Doi:10.1016/j.clinph.2007.11.177

Clinical Neurophysiology 119 (2008) 842–852 Non-provocative diagnostics of photosensitivity using visual evoked potentials Joost Vermeulen a,1, Stiliyan Kalitzin b,*, Jaime Parra c, Erwin Dekker c, Albert Vossepoel f, Fernando Lopes da Silva d,e a Quantitative Imaging Group, Department of Imaging Science and Technology, Faculty of Applied Sciences, Delft University of Technology, Lorentzweg 1, 2628 CJ Delft, The Netherlands b Medical Physics Department, Epilepsy Institutes of The Netherlands Foundation (SEIN), Lorentz de Haas Laboratory, Achterweg 5, 2103 SW Heemstede, The Netherlands c Department of Clinical Neurophysiology, Epilepsy Institutes of The Netherlands Foundation (SEIN), Lorentz de Haas Laboratory, Achterweg 5, 2103 SW Heemstede, The Netherlands d Center of Neuroscience, Swammerdam Institute for Life Sciences, University of Amsterdam, The Netherlands e Instituto Superior Te´cnico, Lisbon Technical University, Portugal f Biomedical Imaging Group Rotterdam, Erasmus MC – University Medical Center Rotterdam, The Netherlands Accepted 8 November 2007 Available online 7 February 2008 Objective: Photosensitive epilepsy (PSE) is the most common form of reflex epilepsy. Usually, to find out whether a patient is sensitive,he/she is stimulated visually with, e.g. a stroboscopic light stimulus at variable frequency and intensity until a photo paroxysmal response(PPR) occurs. The research described in this work aims to find whether photosensitivity can be detected without provoking a PPR.
Methods: Twenty-two subjects, 15 with known photosensitivity, were stimulated with visual stimuli that did not provoke a PPR. Usingan ‘‘evoked response representation'', 18 features were analytically derived from EEG signals. Single- and multi-feature classificationparadigms were applied to extract those features that separate best subjects with PSE from controls.
Results: Two variables in the ‘‘evoked response representation'', a frequency term and a goodness of fit term to a particular template,appeared to be best suited to make a prediction about the photosensitivity of a subject.
Conclusions: Evoked responses appear to carry information about potential PSE.
Significance: This result can be useful for screening patients for photosensitivity and it may also help to assess in a quantitative way theeffectiveness of medical therapy.
 2007 International Federation of Clinical Neurophysiology. Published by Elsevier Ireland Ltd. All rights reserved.
Keywords: Epilepsy; Photosensitivity; Visual evoked potentials; EEG Diagnosing photosensitivity ) usually involves provoking photo paroxysmal responses (PPRs) Photosensitive epilepsy (PSE) is the most common form with intermittent photic stimulation (IPS). If stimulation is of stimulus induced epilepsy ( stopped as soon as PPRs appear, there is a small risk of ). It is found in 5% of the adult epileptic inducing an actual seizure. Some patients, however, might patients and in 10% of all children with epilepsy refuse to have photic stimulation due to this small risk. In ). It is estimated that it occurs in approx- any case it would be desirable to develop an analytical imately 1 in 4000 of the population.
method that could predict whether the responses to IPS inthe sensitive patients would differ from those of normal subjects.
In this context one possibility would be to determine Corresponding author. Tel.: +31 235588248.
1 Tel.: +31 15 27 81416; fax: +31 15 27 86740.
whether some properties of the visual evoked potentials 1388-2457/$34.00  2007 International Federation of Clinical Neurophysiology. Published by Elsevier Ireland Ltd. All rights reserved.
