</strong>mation pointing consistently to coil, electrodes, or speech processor. Thus, at<br />

present the electrical properties of the artifact remain poorly understood. Instead we believe<br />

that the residual variance is a valid criterion to distinguish ICs likely to represent brain<br />

activity from others and it is a necessary step in our procedure.<br />

22) Page 10, Line 244: ".expert that was not involved in neither."<br />

Reply: Revised.<br />

23) Page 12-13, Lines 321-322: ".the duration of the auditory stimulus should be longer than<br />

the cortical response interval of interest. This seems to be another major limitation of the<br />

procedure. There are a very large number of studies that use much shorter stimuli. For<br />

example, those interested in clinical protocols often use short tone-bursts, or even click<br />

stimuli. Experimentalists may be interested in very short acoustic changes. Further, a child<br />

with a late P1 or N1 response that may occur at the edge would have a disadvantage when it<br />

comes to observing the true brain activity; and such a patient would be a case where this<br />

in<strong>for</strong>mation may be most useful. In this case, is there a disadvantage to using, <strong>for</strong> example,<br />

short stimuli that end be<strong>for</strong>e the window of interest? Do these transitions between CI artifact<br />

and brain response have a unique IC? A unique EEG signature? Is there some other<br />

in<strong>for</strong>mation that could be used to help parse such activity?<br />

Reply: We agree that the way we wrote the sentence is misleading. The critical latencies are<br />

those that include the onset and offset of the CI artifact and are related to the duration of the<br />

stimuli. As reported <strong>for</strong> TNS the offset of the CI artifact coincided with the P2 response time<br />

range. There<strong>for</strong>e <strong>for</strong> some CI users the P2 latency window was still corrupted by residual CI<br />

artifact. When using short stimuli, such as clicks and tone-bursts, we do not expect the<br />

responses of interest to coincide with the onset/offset of the CI artifact. However our<br />

experience reveals that responses falling in one of these edges can be critical.<br />

In our data we have not found ICs representing particularly transitions between CI artifact and<br />

brain responses. In case of a poor ICA unmixing it is possible that ICs representing brain<br />

responses would also contain residual artifact in the onset/offset time window. If that is the<br />

case, these ICs should not be corrected, in order to preserve the brain response. We would like<br />

to suggest that stimuli durations between 90 and 300 ms should be avoided when studying the<br />

N1-P2 complex or the MMN in adults. This procedure would ensure that the likelihood that<br />

the response of interest coincides with the onset/offset of the artifact is low. The discussion of<br />

these aspects has been revised in the manuscript.<br />

24) Pages 14-15, Conclusion: See "General Comments" above. A discussion of the types of<br />

in<strong>for</strong>mation that would be helpful in improving this procedure would be nice. If, as the<br />

authors suggest, outside testing and use would improve the overall procedure, then some<br />

guidance (based on their expertise and experience) about the types of questions to ask, would<br />

certainly do more to spur such work.<br />

Reply: As described previously (#15b), we have provided a broader discussion where we have<br />

highlighted <strong>for</strong> instance the importance of further validating CIAC using other EEG montages<br />

and types of stimulation.<br />

9


25) The ESS in<strong>for</strong>mation in figures 3 and 4 seems largely redundant, as the AEP responses in<br />

Figure 4 are the 12 subject sub-set of responses from Figure 3. Is this correct? Why not<br />

show comparison traces <strong>for</strong> the manual versus automatic rejection <strong>for</strong> both experimental<br />

studies in Figure 4? This could largely eliminate the need to plot the same responses in<br />

Figure 3. (of course, without the direct comparison <strong>for</strong> an additional six subjects). Also, it<br />

would be helpful to include some labeling of peaks and landmarks in the figures; <strong>for</strong> example<br />

arrows or lines showing where P1, N1, and P2 were approximated <strong>for</strong> each subject. Arrows<br />

showing implant artifact (and possibly residual artifact as might be apparent <strong>for</strong> some<br />

subjects in Figure 4 <strong>for</strong> the TNS study). In other words, some way to better alert the reader to<br />

the key in<strong>for</strong>mation in each figure.<br />

Reply: Fig. 4 has been revised according to the reviewers’ suggestions.<br />

Reviewer #3:<br />

26) Manual selection of ICs is tedious, subjective and usually inaccurate, so method <strong>for</strong><br />

automatic selection is welcome. However, the IC corresponding to the CI electrical artifact is<br />

usually very obvious (e.g. usually appear as the first IC using FAST ICA or JADER or by<br />

plotting its topography) such that an elaborate procedure as the one suggested may not be<br />

necessary. At least it needs to be more clearly justified than what is presented -lines 100-118,<br />

(e.g. by providing an example where a simple IC selection fails). A comparison between the<br />

CIAC method and standard methods of CI artifact correction would highlight the benefit of<br />

the technique. Comparing to manual selection is not really appropriate or valid.<br />

Reply: We are not aware of any solid, published evidence showing that FASTICA or JADER<br />

(or any other ICA algorithm) consistently identifies the CI electrical artifact as the first IC. In<br />

most algorithms ICs are sorted by descending order regarding the amount of variance they<br />

explain, and whether this is the case depends, we believe, on a number of issues (like the<br />

stimulus presentation duration relative to the total recording time, among others). The figure<br />

below shows the position of identified ICs in each individual, clearly demonstrating that<br />

components are relatively widely distributed.<br />

10


As requested we show in the next figure a typical case (subject #1 of study ESS) where the<br />

simplest IC selection suggested by the reviewer, that is, removing only the first ranked<br />

component, clearly does not produce reasonable AEPs results. We would be happy to add this<br />

analysis to the revised paper if really requested by the reviewer.<br />

The reviewer also questions the expert classification evidence as a reference base against<br />

which CIAC is compared. Un<strong>for</strong>tunately the ground truth in real recordings is never known,<br />

and there<strong>for</strong>e many papers in the context of EEG linear decomposition development and<br />

application rely on expert judgments as a reference base (e.g., Winkler et al., 2011,<br />

11


Behavioral and Brain Functions, 7; 30; <strong>for</strong> a very recent example). Here we have adopted<br />

exactly the same strategy and believe that this is in fact a valid approach. We are confident<br />

that aiming towards objective methods to identify ICs representing the CI artifact is necessary<br />

and constitutes a clear improvement over manual component selection.<br />

Furthermore the reviewer also requests a comparison to other tools that classify ICs. Please<br />

see our detailed response below (issue #28). CIAC was developed to identify ICs objectively<br />

across subjects and although it requires the evaluation of spatial and temporal properties of<br />

the ICs, we do not consider it to be “very elaborated”, as it is based on only two features and<br />

the definition of only three parameters. Other studies explored a much larger feature space,<br />

e.g. Winkler et al., 2011 (starting with 38 features) or Nolan et al., 2010 (FASTER includes<br />

six main steps and <strong>for</strong> each step more than five parameters can be adjusted). Accordingly,<br />

CIAC is a rather simple, robust yet flexible tool.<br />

27) The steps of CIAC seem ad hoc and not grounded in any understanding of the biophysics<br />

or features of the CI electrical artifact, and as such does not offer any new contribution over<br />

and above present attempts <strong>for</strong> 'automatic' selection.<br />

Reply: As discussed above, it is indeed the case that the electrical artifact is not well<br />

understood at present. Here the problem is again that the ground truth is not known, since<br />

pure artifact only recordings cannot be made. At least we are not aware of any procedure<br />

measuring the artifact exclusively, using an adequate <strong>for</strong>ward model. We would greatly<br />

appreciate if the reviewer could point us to any reference describing the electrical CI artifact<br />

in detail. Accordingly it is not well understood which electrical properties of the CI device are<br />

represented in the EEG recording. The steps implemented in CIAC are grounded in the<br />

properties of the ICs representing the CI artifact, since it is an ICA-based tool. We believe<br />

that CIAC is a significant contribution towards a more objective, reliable and efficient<br />

recovery of AEPs from CI users. Other available tools were tested and the results obtained<br />

were not satisfactory (please see #28 below).<br />

28) While the focus in automatic artifact/IC detection has been on eye movement and muscle<br />

artifacts, the *procedures* of automatic identification of (un)desired features should be<br />

generic, and should be possible to tune <strong>for</strong> any required features, and as such comparison<br />

with existing tools is required. So a comparison with Ting et al, Morgan et al or even the<br />

authors own previous work Viola et al, 2009 should be considered. Alternatively, a clear<br />

justification of a new method based, <strong>for</strong> example, on the uniqueness of the properties or<br />

biophysics of electric artifact would be needed (over an above just stating " electrical CI<br />

artifacts have a particular signature in the spatial and temporal domain as illustrated in<br />

Figure 1"). The setting of the thresholds seems arbitrary and their selection may, in some<br />

cases, require tuning, which questions the claim of automaticity, which perhaps should be<br />

revised.<br />

Reply: We had run comparisons between our method and other available tools but had<br />

decided to not include them in the manuscript. The reason <strong>for</strong> that is that we do not find it fair<br />

reporting poor per<strong>for</strong>mance of a tool when a priori in<strong>for</strong>mation makes it very likely that it<br />

cannot work. For instance, as explained previously (see #9), CORRMAP is not expected to<br />

find ICs representing the CI artifact (or ICs representing neural activity; see Wessel et al.,<br />

12


2011, Neuroimage, <strong>for</strong> an unfair comparison using CORRMAP) across subjects and indeed,<br />

that is the case. What exactly the value of reporting this would be we do not know.<br />

CORRMAP operates on a map correlation approach and as it is shown in Fig. #9 and Fig.<br />

#15a, the ICs related to CI artifact are not consistently similar across subjects, thus<br />

CORRMAP will per<strong>for</strong>m poorly. In the case of ADJUST (Mognon et al 2010), this tool was<br />

developed with the aim of clustering ICs representing eye blinks, lateral and vertical eye<br />

movements and discontinuities. Our prediction was that if ADJUST would be able to identify<br />

CI artifact related ICs, these ICs would be labeled as “discontinuities”. We ran ADJUST <strong>for</strong><br />

the ESS. Since we per<strong>for</strong>med our artifact correction in two steps: CORRMAP (used to find<br />

blinks, lateral eye movements and heartbeat artifacts) + CIAC (used to find CI artifact related<br />

ICs), we believe it is reasonable to compare the selection from CIAC with the sub-selection<br />

from ADJUST <strong>for</strong> “discontinuities”. The results are summarized in the table below, where we<br />

show the total number of ICs found by each method, as well as the number of ICs found by<br />

both methods (“common in both”). As expected the per<strong>for</strong>mance of ADJUST on finding<br />

ocular artifacts was good but on CI artifact related ICs was poor. When correcting all ICs<br />

found using ADJUST it was not possible to recover AEPs free from CI artifact, <strong>for</strong> any of the<br />

18 CI users that comprised ESS.<br />

Nr. ICs identified<br />

CI CIAC ADJUST Common in<br />

user<br />

both<br />

1 10 0 0<br />

2 15 2 0<br />

3 9 8 3<br />

4 22 3 2<br />

5 16 4 1<br />

6 9 5 0<br />

7 5 6 0<br />

8 5 2 0<br />

9 4 11 0<br />

10 11 4 2<br />

11 5 3 0<br />

12 2 9 0<br />

13 6 6 0<br />

14 4 8 0<br />

15 5 7 2<br />

16 13 9 0<br />

18 6 11 0<br />

19 7 10 0<br />

In the case of FASTER (Nolan et al, 2010), this tool was developed to receive as input raw<br />

data and run a sequence of steps that including different types of pre-processing and artifact<br />

correction. We could not run only one of the steps, there<strong>for</strong>e it was not possible to test the<br />

13


artifact correction only and compare it with CIAC. In the case of the method developed by<br />

Ting el al, 2006, it was optimized <strong>for</strong> blind source separation methods using second order<br />

statistics, which differ substantially from the Infomax ICA-approach used here. Thus in our<br />

opinion a comparison would not be valid. Accordingly we decided to not expand on this issue<br />

and not report these results in the revised paper, but would be a happy to do so in a further<br />

revision if specifically requested by the reviewer.<br />

14


*Highlights<br />

Highlights<br />

New algorithm selects independent components related to cochlear implant artifacts<br />

Auditory evoked potentials (AEPs) with reasonable quality could be reconstructed<br />

High correlation between age and N1-P2 peak-to-peak amplitude in the AEPs<br />

High test-retest reliability <strong>for</strong> AEP N1 amplitudes and latencies<br />

Preservation of individual response characteristics after automatic attenuation


*<strong>Manuscript</strong><br />

Click here to view linked References<br />

1<br />

2<br />

3<br />

4<br />

5<br />

6<br />

7<br />

8<br />

9<br />

10<br />

11<br />

12<br />

13<br />

14<br />

15<br />

16<br />

17<br />

18<br />

19<br />

20<br />

21<br />

22<br />

23<br />

24<br />

25<br />

26<br />

Abstract<br />

Electrical artifacts caused by the cochlear implant (CI) contaminate electroencephalographic (EEG)<br />

recordings from implanted individuals and corrupt auditory evoked potential (AEPs). Independent<br />

component analysis (ICA) is efficient in attenuating the electrical CI artifact and AEPs can be<br />

successfully reconstructed. However the manual selection of CI artifact related independent<br />

components (ICs) obtained with ICA is unsatisfactory, since it contains expert-choices and is time<br />

consuming.<br />

We developed a new procedure to evaluate temporal and topographical properties of ICs and semi-<br />

automatically select those components representing electrical CI artifact. The CI Artifact Correction<br />

(CIAC) algorithm was tested on EEG data from two different studies. The first consists of published<br />

datasets from 18 CI users listening to environmental sounds. Compared to the manual IC selection<br />

per<strong>for</strong>med by an expert the sensitivity of CIAC was 91.7% and the specificity 92.3%. After CIAC-<br />

based attenuation of CI artifacts, a high correlation between age and N1-P2 peak-to-peak amplitude<br />

was observed in the AEPs, replicating previously reported findings and further confirming the<br />

algorithm’s validity.<br />

In the second study AEPs in response to pure tone and white noise stimuli from 12 CI users that had<br />

also participated in the other study were evaluated. CI artifacts were attenuated based on the IC<br />

selection per<strong>for</strong>med semi-automatically by CIAC and manually by one expert. Again, a correlation<br />

between N1 amplitude and age was found. Moreover, a high test-retest reliability <strong>for</strong> AEP N1<br />

amplitudes and latencies suggested that CIAC based attenuation reliably preserves plausible individual<br />

response characteristics.<br />

We conclude that CIAC enables the objective and efficient attenuation of the CI artifact in EEG<br />

recordings, as it provided a reasonable reconstruction of individual AEPs. The systematic pattern of<br />

individual differences in N1 amplitudes and latencies observed with different stimuli at different<br />

sessions, strongly suggests that CIAC can overcome the electrical artifact problem. Thus CIAC<br />

facilitates the use of cortical AEPs as an objective measurement of auditory rehabilitation.<br />

1


27<br />

28<br />

29<br />

30<br />

31<br />

32<br />

33<br />

34<br />

35<br />

36<br />

37<br />

38<br />

39<br />

40<br />

41<br />

42<br />

43<br />

44<br />

45<br />

46<br />

47<br />

48<br />

49<br />

50<br />

51<br />

52<br />

53<br />

Keywords: cochlear implant, AEPs, N1, artifact attenuation, test-retest reliability<br />

1. Introduction<br />

Auditory evoked potentials (AEPs) are important <strong>for</strong> the evaluation of auditory cortex functions in<br />

normal hearing and hearing impaired humans. Several studies have used AEPs to investigate how the<br />

auditory cortex adapts to the artificial input provided by a cochlear implant (CI). Examples are the<br />

measurement of the P1 response to investigate the functional development of the auditory cortex in<br />

children fitted with CIs (Gilley et al., 2008; Sharma et al., 2005), the study of brain asymmetries in the<br />

auditory cortex (Debener et al., 2008; Sandmann et al., 2009), the investigation of neural correlates of<br />

musical sound perception (Koelsch et al., 2004; Sandmann et al., 2010), and the relationship of AEPs<br />

to speech perception (Henkin et al., 2009; Kelly et al., 2005; Lonka et al., 2004; Zhang et al., 2010;<br />

Zhang et al., 2011). Based on those and other studies it has been suggested that the functional integrity<br />

of the auditory system from CI users, which varies widely across patients, as well as the capacity <strong>for</strong><br />

cortical plasticity, deserves more attention when investigating implantation outcome (Moore and<br />

Shannon, 2009; Wilson and Dorman, 2008).<br />

One of the limitations of using AEPs as a routine research or clinical tool is the fact that the EEG<br />

recordings taken from CI users are contaminated by electrical artifacts which coincide in time with<br />

auditory stimulation. Other authors have already described in detail the characteristics of the CI<br />

artifact (Gilley et al., 2006). The CI artifact properties vary widely across devices, individuals, and<br />

types of stimulation (Gilley, et al., 2006; Viola et al., 2011) and the literature is inconsistent<br />

concerning the prevalence of the artifact. In some CI users the absence of artifacts in EEG recordings<br />

has been reported (Zhang, et al., 2010). Moreover, at least one study suggested electrical artifacts only<br />

occur at response latencies different from cortical AEPs and thus did not report difficulties in the<br />

measurement of AEPs (Koelsch, et al., 2004). It has also been speculated that the CI artifact may be<br />

present until one year after CI activation (Lonka, et al., 2004). Despite these reports, most studies<br />

presenting multi-channel EEG data have found that AEPs from CI users are strongly corrupted by a<br />

large electrical artifact generated by the CI device, thus impairing any type of analysis unless tailored,<br />

sophisticated and often time-consuming artifact processing techniques are applied. Accordingly,<br />