doi:10.1016/j.clinph.2007.11.177 J. Vermeulen et al. / Clinical Neurophysiology 119 (2008) 842–852 would yield useful information in this respect. Research on ). A sub-threshold paradigm, i.e. without eliciting visual evoked potentials (VEPs) in these cases, however, is PPRs, was used. The stimuli were black and white, and rather sparse. Gokcay et al. showed that there were differ- red and green sinusoidal gratings. The essence of their find- ences in the visual evoked potentials of patients with two ings is that spatial contrast can play a dominant role in forms of epilepsy: childhood epilepsy with occipital parox- revealing the potential sensitivity of these patients to visual ysms (CEOP) and symptomatic occipital epilepsy (SOE) inputs. No attempt for prognostic classification, however, (The VEPs of both groups were also was reported in their paper.
found to be different from control subjects. This research Our approach focuses on quantifying features of the was done with IPS at 1 Hz, which in general does not pro- VEPs to IPS. A set of these features is then used to test voke a photo paroxysmal response (PPR).
whether photosensitive epileptic patients and non-sensitive In recent years, an increase of incidence of photosensi- control subjects can be discriminated. Our strategy is to use tive epilepsy, caused predominantly by increased exposure non-provocative stimulation that is the closest possible to to video games and television, has prompted researchers to the most common provocative protocol used clinically.
deepen the study of the pathophysiology of human photo- Accordingly, we use here responses to simple flashes (and not checker boards) delivered at the generally non-provoc- In our group (), we ative rate of 2 Hz.
found that photosensitive and control subjects could be dis- Instead of making an ad hoc selection of features that may criminated using the phase clustering index (PCI) of the be expected to yield a good classification, we started with a EEG, or MEG, recorded during intermittent light stimula- set of eighteen features, and afterwards we selected, those tion at various frequencies, before a PPR was elicited. Using that provide the best classification using statistical tests.
this stimulation paradigm, however, a PPR was ultimatelytriggered. The question we address in this work is whether it is possible to make a prediction about whether or not thepatient is photosensitive using non-provocative stimulation.
Porciatti demonstrated a defect in the contrast gain con- trol in a selected group of patients with idiopathic occipital Twenty-two subjects were studied. Information of the epilepsy compared to normal controls ( patients can be found in .There were three normal Table 1This table gives information about the patients Posttraumatic focal epilepsy Abbreviations for clinical responses: MS, myoclonic seizure; GTSC, generalized tonic–clonic seizures; EM, eyelid myoclonia; CPS, complex partial seizures;FC, febrile convulsions; AS, absence seizures; S, subclinica; U, uncomfortable sensation. Abbreviations for AED (Anti Epileptic Drug): LTG, lamotrigine;VPA, valproic acid; GBP, gabapentine; LEV, levetiracetam; OCB, oxcarbazepine; CBZ, carbamazepine; ESM, ethosuximide; CLB, clobazam; LZP,lorazepam. Abbreviations for syndromes: JME, juvenile myoclonic epilepsy; JAE, juvenile absence epilepsy; IGE, idiopathic generalized epilepsy; SGE,symptomatic generalized epilepsy; GEFS+, generalized epilepsy with febrile seizures plus; PSE, photosensitive epilepsy; TLE, temporal lobe epilepsy. The‘*' indicates that patients were photosensitive in the past, but not anymore at the time of the recording.
J. Vermeulen et al. / Clinical Neurophysiology 119 (2008) 842–852 subjects, with no familiar antecedents of epilepsy. Two ear fit of each trial is subtracted (‘‘detrending''). Our basic patients had focal epilepsy (one posttraumatic and one stimulus response model is: with lesional temporal lobe epilepsy), without evidence of RðtÞ ¼ N ðtÞ þ 1 photosensitivity, neither during several EEG recordings, nor on clinical grounds. In addition, two patients with idi- here R, N and S represent, respectively, the measurement, opathic generalized epilepsy (one with juvenile absence epi- the noise and the ‘‘true'' response signal. In the above rep- lepsy and the other with juvenile myoclonic epilepsy) did resentation, S(t) is the time-evolution factor while the coef- show some degree of photosensitivity in previous record- ficient ac quantifies the distribution of the response over the ings, but this condition was controlled with medication, EEG derivations and the coefficient ss represents possible such that no evidence of photosensitivity was found during trial-to-trial jitter in the responses. The latter can be due the day of the recording. These patients were considered to imprecise trigger localization or due to jitter caused by for the purpose of this study as non-sensitive and are indi- fluctuations of the EEG on-going activity. We assume fur- cated with an ‘*' in .
ther that the noise N has all the 3 dimensions: channels, The other 15 patients showed photo paroxysmal dis- stimuli and time. We have also assumed here that the jitter charges during appropriate intermittent photic stimulation.
can depend on the stimulus but not on the channels which Two patients had symptomatic generalized epilepsy, the limits our interpretation to non-propagating patterns.
others belonged to the group of idiopathic generalized epi- Notice that for all stimulation trials the same acS(t) rep- lepsies, mostly juvenile myoclonic epilepsy.
resents the stationary physiological response. To accountfor this, the symbol 1s " [1, 1, . . , 1] is introduced.
In short, Eq. assumes that the evoked response of a 2.2. Data acquisition patient is stationary and that the system resets itself aftereach response.
Most data were acquired at SEIN using a LaMont 32 The objective of the following pre-processing phase was channel amplifier and Harmonie 6.0 software (Stellate Sys- to determine or estimate the individual factors in In the tems, Montreal). The international 10/20 system was used pre-processing phase several techniques were used as for the electrode positions and the EEG data were sampled described in Sections Different strategies at 800 Hz. Some data were obtained from the VUMC (Free were followed in order to find an optimal representation of University Medical Center in Amsterdam) magneto the response signal for our classification objective.
encephalography (MEG) center. The VUMC data wereacquired at different sampling rates, varying from 250 Hzto 625 Hz. To overcome this inequality, all data were 2.3.1. Alignment procedure downsampled to 250 Hz.
Our first optional strategy was to find an adequate esti- VEP investigation was carried on by means of a com- mation for the ‘‘misalignments'' ss in To account for mercially available stroboscopic photic stimulator (Grass possible misalignment a computational scheme was devel- PS33). The patients were seated on a chair in front of the oped ) based on estimating the corre- stroboscopic stimulator that was placed at 30 cm from lation functions between the evoked responses. Here an their eyes. The recordings obtained in the MEG center alternative alignment procedure was applied that is also were performed with a stimulator specifically designed to correlation-based but less computationally expensive. The project images in the shielded MEG environment procedure can be found in .
In the MEG center, EEG data were measured The ss found here can be inserted in formula to compen- in addition to MEG data. Only the EEG data were used for sate if required for the initial misalignment.
this analysis.
In order to record visual evoked potentials (VEPs) the 2.3.2. Taking the mean over trials patients were stimulated with a flashing light at 2 Hz and To further reduce the complexity of our data and an analog trigger signal was recorded synchronously with according to the assumption that the terms acS(t) are the the EEG traces; only the first 333 ms of the evoked actual evoked responses per channel, we average over responses was used.
all trials to obtain P cðtÞ  hRðtÞ i ¼ hN ðtÞ i þ h1 sSðt þ ssÞis c 2.3. Signal representation ¼ acSðtÞ þ g ðtÞ The interesting parts of the EEG recordings for this Here, the notation Æ æa is used to indicate averaging. In the study are the sequences immediately after the stimulus.
last formula the left-hand side defines the channel-depen- Using the trigger timing information the data are cut and dent triggered response, the average gc(t) " ÆN(t)csæs is the arranged into a 3D response ‘‘cube'': channels (c) · stimu- lus number (s) · time after stimulus (t), R(t)cs. For each h1sSðt þ ssÞi  SðtÞ has been applied. In the case when channel one can select the response to any individual stim- no jitter is assumed (ss " 0) we have simply SðtÞ ¼ SðtÞ. In ulus; this is called a trial. After this procedure, the best lin- the sequel we use the notation S(t) both for the aligned

J. Vermeulen et al. / Clinical Neurophysiology 119 (2008) 842–852 and the non-aligned estimations of the response's shape tion obtained with PCA (green line). From the red line it appears as if nothing was measured. The green line shows In addition to the averaged response , the response that with the PCA weights, the shape of the signal is statistical properties can be quantified by the phase cluster- ing index (PCI) ().