2


54<br />

55<br />

56<br />

57<br />

58<br />

59<br />

60<br />

61<br />

62<br />

63<br />

64<br />

65<br />

66<br />

67<br />

68<br />

69<br />

70<br />

71<br />

72<br />

73<br />

74<br />

75<br />

76<br />

77<br />

78<br />

79<br />

80<br />

adequate artifact attenuation seems crucial <strong>for</strong> the AEP-based study of auditory cortex rehabilitation in<br />

CI users.<br />

A traditional approach to attenuate the CI artifact is the subtraction technique, where the presentation<br />

of the auditory stimuli is manipulated to create experimental conditions where the AEP response<br />

varies but the CI artifact remains constant (Friesen and Picton, 2010). Un<strong>for</strong>tunately this approach<br />

limits the type of experimental paradigms that can be used and it has only been tested <strong>for</strong> a small<br />

population in the context of multi-channel EEG recordings. Other authors used linearly constrained<br />

minimum variance beam<strong>for</strong>mers to reconstruct cortical activity with minimal artifact interference<br />

(Wong and Gordon, 2009). This approach has been reported to work in a single case study. It is also<br />

possible to minimize the CI artifact by using an optimized differential reference (ODR) technique.<br />

Here the reference of the EEG montage is placed in a location that allows recording a particular<br />

electrode of interest free of artifact (Gilley, et al., 2006). A shortcoming of the ODR technique is<br />

finding and validating the best location <strong>for</strong> the reference <strong>for</strong> each CI user, which is time consuming.<br />

The ODR approach makes it also difficult to analyze AEPs on the cortical source level. A more<br />

generic and promising approach is the use of independent component analysis (ICA) to separate the<br />

EEG signals into statistically maximally independent components (Makeig et al., 2004; Onton et al.,<br />

2006). These independent components (ICs) need to be evaluated by an expert in order to select those<br />

representing the CI artifact. It has been shown in various CI users using different types of devices that<br />

the ICA method allows good attenuation of the CI artifact and the reconstruction of individual AEPs<br />

(Debener, et al., 2008; Gilley, et al., 2006; Gilley, et al., 2008; Sandmann, et al., 2009; Sandmann, et<br />

al., 2010; Viola, et al., 2011; Zhang, et al., 2010; Zhang, et al., 2011). Furthermore it has been reported<br />

that after attenuation of CI artifacts, individual differences were reasonably well reconstructed, as<br />

evidenced by a high correlation between age and AEP amplitudes (Viola, et al., 2011).<br />

However, one significant limitation of the ICA approach is the laborious selection of the ICs<br />

representing electrical CI artifact (Viola, et al., 2011). This process is subjective and time consuming,<br />

since it requires extensive visual inspection of all ICs by a trained operator. Although automatic<br />

methods have been developed that reasonably well identify ICs representing conventional EEG<br />

3


81<br />

82<br />

83<br />

84<br />

85<br />

86<br />

87<br />

88<br />

89<br />

90<br />

91<br />

92<br />

93<br />

94<br />

95<br />

96<br />

97<br />

98<br />

99<br />

100<br />

101<br />

102<br />

103<br />

104<br />

105<br />

106<br />

107<br />

artifacts (Mognon et al., 2010; Nolan et al., 2010; Viola et al., 2009), they are not optimal to identify<br />

components representing electrical CI artifacts, which have a particular signature in the spatial and<br />

temporal domain as illustrated in Figure 1.<br />

Accordingly we aimed at developing and validating a novel, semi-automatic and user-friendly<br />

algorithm that screens the temporal and spatial properties of ICs and identifies the components<br />

representing the CI artifact. The Cochlear Implant Artifact Correction (CIAC) tool presented here<br />

provides a faster and, more importantly, more objective CI artifact attenuation, and thus facilitates the<br />

reconstruction of AEPs. In a first step the CIAC approach was validated using a published AEP study<br />

from 18 adult CI users. In this study IC classification was per<strong>for</strong>med manually by a well-trained<br />

researcher (FCV) (Viola, et al., 2011), and the resulting AEPs served as a reference <strong>for</strong> the<br />

development of CIAC. In a second step, an unpublished set of EEG recordings from a subgroup of the<br />

same CI users (N = 12) presented with different auditory stimuli and recorded 12 months earlier was<br />

evaluated. CI artifacts were attenuated based on the selection per<strong>for</strong>med by CIAC and by another well-<br />

trained expert (PS) with several years of experience in using ICA <strong>for</strong> the evaluation of AEPs from CI<br />

users (Sandmann, et al., 2009; Sandmann, et al., 2010). For both studies the AEPs after semi-<br />

automatic selection with CIAC were evaluated. We adjusted CIAC parameters in both studies, aiming<br />

<strong>for</strong> high sensitivity and specificity. We also explored whether AEPs reconstructed after CIAC<br />

selection would show good temporal stability.<br />

2. Methods<br />

2.1. CIAC description<br />

ICs can be characterized by their properties both in the temporal and in the spatial domain and<br />

different criteria can be defined to distinguish between brain related and CI artifact related<br />

components. In the spatial domain the residual variance (RV) between the actual IC topography and<br />

the model projection <strong>for</strong> the equivalent dipole to the same electrode montage can be used as a<br />

differentiation criterion (Gramann et al., 2010; Onton, et al., 2006), since brain related ICs are dipolar<br />

and thus have a much lower RV (Figure 1, top row). On the other hand the topographies of the CI<br />

artifact related ICs are less dipolar and can be differentiated based on RV, as can be seen in Figure 1,<br />

4


108<br />

109<br />

110<br />

111<br />

112<br />

113<br />

114<br />

115<br />

116<br />

117<br />

118<br />

119<br />

120<br />

121<br />

122<br />

123<br />

124<br />

125<br />

126<br />

127<br />

128<br />

129<br />

130<br />

131<br />

132<br />

133<br />

134<br />

135<br />

top row. Contrary to several conventional types of EEG artifacts, which are represented by ICs with<br />

similar topographies across subjects (Viola, et al., 2009), the CI artifact topographies can be<br />

substantially different across individuals. On the other hand, <strong>for</strong> the same CI user, the topographies of<br />

the ICs reflecting the artifact may share a substantial degree of similarity (not shown). We could not<br />

find any relationship between the topographical pattern of the ICs and the type of CI device or other<br />

related properties. However when inspecting the temporal properties of the CI related ICs from<br />

different CI users it is evident that they share very similar profiles. The largest activity takes place<br />

during the onset and/or offset of the artifact as can be seen both <strong>for</strong> the time-locked average of the IC<br />

activation (IC ERP) and its first temporal derivative (Figure 1, middle and bottom rows). On the other<br />

hand, ICs representing <strong>for</strong> instance late auditory cortex related activity usually have largest deflections<br />

in the time window corresponding to the N1-P2 responses (100-250 ms) as illustrated in Figure 1.<br />

Based on these observations we implemented an algorithm that combines spatial and temporal<br />

in<strong>for</strong>mation and selects CI related ICs using three steps and three thresholds. Figure 2 shows a<br />

schematic flow chart of the cochlear implant artifact correction (CIAC) algorithm. As a starting point<br />

three user inputs are required: 1) ICs from one or more EEG dataset, epoched to the same auditory<br />

stimuli of interest; 2) the duration of the auditory stimuli; 3) a time window of interest <strong>for</strong> the AEP<br />

response. A loop of three steps is then computed <strong>for</strong> the ICs from each single dataset. In the first step<br />

CIAC selects the ICs with RV larger than a pre-defined threshold (RV > Thr_rv). This is an exclusion<br />

step to avoid the further selection of ICs that would likely represent brain related activity. In the<br />

second step the temporal derivative of the ERP is calculated <strong>for</strong> each of the ICs which are part of the<br />

subset selected in the first step. Then the ratio is computed between the root mean square (RMS)<br />

amplitude of the IC temporal derivative in the artifact onset/offset time window (derived from the<br />

duration of the auditory stimuli – user input) and the RMS amplitude <strong>for</strong> the time window where the<br />

responses of interest are expected (user input). The IC with the largest ratio is selected as a<br />

topographical template <strong>for</strong> that particular CI user. This IC is the one reflecting the strongest artifact<br />

profile, and its topography is then going to be correlated with all other topographies from the same CI<br />

user. In the third step ICs are selected if at least one of these two criteria is met: having a ratio larger<br />

than a pre-defined threshold (ratio > Thr_deriv) or the correlation between topography and the CI<br />

5


136<br />

137<br />

138<br />

139<br />

140<br />

141<br />

142<br />

143<br />

144<br />

145<br />

146<br />

147<br />

148<br />

149<br />

150<br />

151<br />

152<br />

153<br />

154<br />

155<br />

156<br />

157<br />

158<br />

159<br />

160<br />

161<br />

artifact template being above a pre-defined threshold (correlation > Thr_corr). After CIAC per<strong>for</strong>med<br />

these three steps, visual representations are presented to the user, displaying the topographies of the<br />

selected ICs as well as the original AEP and the corrected AEP. The user is thus provided with a visual<br />

representation of the degree of CI artifact attenuation that could be achieved <strong>for</strong> each individual<br />

dataset as well as a grand average summary plot. In case of an unsatisfactory result CIAC could be re-<br />

run using different, user-defined thresholds <strong>for</strong> the whole study set.<br />

For the residual variance threshold we experienced consistently good results using values between<br />

15% < Thr_rv < 25%. The recommended value is Thr_rv = 20%, which is close to residual variance<br />

thresholds <strong>for</strong> the evaluation of IC quality as described in other studies (Gramann, et al., 2010; Onton,<br />

et al., 2006). The recommended values <strong>for</strong> the derivative ratio threshold are 1.5 < Thr_deriv < 3.5.<br />

This range is motivated by a previous study (Viola, et al., 2011) where the ICs manually labeled as CI<br />

artifact were characterized by a ratio range between 1.5 and 12. For the correlation threshold the<br />

recommended values are 0.85 < Thr_corr < 0.95, which represents a rather conservative range, since<br />

in our experience ICs reflecting CI artifact have either very similar or uncorrelated topographies. Note<br />

that a set of thresholds need to be defined only once. The set of thresholds depends only on the type<br />

and duration of auditory stimuli used and is not related to the type of CI device. The set of thresholds<br />

chosen is applied always to all datasets comprised in a study set.<br />

To validate the CIAC algorithm EEG recordings from CI users comprising two study sets were<br />

evaluated. In section 2.2 we report the results of a study with environmental sounds and hence<strong>for</strong>th we<br />

refer to it as Environmental Sounds Study (ESS). In section 2.3 we report the results of a study with<br />

tonal and noise stimulation and hence<strong>for</strong>th labeled this as Tones and Noise Study (TNS). Note that,<br />

chronologically, the TNS study was recorded first and the ESS study recorded approximately 12<br />

month later, but, because the ESS data were already available (Viola, et al., 2011), they were used <strong>for</strong><br />

the development of CIAC and thus are reported first.<br />

2.2. Environmental Sounds Study (ESS)<br />

2.2.1. Subjects<br />

6


162<br />

163<br />

164<br />

165<br />

166<br />

167<br />

168<br />

169<br />

170<br />

171<br />

172<br />

173<br />

174<br />

175<br />

176<br />

177<br />

178<br />

179<br />

180<br />

181<br />

182<br />

183<br />

184<br />

185<br />

186<br />

187<br />

188<br />

Eighteen post-lingually deafened, adult CI users (10 females, M = 59.89, SD = 13.06 years) were<br />

recruited from the South of England Cochlear Implant Centre in Southampton, UK. All CI users were<br />

right-handed and had no history of neurological or psychiatric disorders. All but one CI user were<br />

unilaterally implanted and all CI users had at least 6 months experience with the implant (CI<br />

experience: M = 40.22, SD = 36.60 months). Details about the clinical profile of the CI users are<br />

shown in Table 1. Procedures were approved by the local ethics committee and con<strong>for</strong>med to the<br />

declaration of Helsinki.<br />

2.2.2. Stimuli and Task<br />

CI users were stimulated with 80 environmental sounds with 400 ms duration and a sampling rate of<br />

22 kHz via two loudspeakers (Quad L12) positioned at an azimuth of 45°/135° in front of the<br />

participant. Sounds were presented at a com<strong>for</strong>table level adjusted individually <strong>for</strong> each participant<br />

using a five level loudness com<strong>for</strong>t rating scale varying from 60 to 80 dB SPL in steps of 5 dB SPL.<br />

CI users were using their implant on their standard everyday settings. The environmental sounds were<br />

used as primes in a semantic priming paradigm (Viola, et al., 2011).<br />

2.2.3. Electrophysiology recordings<br />

CI users were seated in an electrically shielded, sound attenuated and dimly lit booth (Industrial<br />

Acoustics, Winchester, UK) and EEG data were recorded from 68 channels using a high-input<br />

impedance amplifier system (Compumedics Neuroscan, Charlotte, NC, USA) and a customized,<br />

infracerebral electrode cap with an equidistant electrodes lay-out (Easycap, Herrsching, Germany).<br />

Data were recorded with a sampling rate of 1000 Hz using the nose-tip as reference, and were<br />

analogue filtered between 0.1 and 200 Hz. Electrode impedances were maintained below 20 kΩ prior<br />

to data acquisition.<br />

2.2.4. Data processing<br />

EEG data were processed using custom scripts and EEGLAB (Delorme and Makeig, 2004) running in<br />

the MATLAB (Mathworks, Natick, MA) environment. Data were filtered offline from 1 to 40 Hz<br />

using sinc FIR filters windowed with a Hann window (courtesy of A. Widmann: www.uni-<br />

leipzig.de/~biocog/content/widmann/eeglab-plugins/). Data were then down-sampled to 500 Hz and<br />

7


189<br />

190<br />

191<br />

192<br />

193<br />

194<br />

195<br />

196<br />

197<br />

198<br />

199<br />

200<br />

201<br />

202<br />

203<br />

204<br />

205<br />

206<br />

207<br />

208<br />

209<br />

210<br />

211<br />

212<br />

213<br />

214<br />

215<br />

pruned of unique, non-stereotyped artifacts using the EEGLAB function jointprob.m (Delorme et al.,<br />

2007). Extended infomax ICA as implemented in EEGLAB was then applied to the remaining data in<br />

order to achieve a reliable decomposition (Debener et al., 2010). Independent components (ICs)<br />

representing eye-blinks and electrocardiographic artifacts were semi-automatically identified using<br />

CORRMAP (Viola, et al., 2009) and then corrected from all datasets. Since one of the parameters<br />

assessed in the CIAC algorithm is the residual variance (RV) equivalent current dipole modeling was<br />

then computed <strong>for</strong> the remaining ICs using a four-shell spherical head model and procedures<br />

implemented in the EEGLAB DIPFIT toolbox (Oostenveld and Oostendorp, 2002). Afterwards all<br />

datasets were segmented into epochs from -200 to 600 ms relative to sound onset. AEPs were obtained<br />

by time-domain averaging and the pre-stimulus interval (-200 to 0 ms) was used <strong>for</strong> baseline<br />

subtraction.<br />

2.2.5. Semi-automatic identification of CI artifact related ICs using CIAC<br />

CIAC was tested using the following input parameters: ICs from 18 EEG datasets recorded from the<br />

CI users, 400 ms <strong>for</strong> the duration of the auditory stimuli and the interval from 80 to 250 ms<br />

(corresponding to the N1-P2 responses) was selected as time window of interest. The following<br />

threshold values were used: Thr_rv = 20%; Thr_deriv = 2.5; Thr_corr = 0.95. After semi-automatic<br />

attenuation of CI artifacts, AEPs at missing electrodes were interpolated with the EEGLAB function<br />

eeg_interp.m, which implements a smoothed inverse distance approach. The sensitivity and specificity<br />

of CIAC were evaluated, taking as “gold standard” the manual selection previously per<strong>for</strong>med by an<br />

expert (Viola, et al., 2011). Sensitivity was defined as the ratio between the number of ICs selected<br />

both by CIAC and by the expert (Hits) divided by the sum of Hits and Misses, the latter corresponds to<br />

the ICs identified only by the expert. Specificity was defined as the ratio between the number of ICs<br />

not selected both by the expert and CIAC (Correct Rejects) divided by the sum of Correct Rejects and<br />

False Alarms, the latter corresponds to the ICs selected by CIAC only.<br />

2.2.6. Auditory evoked potential quantification<br />

AEP amplitude and latency analyses were per<strong>for</strong>med <strong>for</strong> the fronto-central channel closed to Fcz,<br />

showing the largest grand average N1 amplitude. AEP peak amplitudes and latencies were determined<br />

8


216<br />

217<br />

218<br />

219<br />

220<br />

221<br />

222<br />

223<br />

224<br />

225<br />

226<br />

227<br />

228<br />

229<br />

230<br />

231<br />

232<br />

233<br />

234<br />

235<br />

236<br />

237<br />

238<br />

239<br />

240<br />

241<br />

using a semi-automatic procedure as implemented in peakdet.m (www.billauer.co.il/peakdet.h<strong>tm</strong>l).<br />

The N1 and P2 peak amplitudes and latencies obtained after manual and semi-automatic selection of<br />

CI related ICs were compared.<br />

2.3. Tones and Noise Study (TNS)<br />

2.3.1. Subjects<br />

Twelve post-lingually deafened cochlear implant (CI) users (5 females, M = 61.75, SD = 13.46 years)<br />

implanted unilaterally were recruited from the South of England Cochlear Implant Centre in<br />