Explicitly, 2.3.4. Representation approaches for response classification In this work four different approaches for the quantifica- tion of the evoked potential (EP) data are pursued. They consist of combinations of what has been described in Sec- tions . The text in bold is used to designatea specific representation for future reference.
F ðxÞ ¼ FT fRðtÞ g 1. Basic representation. Only the procedure of Section is the Fourier transform of the signal. We consider here the is performed. For each occipital channel (O1 and overall PCI as in formula 3 with the following channel- O2) the mean is taken over all responses. This yields 2 evoked responses per patient, one for each occipital chan- nel. The subsequent features (Section ) are calculated separately and averaged at the end.
2. PCA representation. Like the first method, but now the resulting signals of the different channels are pro- 2.3.3. Principal component analysis between the channel cessed according to Section . This representation yields one evoked response (the principal component) Since the EEG is measured at several positions on the per patient.
head, one expects the amplitude to differ between channels 3. Aligned basic representation. The same as the basic resulting in factors ac different for each channel. To sepa- representation but after applying the alignment procedure rate the common response shape S(t) from the weight fac- described in Section .
tors ac in formula we can use the singular value 4. Aligned PCA representation. The fourth and most decomposition technique that minimizes the total qua- computational expensive method involves all three pre-pro- dratic variation of the channel-specific noise gc(t). This technique uses the covariance matrix C ¼ covtðP cðtÞÞ Selecting the eigenvector Vc corresponding to its largesteigenvalue we define the principal component (PC) The remaining components, orthogonal to may then beinterpreted as ‘‘noise''. From formula it can be seen thatthe factor ac can be calculated as follows: Finally, to give an estimate of the three-dimensional noise,we write in the form N ðtÞ ¼ RðtÞ  1 t0 P cðt0Þ  Sðt0Þ t0 Sðt0Þ  Sðt0Þ In the above formula all quantities are known from the Fig. 1. Demonstration of the effect of the PCA procedure. There are two steps described in Sections . We assume artificial ‘‘signals'' (blue lines), their mean (red line) and the representation obtained with PCA (green line). In this particular case, it can be seen that SðtÞ  SðtÞ  1. Given the Pythagorean theorem, this two signals with the same shape, but with only a different scaling, are means that all spatial amplitude variation can be found considered to be almost the same. The red line shows the mean of the two in the channel weights a.
signals. According to the red line, nothing or very little was measured.
Maintaining the shape is beneficial when for example one measures the In the effect of what has been described in Section same signal at opposite sides of a dipole. (For interpretation of the is demonstrated. The image shows two artificial ‘‘sig- references to colour in this figure legend, the reader is referred to the web nals'' (blue lines), their mean (red line) and the representa- version of this article.)

J. Vermeulen et al. / Clinical Neurophysiology 119 (2008) 842–852 2.4. Quantification of the EP After obtaining a representation of the EP, we proceed with quantifying the features of the signal. In the previouschapter we have reduced the response signal complexity,according to the chosen representation from Sectionto several lower dimensional objects. For representa-tions 2 and 4 for example, we are left with the responsefunction , weights , jitter delays ss and the quantitiesPCI(x) defined in . In representations 1 and 3 theresponses are kept separate for the individual channelsbut the processing is the same. For these two representa-tions, the features per channel are averaged.
To compare different responses, additional data reduc- tion is useful. In an explanationof features that are used is given. To simplify the notations,the proposed features are denoted by: (f.n) = {feature n,‘‘n'' stands for the number of the feature}.
Fig. 2. Illustration of a response and the corresponding response function.
Features were derived from the representation described The blue line shows the measured response. The red line shows the result in Section an envelope function and from a response of the parametrical fit. In this case a 2nd order fit was applied. (For interpretation of the references to colour in this figure legend, the reader isreferred to the web version of this article.) To obtain the envelope function, we first compute the analytic signal.
dure to obtain the parameters of the second order fit. H ¼ SðtÞ þ iHilbertðSðtÞÞ shows a complete list of the features that are used in the The absolute of the analytic signal is the envelope function.
Because of the uncertainties intrinsic to the source local- AðtÞ  jH ðtÞj ization inverse problem, as well as questions concerning This function has no zero crossings and is therefore well parameters such as unequal contact impedances, we have suited for computing features such as moments.
not attempted to extract any information related to the The response model is computed by trying to minimize spatial distribution of the response. Therefore we have the distance between the actual response from Section ignored the information contained in the coefficients ac and a model function by means of which it is from formula .
attempted to accurately describe the response.
The response model, k, is: 2.5. Classification kðtÞ ¼ ekt  sinð2  p  m  tÞ 2.5.1. Introduction where m represents the frequency and k is a measure for In this section we explain how the above quantities are damping. Note that the frequency m has units Hertz be- used to classify signals of different conditions. In this work cause of the 2p factor in k(t).
the statistical power of the VEP features to discriminate We assume that shape is more important than the over- photosensitive from non-sensitive subjects is analyzed.
all scale factor, hence we quantify the goodness of fit by thedistance function D: 2.5.2. Kolmogorov–Smirnov test ðSðtÞ  SF  kðtÞÞ2 The Kolmogorov–Smirnov test ) is used to compare the distributions of feature values of sensitiveand non-sensitive subjects. This particular test has the advantage that a priori no assumptions need to be made about the distributions.
and the dot denotes a scalar product. A direct parametersearch method is used to minimize D, i.e. a range of values 2.5.3. One-dimensional classification for k and m are tested. m values were tested between 0.25 and To further investigate the features, the discriminatory 100 Hz in steps of 0.25 Hz; k values were tested between information of features is tested directly. This was done 0.1 and 0.1 in steps of 0.001. illustrates an example using simple one-dimensional threshold classifiers where of an actual response and its model.
the values above a certain threshold are considered to Note that this can be iterated several levels deep. One belong to one class and those below the threshold to the can subtract k(t) from the response and repeat the proce- other. The classifier selects an optimal threshold for each J. Vermeulen et al. / Clinical Neurophysiology 119 (2008) 842–852 Table 2A list of the 18 features that are used in the analysis Average time of the response Variance of the signal Misalignment of the response PCI of the driving frequency Maximal phase clustering index Frequency corresponding to f.8 PCI weighted mean frequency Relative phase clustering index First order damping factor of the response model Second order damping factor of the response model First order frequency of the response model Second order frequency of the response model Distance between the first order response model and the evoked response Distance between the second order response model and the evoked response feature separately so that the number of misclassified sub- In this case also a threshold is set for Ca to minimize the jects is minimal.
Although the voting scheme and linear classifier formu- 2.5.4. Feature combinations: voting scheme las look the same, they are different with respect to how the Using combinations of features may yield better classifi- features are interpreted. The voting scheme translates the cation results. One simple method to combine the discrim- values of the features to a yes–no (0 1) vote per feature.
ination power of more than one quantity is to perform In case of the linear classifier the values of the features independently a classification according to each ‘‘member'' are used to build a feature space in which it is attempted and then to postulate certain voting rules for a final to separate the photosensitive from the control subjects.
If Via is the vote of feature i on patient a and Wi is a 2.6. Statistical significance of the classification weight of feature i, then for Ca, the probable classificationof patient a we have: 2.6.1. Determining a significance threshold For any classification approach, the significance level is determined by the probability that the same results can be achieved by a ‘‘random guess'' procedure. We require in ia contains the votes (0 or 1) of each feature on each patient based on the optimal threshold classification de- this work statistical significance that limits the latter prob- scribed in Section and A is a constant representing ability to at most 5%.
possible bias in the classification.