Southampton, UK. All CI users were right-handed and had no history of neurological or psychiatric<br />

disorders. Procedures were approved by the local ethics committee and con<strong>for</strong>med to the declaration<br />

of Helsinki. All CI users had at least 6 months experience with the implant (CI experience: M = 41.75,<br />

SD = 37.69 months). The CI users participated also in the Environmental Sounds Study (ESS)<br />

approximately one year later (Interval between recordings: M = 12.33, SD = 0.98 months).<br />

2.3.2. Stimuli and Task<br />

The procedures were the same as described <strong>for</strong> the single case previously reported (Debener, et al.,<br />

2008). Stimuli were 1-kHz tones and white noise, 220 ms long with 10 ms rise and fall time, sampled<br />

at 44.1 kHz and presented at 70 dB SPL. The stimuli were presented using two loudspeakers (Quad<br />

L12) as described in the ESS, while the participants watched a silent movie. CI users were using their<br />

implant on their standard everyday settings.<br />

2.3.3. Electrophysiology recordings<br />

The procedure used was the same as described <strong>for</strong> the ESS (section 2.2.3.).<br />

2.3.4. Data analysis<br />

EEG data were processed the same way as described <strong>for</strong> the ESS (section 2.2.4.).<br />

2.3.5. Semi-automatic identification of CI artifact related ICs using CIAC<br />

CIAC was applied to the twelve CI user datasets using as input parameters 220 ms <strong>for</strong> the duration of<br />

the auditory stimuli and the interval from 54 to 180 ms (corresponding to the P1-N1 responses) <strong>for</strong> the<br />

time window of interest. After running CIAC with the thresholds used in the ESS it was observed that<br />

9


242<br />

243<br />

244<br />

245<br />

246<br />

247<br />

248<br />

249<br />

250<br />

251<br />

252<br />

253<br />

254<br />

255<br />

256<br />

257<br />

258<br />

259<br />

260<br />

261<br />

262<br />

263<br />

264<br />

265<br />

266<br />

267<br />

some of the reconstructed AEPs were still contaminated by a large CI artifact. The thresholds were<br />

then adjusted (Thr_rv = 20%; Thr_deriv = 1.5; Thr_corr = 0.9) within the range proposed above and<br />

CIAC was run again. It is worth noting that the Thr_deriv is largely dependent on the strength of the<br />

CI artifact present in the data, which can vary depending on the type of stimuli and stimulus<br />

presentation details. The ICs selected semi-automatically were corrected and the AEPs were<br />

reconstructed <strong>for</strong> all datasets. AEPs at missing electrodes were interpolated with the EEGLAB<br />

function eeg_interp.m, which implements a smoothed inverse distance approach. The sensitivity and<br />

specificity of CIAC were again evaluated, taking as “gold standard” the manual selection per<strong>for</strong>med<br />

by an independent expert that was not involved in data collection and processing or development of<br />

the algorithm (PS).<br />

2.3.6. Auditory evoked potential quantification<br />

AEP amplitude and latency parametrization was per<strong>for</strong>med as described in the ESS. Since the stimuli<br />

used in the TNS had the duration of 220 ms, <strong>for</strong> some CI users the P2 response was contaminated by<br />

residual offset CI artifact. There<strong>for</strong>e the focus of the analysis was on the P1-N1 time window with a<br />

particular focus on the N1 response, which could be identified <strong>for</strong> all CI users.<br />

2.4 Statistical analysis<br />

All variables were tested <strong>for</strong> normality using Shapiro-Wilk tests. For the ESS comparisons between<br />

manual and semi-automatic CI artifact selection N1 and P2 peak amplitudes and latencies were<br />

evaluated using two-tailed paired t-tests. For four out of the eighteen CI users it was not possible to<br />

identify a P2 component in the AEPs after semi-automatic attenuation, similar to previous<br />

observations (Viola, et al., 2011). For these participants P2 amplitude was taken as 0 V and P2<br />

latency as the mean value of the other CI users (245 ms). In order to investigate if the semi-automatic<br />

selection of CI related ICs would preserve the individual differences found <strong>for</strong> the manually corrected<br />

datasets (Viola, et al., 2011), the Spearman correlation, indicated with rS, between age and N1-P2<br />

peak-to-peak amplitude was calculated. For the TNS study the Spearman correlation between age and<br />

N1 amplitude was computed. The same correlation was calculated <strong>for</strong> the ESS using only the datasets<br />

10


268<br />

269<br />

270<br />

271<br />

272<br />

273<br />

274<br />

275<br />

276<br />

277<br />

278<br />

279<br />

280<br />

281<br />

282<br />

283<br />

284<br />

285<br />

286<br />

287<br />

288<br />

289<br />

290<br />

291<br />

292<br />

293<br />

from the 12 CI users that participated in both studies. For all tests differences were considered<br />

significant when p < .05.<br />

2.5. Test-retest reliability<br />

The test-retest reliability was also assessed by computing the coefficient of determination (R 2 ) <strong>for</strong> the<br />

correlation between N1 peak amplitudes and latencies in the TNS (first test) and in the ESS (retest). It<br />

is worth noting that the studies were not planned as retests and there<strong>for</strong>e different stimuli and<br />

paradigms were used. In addition, since the participants had at least 6 months experience with the CI<br />

device prior to taking part in either of the studies reported here, we assumed that the auditory system<br />

would have re-organized to a substantial extent and AEPs would have been established (Pantev et al.,<br />

2006). Thus, we predicted that, given careful artifact attenuation, the AEP N1 response should be<br />

characterized by a reasonable temporal stability.<br />

3. Results<br />

3.1. Environmental Sounds Study (ESS)<br />

After running CIAC, the reconstructed AEPs were similar to the ones obtained after manual selection<br />

of CI artifacts. Figure 3 shows the individual AEPs <strong>for</strong> a fronto-central electrode, as well as the grand<br />

average AEP and the N1 and P2 peak topographies. No significant differences were found <strong>for</strong> the N1<br />

peak latency obtained after manual (MAN) and semi-automatic (S-AUTO) selection (t(17) = 1.27,<br />

n.s.). The P2 peak latency was also not significantly different (t(17) = 0.81, n.s.). When comparing the<br />

N1-P2 peak-to-peak amplitude, no significant differences were found (t(17) = -1.42, n.s.). Table 2<br />

shows a summary of these comparisons. Interestingly, CIAC revealed a mean sensitivity of 91.7 % ±<br />

0.12 and a mean specificity of 92.3 % ± 0.07, when its per<strong>for</strong>mance was compared to the manual<br />

selection. The correlation between age and N1-P2 peak-to-peak amplitude after the two types of CI<br />

artifact selection is shown in Figure 3-c. A significant negative correlation between age and peak-to-<br />

peak amplitude was found after automatic selection (AUTO: rs = -0.67, p = .002), replicating previous<br />

findings (MAN: rs = -0.70, p =.001) (Viola, et al., 2011).<br />

3.2. Tones and Noise Study (TNS)<br />

11


294<br />

295<br />

296<br />

297<br />

298<br />

299<br />

300<br />

301<br />

302<br />

303<br />

304<br />

305<br />

306<br />

307<br />

308<br />

309<br />

310<br />

311<br />

312<br />

313<br />

314<br />

315<br />

316<br />

317<br />

318<br />

319<br />

320<br />

After running CIAC, AEPs were reconstructed <strong>for</strong> the 12 CI users. The residual artifacts in this study<br />

were larger than when the same CI users were stimulated with environmental sounds. Figure 4 shows<br />

the AEPs <strong>for</strong> a fronto-central electrode be<strong>for</strong>e and after semi-automatic attenuation of CI artifacts <strong>for</strong><br />

the two studies (ESS, left and TNS, right). Age and N1 amplitude were systematically correlated in<br />

both studies (ESS: rS = -0.48, p =.12; TNS: rS = -0.52, p =.09) but failed to reach significance.<br />

In this study CIAC revealed a mean sensitivity of 87.3 % ± 0.13 and a mean specificity of 65.6 % ±<br />

0.10, when its per<strong>for</strong>mance was compared to the manual selection per<strong>for</strong>med by an independent<br />

expert. The specificity was lower than in the previous study. After investigating the properties of the<br />

ICs selected by the algorithm and the ones selected by the expert, it was found that the expert seemed<br />

to be more conservative, while the algorithm selected also noise related ICs affected by residual CI<br />

artifact.<br />

3.3. Test-retest reliability<br />

In order to investigate test-retest reliability, the coefficient of determination was computed <strong>for</strong> N1 peak<br />

amplitudes and latencies, as shown in Figure 5. In line with what was reported in previous studies<br />

where normal hearing individuals with a broad age range were assessed (Walhovd and Fjell, 2002), a<br />

high test-retest reliability was found <strong>for</strong> N1 latencies (rS = 0.77; R 2 = 0.59) and <strong>for</strong> N1 amplitudes (rS<br />

=0.69; R 2 = 0.48), suggesting that both parameters were reliably reconstructed in both studies.<br />

4. Discussion<br />

CIAC is a tool that semi-automatically selects ICs representing CI artifacts, thus aiming <strong>for</strong> a more<br />

objective and efficient attenuation of electrical CI artifacts and facilitating the reconstruction of AEPs<br />

in CI users. The algorithm was optimized taking into account known properties of the ICs representing<br />

CI artifacts. We also aimed at reducing the number of computational steps and thresholds to a<br />

minimum. As a result we consider the selection of CI artifacts with CIAC a quick, comprehensive<br />

procedure.<br />

The per<strong>for</strong>mance of CIAC was evaluated using a total of 30 EEG recording from 18 adult CI users<br />

using different types of CI devices, and which were stimulated in two studies with either<br />

environmental sounds or tones and noise, separated in time approximately one year. CIAC revealed a<br />

12


321<br />

322<br />

323<br />

324<br />

325<br />

326<br />

327<br />

328<br />

329<br />

330<br />

331<br />

332<br />

333<br />

334<br />

335<br />

336<br />

337<br />

338<br />

339<br />

340<br />

341<br />

342<br />

343<br />

344<br />

345<br />

346<br />

347<br />

high sensitivity and a good specificity when compared to the results of classification by two experts. It<br />

is worth noting that the ability of semi-automatically identifying CI artifact related ICs relies mainly<br />

on the general quality of the ICA decomposition, which depends on EEG preprocessing and other<br />

aspects not covered here (Debener, et al., 2010). The existence of ICs where the artifact is not well<br />

disentangled from brain activity (or other types of artifact) may present some challenges. It is not<br />

known how many sensors may be needed to accurately identify CI artifacts and it is also not well<br />

understood which components of the device contribute more to the artifact. There<strong>for</strong>e it would be<br />

beneficial to validate CIAC with other EEG montages and types of stimuli.<br />

It is important to highlight that CIAC benefits from experimental designs where the duration of the<br />

stimuli does not overlap with the responses of interest. This was not the case in the TNS data where<br />

the P2 responses were difficult to reconstruct <strong>for</strong> some CI users. Accordingly an experimental design<br />

where duration of the auditory stimuli is longer than the cortical response interval of interest enlarges<br />

the probability to reconstruct good quality AEPs. The use of short stimuli may prevent this type of<br />

issues. However this also limits the type of studies that can be implemented. It is expected that some<br />

of these current limitations may be overcome with the implementation of new ICA algorithms that<br />

may allow a better separation between artifacts and other sources. When it comes to the comparison<br />

between CIAC and experts, the results should be considered preliminary, since only two experts<br />

participated in the validation procedure. Given the large number of decisions necessary (approximately<br />

number of electrodes x number of individuals) it is likely that users show some degree of<br />

inconsistency, thus limiting the reliability of the resulting AEPs. Different experts may also apply<br />

different criteria. For instance, it is our experience that experts may ignore noise related ICs<br />

contaminated with residual CI artifact since these normally explain a small amount of variance in the<br />

AEPs. Moreover experts could be biased by their past experience, if they, <strong>for</strong> instance, only had<br />

experience with datasets collected from CI users using devices from a specific manufacturer or<br />

collected with a particular electrode montage. It is also worth noting that the number of researchers<br />

experienced with the selection of CI artifact related ICs is likely small, which hinders the wider use of<br />

AEPs <strong>for</strong> the assessment of auditory rehabilitation. Accordingly a comparison of CIAC with more<br />

13


348<br />

349<br />

350<br />

351<br />

352<br />

353<br />

354<br />

355<br />

356<br />

357<br />

358<br />

359<br />

360<br />

361<br />

362<br />

363<br />

364<br />

365<br />

366<br />

367<br />

368<br />

369<br />

370<br />

371<br />

372<br />

373<br />

374<br />

than two experts, as provided <strong>for</strong> the CORRRMAP plug-in <strong>for</strong> instance, would have not been feasible<br />

(Viola, et al., 2009).<br />

In terms of data quality, after semi-automatic selection of CI artifacts, AEPs with reasonable quality<br />

could be reconstructed. However <strong>for</strong> the TNS study the amount of residual artifact was larger than in<br />

the ESS. One reason could be the fact that in the TNS study the CI users per<strong>for</strong>med a passive listening<br />

task. For this and other reasons it is likely the resulting AEPs were of a lower signal-to-noise ratio,<br />

which may cause more difficulty <strong>for</strong> ICA in separating AEPs from artifact.<br />

When comparing manual and semi-automatic selection no significant differences were found between<br />

AEP N1 and P2 peak amplitudes and latencies <strong>for</strong> datasets from the ESS. We also found a high<br />

correlation between age and the N1-P2 peak-to-peak amplitude, corroborating previously reported<br />

results that the attenuation of the CI artifact does not eliminate potentially in<strong>for</strong>mative individual<br />

differences in the AEPs (Viola, et al., 2011). Age and N1 amplitude were also correlated in the sub-<br />

group of 12 CI users, showing that the individual differences between users were conserved<br />

independent of the type of stimuli. As this sub-group was evaluated at two different points of time<br />

after implantation, it was possible to evaluate the test-retest reliability of the AEPs. As predicted, a<br />

high test-retest reliability was observed <strong>for</strong> both N1 peak amplitude and latency. The amplitude<br />

reliability was comparable to values reported in a comparable study from young and old normal<br />

hearing listeners that were re-tested after one year (Walhovd and Fjell, 2002). Our reliability results<br />

accordingly suggest that by measuring standard AEP markers such as the N1 peak amplitude and<br />

latency, insights could be obtained about the rehabilitation state of the auditory system, in particular<br />

within the first few months after implant device switch on (Pantev, et al., 2006). We consider<br />

important that, despite the electrical artifact that can be orders magnitude larger than the auditory<br />

cortex response (e.g., Debener, et al., 2008), AEPs can be reliably reconstructed with ICA, and thus be<br />

used in the context of investigating auditory rehabilitation from CI users as proposed previously<br />

(Kileny, 2007; McNeill et al., 2009).<br />

Our work is in line with a broader field of research that has been moving towards the objective and<br />

automatic correction of artifacts in EEG recordings (Mognon, et al., 2010; Nolan, et al., 2010; Viola,<br />

14


375<br />

376<br />

377<br />

378<br />

379<br />

380<br />

381<br />

382<br />

383<br />

384<br />

385<br />

386<br />

387<br />

388<br />

389<br />

390<br />

391<br />

392<br />

393<br />

394<br />

395<br />

396<br />

397<br />

398<br />

399<br />

400<br />

401<br />

et al., 2009) using ICA as the main pre-processing method. The development of CIAC opens new<br />

doors in the use of EEG as a routine tool to assess auditory cortical function in CI users, since the<br />

types of auditory stimuli, as well as the experimental design do not need to be strongly conditioned in<br />

order to minimize CI artifacts. In the context of CI rehabilitation further research is needed to evaluate<br />

the value of late AEPs, <strong>for</strong> instance in helping with the CI fitting procedure, or in response to speech<br />

sounds in patient groups that cannot voluntarily report their electrical hearing experience, such as early<br />

implanted children (Kileny, 2007). Accordingly we envision that measures of auditory cortex function<br />

as assessed with late AEPs can be of use in CI configuration from initial setup to the long-term<br />

monitoring of rehabilitation progress. We fully agree with the view that developing the ability of the<br />

brain to learn how to use an implant may be as important as further improvements of CI device<br />

technology (Moore and Shannon, 2009). In this context CIAC may be an improvement, as it facilitates<br />

the investigation of auditory cortex functions in CI users.<br />

5. Conclusion<br />

The CIAC algorithm reported here provides a fast, user-friendly and objective method to correct<br />

electrical CI artifacts from AEP recordings. We hope that this freely available tool will support<br />

research investigating auditory cortex reorganization during CI adaptation and rehabilitation, since it is<br />

a significant step towards the objective and efficient study of late AEPs. As CIAC will be provided as<br />

an open source plugin to be used with the popular EEGLAB toolbox, we hope that other researchers<br />

will contribute to its further development, validation and, ultimately, its clinical application.<br />

Acknowledgements<br />

F.C.V. was funded by the Fundacao para a Ciencia e Tecnologia, Lisbon, Portugal<br />

(SFRH/BD/37662/2007). M.D.V. was supported by a Alexander von-Humboldt stipendium. P.S. was<br />

supported by the Swiss National Science Foundation (grant number PBZHP3-128462). The authors<br />

would like to thank A. Barks <strong>for</strong> assistance with recording the data and J.D. Thorne <strong>for</strong> helpful<br />

discussions.<br />

15


402<br />

403<br />

404<br />

405<br />

406<br />

407<br />

408<br />

409<br />

410<br />

411<br />

412<br />

413<br />

414<br />

415<br />

416<br />

417<br />

418<br />

419<br />

420<br />

421<br />

422<br />

423<br />

424<br />

425<br />

426<br />

427<br />

428<br />

429<br />

References<br />

Debener, S., Hine, J., Bleeck, S., & Eyles, J., 2008. Source localization of auditory evoked potentials<br />

after cochlear implantation. Psychophysiology. 45, 20-24.<br />

Debener, S., Thorne, J., Schneider, T. R., & Viola, F. C. (2010). Using ICA <strong>for</strong> the Analysis of Multi-<br />