To make sure that the chance is <5%, that the results are The weight coefficients W as good as random, 10,000 classification attempts on ran- i and the bias term A are deter- mined to minimize the overall classification error, as dom data were done with a classifier. Since 10,000 classifi- explained in .
cation attempts are done, one can assume that any result the error of which ranks among the best 500 results, gives and A determined, Ca can be calculated.
Finally, we find the common threshold for the class identi- <5% chance to be random. Therefore the system is allowed to make as many errors as the 500th best attempt. This a that minimizes the classification error.
gives a distinct maximum allowed error. This maximum 2.5.5. Feature combinations: linear classifier allowed error is called error(max).
The results with a linear classifier are also used in this work. If, as in the previous paragraph, Wi is some weight 2.6.2. Procedure to find good feature combinations of feature i and Qia is the actual value of feature i on To find the good feature sets, all possible 2D, 3D, 4D patient a, then we postulate and 5D combinations of the 18 ) features are tested. This is done with both the voting scheme and the linear classifiers. From these feature sets a selection is made based on the significance from Section .
To determine the optimal weights Wi and bias A, we refer First, the size of the feature space is determined by look- ing at the error(max) from Section and the error of With Wi and A determined, Ca can be calculated.
the best feature sets for each feature space size. The error J. Vermeulen et al. / Clinical Neurophysiology 119 (2008) 842–852 of this best set needs to be as low as possible. In addition, the results need to be significant.
The training and testing here and in Section are done on the same sets. This is no problem as this test serves 2.9. Robustness check only to determine the size of the feature space and toremove feature sets that do not perform in accordance with To get an impression about the consistency of our the minimum level that is set by error(max).
results, multiple measurements of subjects are used. Of Once the proper dimensionality of the feature set is two subjects, one of which is photosensitive, there were determined, all combinations that have an error <error two data sets. These subjects will be referred to as subjects (max) are used for further investigation.
A and B. From each of these two subjects, the measured The results of this procedure are feature sets of proper evoked responses are inspected visually. Additionally, it dimensionality that can separate the two classes with a cer- is verified that both measurements appear near each other tain degree of significance.
in the feature space.
2.7. Cross-validation tests To find which feature sets yield the best classification 3.1. Kolmogorov–Smirnov test results, two tests are applied. First a leave-one-out cross-validation test is done. Leave-one-out cross-validation In the probability is given, involves using a single observation from the data set as for each feature (f.n), that the distributions of sensitive and the test data, and the remaining observations as the train- non-sensitive subjects are the same. This is done for the ing data. This is repeated until each observation in the sam- four different representations.
ple is used once as test data. The errors of all repetitions are shows that the align- Additionally, a hold-out validation test is done. In the ment procedure is not very beneficial. This suggests that hold-out test a set of measurements is randomly split in a the alignment destroys discriminatory information and/or training (70%) and a testing (30%) part following prior probabilities. The hold-out cross-validation procedure is repeated 50 times and the results are averaged.
The best results can be seen in .
Only those feature combinations that have significant In this case the basic representation is used.
classification performance and of proper dimensionality The figures show clearly that the distributions of feature (see Sections are used in this test.
f.15 (m) for the two groups of subjects are very different.
This is also the case for features f.8, f.11, f.17 and f.18.
2.8. Inspection of relevant features From this point onwards we consider only the EP repre- sentations where alignment is not involved.
The best sets from the cross-validation tests are inspected taking into account the occurrence of specific fea- 3.2. One-dimensional classification tures and feature combinations. The feature sets that had across-validation error <25% are further considered. The In 1D classification the simple threshold from Section percentage (25%) is chosen arbitrarily.
is used. The error for f.15 is 13.6% (=3/22, cf. Additionally the best features are studied using ROC This is when the basic representation is (Receiver Operating Characteristic) curves. ROC curves used. The result of the method that involves PCA appears a are used to quantify the performance of a classifier in terms little less favorable. In this case, f.8, f.15 and f.17 were the of specificity and sensitivity. The last are determined as best features. Each of them had an error of 18% (=4/22 cf. ). The Kolmogorov–Smirnovtest also gave some indication of this result. Notice that all errors are fractions of 22, i.e. 1/22, 2/22, etc.