Channel EEG Data. In M. Ullsperger & S. Debener (Eds.), Simultaneous EEG and fMRI (pp.<br />

121-135). New York: Ox<strong>for</strong>d University Press.<br />

Delorme, A., & Makeig, S., 2004. EEGLAB: an open source toolbox <strong>for</strong> analysis of single-trial EEG<br />

dynamics including independent component analysis. J Neurosci Methods. 134, 9-21.<br />

Delorme, A., Sejnowski, T., & Makeig, S., 2007. Enhanced detection of artifacts in EEG data using<br />

higher-order statistics and independent component analysis. Neuroimage. 34, 1443-1449.<br />

Friesen, L. M., & Picton, T. W., 2010. A method <strong>for</strong> removing cochlear implant artifact. Hear Res.<br />

259, 95-106.<br />

Gilley, P. M., Sharma, A., Dorman, M., Finley, C. C., Panch, A. S., & Martin, K., 2006. Minimization<br />

of cochlear implant stimulus artifact in cortical auditory evoked potentials. Clin<br />

Neurophysiol. 117, 1772-1782.<br />

Gilley, P. M., Sharma, A., & Dorman, M. F., 2008. Cortical reorganization in children with cochlear<br />

implants. Brain Res. 1239, 56-65.<br />

Gramann, K., Onton, J., Riccobon, D., Mueller, H. J., Bardins, S., & Makeig, S., 2010. Human brain<br />

dynamics accompanying use of egocentric and allocentric reference frames during navigation.<br />

J Cogn Neurosci. 22, 2836-2849.<br />

Henkin, Y., Tetin-Schneider, S., Hildesheimer, M., & Kishon-Rabin, L., 2009. Cortical neural activity<br />

underlying speech perception in postlingual adult cochlear implant recipients. Audiol<br />

Neurootol. 14, 39-53.<br />

Kelly, A. S., Purdy, S. C., & Thorne, P. R., 2005. Electrophysiological and speech perception<br />

measures of auditory processing in experienced adult cochlear implant users. Clin<br />

Neurophysiol. 116, 1235-1246.<br />

Kileny, P. R., 2007. Evoked potentials in the management of patients with cochlear implants: research<br />

and clinical applications. Ear Hear. 28, 124S-127S.<br />

16


430<br />

431<br />

432<br />

433<br />

434<br />

435<br />

436<br />

437<br />

438<br />

439<br />

440<br />

441<br />

442<br />

443<br />

444<br />

445<br />

446<br />

447<br />

448<br />

449<br />

450<br />

451<br />

452<br />

453<br />

454<br />

455<br />

456<br />

457<br />

Koelsch, S., Wittfoth, M., Wolf, A., Muller, J., & Hahne, A., 2004. Music perception in cochlear<br />

implant users: an event-related potential study. Clin Neurophysiol. 115, 966-972.<br />

Lonka, E., Kujala, T., Lehtokoski, A., Johansson, R., Rimmanen, S., Alho, K., & Naatanen, R., 2004.<br />

Mismatch negativity brain response as an index of speech perception recovery in cochlear-<br />

implant recipients. Audiol Neurootol. 9, 160-162.<br />

Makeig, S., Debener, S., Onton, J., & Delorme, A., 2004. Mining event-related brain dynamics.<br />

Trends Cogn Sci. 8, 204-210.<br />

McNeill, C., Sharma, M., & Purdy, S. C., 2009. Are cortical auditory evoked potentials useful in the<br />

clinical assessment of adults with cochlear implants? Cochlear Implants Int. 10, 78-84.<br />

Mognon, A., Jovicich, J., Bruzzone, L., & Buiatti, M., 2010. ADJUST: An automatic EEG artifact<br />

detector based on the joint use of spatial and temporal features. Psychophysiology.<br />

Moore, D. R., & Shannon, R. V., 2009. Beyond cochlear implants: awakening the deafened brain. Nat<br />

Neurosci. 12, 686-691.<br />

Nolan, H., Whelan, R., & Reilly, R. B., 2010. FASTER: Fully Automated Statistical Thresholding <strong>for</strong><br />

EEG artifact Rejection. J Neurosci Methods. 192, 152-162.<br />

Onton, J., Westerfield, M., Townsend, J., & Makeig, S., 2006. Imaging human EEG dynamics using<br />

independent component analysis. Neurosci Biobehav Rev. 30, 808-822.<br />

Oostenveld, R., & Oostendorp, T. F., 2002. Validating the boundary element method <strong>for</strong> <strong>for</strong>ward and<br />

inverse EEG computations in the presence of a hole in the skull. Hum Brain Mapp. 17, 179-<br />

192.<br />

Pantev, C., Dinnesen, A., Ross, B., Wollbrink, A., & Knief, A., 2006. Dynamics of auditory plasticity<br />

after cochlear implantation: a longitudinal study. Cereb Cortex. 16, 31-36.<br />

Sandmann, P., Eichele, T., Buechler, M., Debener, S., Jancke, L., Dillier, N., Hugdahl, K., & Meyer,<br />

M., 2009. Evaluation of evoked potentials to dyadic tones after cochlear implantation. Brain.<br />

132, 1967-1979.<br />

Sandmann, P., Kegel, A., Eichele, T., Dillier, N., Lai, W., Bendixen, A., Debener, S., Jancke, L., &<br />

Meyer, M., 2010. Neurophysiological evidence of impaired musical sound perception in<br />

cochlear-implant users. Clin Neurophysiol. 121, 2070-2082.<br />

17


458<br />

459<br />

460<br />

461<br />

462<br />

463<br />

464<br />

465<br />

466<br />

467<br />

468<br />

469<br />

470<br />

471<br />

472<br />

473<br />

474<br />

475<br />

476<br />

477<br />

478<br />

479<br />

480<br />

481<br />

482<br />

483<br />

Sharma, A., Dorman, M. F., & Kral, A., 2005. The influence of a sensitive period on central auditory<br />

development in children with unilateral and bilateral cochlear implants. Hear Res. 203, 134-<br />

143.<br />

Viola, F. C., Thorne, J., Edmonds, B., Schneider, T., Eichele, T., & Debener, S., 2009. Semi-automatic<br />

identification of independent components representing EEG artifact. Clin Neurophysiol. 120,<br />

868-877.<br />

Viola, F. C., Thorne, J. D., Bleeck, S., Eyles, J., & Debener, S., 2011. Uncovering auditory evoked<br />

potentials from cochlear implant users with independent component analysis.<br />

Psychophysiology, in press.<br />

Walhovd, K. B., & Fjell, A. M., 2002. One-year test-retest reliability of auditory ERPs in young and<br />

old adults. Int J Psychophysiol. 46, 29-40.<br />

Wilson, B. S., & Dorman, M. F., 2008. Cochlear implants: A remarkable past and a brilliant future.<br />

242, 3-21.<br />

Wong, D. D., & Gordon, K. A., 2009. Beam<strong>for</strong>mer suppression of cochlear implant artifacts in an<br />

electroencephalography dataset. 56, 2851-2857.<br />

Zhang, F., Anderson, J., Samy, R., & Houston, L., 2010. The adaptive pattern of the late auditory<br />

evoked potential elicited by repeated stimuli in cochlear implant users. Int J Audiol. 49, 277-<br />

285.<br />

Zhang, F., Hammer, T., Banks, H. L., Benson, C., Xiang, J., & Fu, Q. J., 2011. Mismatch negativity<br />

and adaptation measures of the late auditory evoked potential in cochlear implant users. Hear<br />

Res. 275, 17-29.<br />

18


484<br />

485<br />

486<br />

487<br />

488<br />

489<br />

490<br />

491<br />

492<br />

493<br />

494<br />

495<br />

496<br />

497<br />

498<br />

499<br />

500<br />

501<br />

502<br />

503<br />

504<br />

505<br />

506<br />

507<br />

508<br />

509<br />

510<br />

Figure captions<br />

Figure 1 – Properties of independent components (ICs). Column a) shows two ICs representing the<br />

cochlear implant (CI) artifact <strong>for</strong> user #1, implanted on the left side, (blue) and user #2, implanted on<br />

the right side (red). Column b) shows another two ICs representing brain related activity <strong>for</strong> the same<br />

CI users. The top row shows the ICs topographies and the respective residual variance (RV) in % after<br />

dipole fitting. 2-D projections of dipole location and orientation are indicated in black on top of the<br />

topographic maps. The middle row shows the ERP of each IC activation. Zero ms represents auditory<br />

onset and amplitude values are expressed in arbitrary units (a.u.). The bottom row shows the temporal<br />

derivative of the ERP <strong>for</strong> each IC. The time windows corresponding to the onset of the CI artifact are<br />

displayed in light grey, whereas the time window representing activity of interest (N1-P2 peaks) is<br />

displayed in dark grey. For each IC the ratio between the root mean square (RMS) amplitude in the<br />

onset/offset window and the RMS amplitude in the time window of interest was calculated (ratio IC).<br />

Figure 2 – Schematic flow chart of the cochlear implant artifact correction (CIAC) algorithm. The<br />

main inputs are the independent components (ICs) after and corresponding in<strong>for</strong>mation from dipole<br />

fitting, the duration of the auditory stimuli used and a time window of interest that should contain the<br />

auditory evoked responses. In the first step a sub group of ICs is selected based in the residual<br />

variance (RV) obtained after dipole modeling and using a pre-defined threshold (Thr_rv). In the<br />

second step the first temporal derivative of the ICs-ERP is computed and a ratio (ratio deriv.) between<br />

the root mean square (RMS) in the time window of the onset/offset of the artifact (derived from the<br />

duration of the auditory stimuli) and the time window of interest is calculated. The IC with the highest<br />

ratio deriv. is selected as a template and its topography map is correlated with all other ICs maps. In<br />

the third step ICs with a ratio deriv. or a correlation with the template larger than pre-defined<br />

thresholds (Thr_deriv and Thr_corr, respectively), are selected as representing the CI artifact.<br />

Figure 3 – Summary of the auditory evoked potentials (AEPs) <strong>for</strong> the Environmental Sounds Study<br />

(ESS). a) Comparison of single subject AEPs reconstructed after manual (black) and semi-automatic<br />

(red) attenuation of cochlear implant artifacts. A fronto-central channel is shown. b) Grand average<br />

AEPs after manual (black) and semi-automatic (red) CI artifact attenuation and respective N1 and P2<br />

19


511<br />

512<br />

513<br />

514<br />

515<br />

516<br />

517<br />

518<br />

519<br />

520<br />

521<br />

522<br />

523<br />

524<br />

525<br />

526<br />

527<br />

528<br />

529<br />

530<br />

531<br />

532<br />

533<br />

534<br />

535<br />

peak topographies. c) Correlation between age (years) and N1-P2 peak-to-peak amplitude (V) <strong>for</strong><br />

data corrected manually (black) and semi-automatically (red).<br />

Figure 4 – Comparison of auditory evoked potentials (AEPs) reconstructed <strong>for</strong> the Environmental<br />

Sounds Study (ESS) (column a) and <strong>for</strong> the Tones and Noise Study (TNS) (column b) be<strong>for</strong>e (top) and<br />

after (bottom) semi-automatic attenuation of cochlear implant artifacts. Grey circles represent the N1<br />

peak and grey bars represent residual cochlear implant artifact.<br />

Figure 5 – Test-retest reliability <strong>for</strong> the N1 peak amplitude and latency. Left: correlation <strong>for</strong> the N1<br />

amplitude expressed in V <strong>for</strong> the Tones and Noise Study (TNS) and the Environmental Sounds Study<br />

(TNS). Right: correlation <strong>for</strong> the N1 peak latency expressed in ms <strong>for</strong> the TNS and the ESS.<br />

20


Figure(s)<br />

Click here to download high resolution image


Figure(s)<br />

Click here to download high resolution image


Figure(s)<br />

Click here to download high resolution image


Figure(s)<br />

Click here to download high resolution image


Figure(s)<br />

Click here to download high resolution image


Table(s)<br />

Table 1. Cochlear implant users’ clinical profile<br />

Stimulation rate<br />

(Pulses/s/chan.)<br />

Sound level<br />

(dB SPL)<br />

Score<br />

(% correct)<br />

CI use<br />

(months)<br />

Sound<br />

coding<br />

strategy<br />

Age<br />

Implantation<br />

(months)<br />

Duration<br />

deafness<br />

(months)<br />

Age<br />

(years)<br />

CI<br />

user<br />

Processor<br />

Device<br />

CI side<br />

Gender<br />

900<br />

65<br />

93<br />

45<br />

ACE<br />

ESPrit 3G<br />

Nucleus<br />

622<br />

150<br />

Right<br />

Male<br />

55<br />

01<br />

720<br />

65<br />

90<br />

54<br />

ACE<br />

ESPrit 3G<br />

Nucleus<br />

704<br />

568<br />

Right<br />

Male<br />

63<br />

02<br />

1200<br />

80<br />

42<br />

33<br />

ACE<br />

Freedom<br />

Nuclues<br />

899<br />

265<br />

Left<br />

Male<br />

77<br />

03<br />

1200<br />

75<br />

93<br />

22<br />

ACE<br />

Freedom<br />

Nucleus<br />

627<br />

614<br />

Left<br />

Male<br />

54<br />

04<br />

1200<br />

75<br />

91<br />

78<br />

ACE<br />

ESPrit 3G<br />

Nucleus<br />

771<br />

531<br />

Left<br />

Female<br />

70<br />

05<br />

900<br />

65<br />

91<br />

46<br />

ACE<br />

ESPrit 3G<br />

Nucleus<br />

837<br />

693<br />

Right<br />

Male<br />

73<br />

06<br />

1200<br />

65<br />

62<br />

21<br />

ACE<br />

Freedom<br />

Nucleus<br />

718<br />

478<br />

Left<br />

Female<br />

61<br />

07<br />

250<br />

70<br />

83<br />

156<br />

SPEAK<br />

ESPrit 3G<br />

Nucleus<br />

572<br />

416<br />

Left<br />

Male<br />

60<br />

08<br />

900<br />

70<br />

66<br />

51<br />

ACE<br />

ESPrit 3G<br />

Nucleus<br />

896<br />

30<br />

Left<br />

Male<br />

79<br />

09<br />

3712<br />

75<br />

99<br />

35<br />

HiRes-S<br />

AB Auria<br />

ABS<br />

491<br />

251<br />

Left *<br />

Female<br />

43<br />

10<br />

1200<br />

70<br />

53<br />

80<br />

ACE<br />

CI24M<br />

Nucleus<br />

883<br />

631<br />

Left<br />

Female<br />

80<br />

11<br />

900<br />

60<br />

51<br />

28<br />

ACE<br />

Freedom<br />

Nucleus<br />

447<br />

447<br />

Left<br />

Female<br />

39<br />

12<br />

2656<br />

65<br />

99<br />

12<br />

HiRes-S<br />

AB Harmony<br />

ABS<br />

679<br />

403<br />

Left<br />

Female<br />

57<br />

13<br />

3712<br />

75<br />

97<br />

6<br />

HiRes-S<br />

AB Harmony<br />

ABS<br />

692<br />

452<br />

Right<br />

Male<br />

58<br />

14<br />

1200<br />

70<br />

88<br />

18<br />

ACE<br />

Freedom<br />

Nucleus<br />

682<br />

41<br />

Right<br />

Female<br />

58<br />

15<br />

884<br />

70<br />

48<br />

8<br />

HiRes-S<br />

AB Harmony<br />

ABS<br />

390<br />

378<br />

Left<br />

Female<br />

33<br />

16<br />

1515<br />

75<br />

93<br />

18<br />

FSP<br />

Medel Opus2<br />

Med El<br />

446<br />

230<br />

Left<br />

Female<br />

38<br />

18<br />

1515<br />

70<br />

97<br />

10<br />

FSP<br />

Medel Opus2<br />

Med El<br />

741<br />

417<br />

Right<br />

Female<br />

62<br />

19<br />

*Subject was implanted bilaterally. Device names and processor names according to manufacture’s labeling. Sound coding strategies legend: advanced combination encoders (ACE),<br />

spectral peak coding (SPEAK), high resolution with fidelity 120 (HiRes-S), fine structure processing (FSP). Score corresponds to the percentage correct on the Bam<strong>for</strong>d-Kowal-Bench<br />

(BKB) speech recognition test in quiet.


Table(s)<br />

Table 2<br />

Mean N1- P2 peak-to-peak- amplitude and N1 and P2 peak latencies <strong>for</strong> datasets where cochlear<br />

implant artifact related independent components (ICs) were selected by an expert (manual) and by<br />

CIAC algorithm (semi-automatic) <strong>for</strong> the Environmental Sounds Study (ESS). All results are presented<br />

as Mean ± 1 SD.<br />

ICs<br />

selection<br />

ESS N1-P2 peak-to-<br />

peak amplitude<br />

ESS Latency [ms]<br />

[V] N1 P2<br />

manual 9.0 ± 4.1 132.3 ± 13.7 244.9 ± 27.6<br />

semi-<br />

automatic 9.1 ± 4.1<br />

130.8 ± 13.2 244.7 ± 28.0

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!