3.3. Procedure to find good feature combinations FP: number of False Positive classifications This procedure is described above in Sections TP: number of True Positive classifications FN: number of False Negative classifications The results in were obtained TN: number of True Negative classifications with the voting scheme and the linear classifier. shows the results of the basic representation Notice that TP + FN represents all subjects that are with the voting scheme. shows the actually positive (photosensitive) and the TN + FP repre- results of the basic representation with the linear classifier.

J. Vermeulen et al. / Clinical Neurophysiology 119 (2008) 842–852 and show equivalent data for more features yield better the PCA representation. These tables only show the best performing feature sets in case that no training/testing split It can be seen that the use of the basic representation is done. As mentioned in Section the tables serve only ) yields better results than to determine appropriate dimensionality.
the PCA representation ( The ‘*' in the tables indicate that more than one feature Previous tests also demonstrated better performance of the combination with similar performance was found. In basic representation. Therefore the PCA representation is the best results are obtained no longer considered.
with 3 or more features; the results do not improve when The results also show that the linear classifier clearly including more features. This means that 3 features can outperforms the voting scheme classifier.
contain the most relevant information. In Like earlier using the Kolmogorov–Smirnov test, f.15, the best results are achieved with S2 or more fea- but also f.18, appears to be important. This is indicated tures. shows that the results with by the fact that they show up in the best performing feature the linear classifier on the basic representation do not combinations. This can also be concluded from the weights improve with more than two features. As expected when they have in combinations in case the voting scheme classi- training and testing are done on the same set, in all Tables fier is used. The weights from the linear classifier are not Fig. 3. The EPs of subjects A and B are shown in (a) and (b). The blue and red line indicate the result of two distinct measurements. The twomeasurements of each subject look very similar to the human eye. It can be seen in the 2D and 3D feature spaces of (c) and (d) that the data sets of thesame subject are positioned near each other. Blue crosses represent control subjects, red asterisks represent photosensitive patients. (For interpretation ofthe references to colour in this figure legend, the reader is referred to the web version of this article.) J. Vermeulen et al. / Clinical Neurophysiology 119 (2008) 842–852 considered, as they hold scale information in addition to binations feature f.15 is used 59 times (see weight information.
). Other frequently occurring features are again Given the information from above it seems reasonable f.8, f.17 and f.18. They are used 23, 24 and 23 times, and safe to focus on finding good 3D feature combinations using the linear classifier on the basic representation.
show graphs of the feature combinations that performed well according to 3.4. Cross-validation our analysis (see, e.g. Cross-validation tests are done according to Section 3.5.3. Sensitivity and specificity Although 161718 ¼ 816 different unique 3D combinations ROC curves of f.15, f.18 and the combination of f.15– can be made with 18 features, only those feature combina- f.18 are shown in . The features tions that have significant classification performance are f.15 and f.18 are combined using the weights of the linear used (see Section ). With the linear classifier used on the basic representation there were 206 3D sets with signif- f :ð15:=18Þ ¼ wðf :15Þ  f :15 þ wðf :18Þ  f :18 icant classification performance. As mentioned in Section, the leave-one-out and hold-out tests are done using where w(f.15)  0.29 and w(f.18)  0.71. This information the linear classifier on the basic representation.
cannot be used to directly draw conclusion about the rela-tive importance of features. To do that scale information 3.4.1. Leave-one-out cross-validation must be incorporated. It can be seen in, e.g. shows the average classifica- that f.15 and f.18 have different scales.
tion error with the leave-one-out cross-validation test.
As expected, the best sensitivity/specificity combination The three best sets had no errors. The corresponding is obtained when f.15 and f.18 are combined.
feature sets are [18 15 1], [18 15 7] and [18 15 14]. shows the best 10 sets and their errors.
3.6. Robustness check 3.4.2. Hold-out validation This check, described in 2.9, gives an impression of shows the average classifica- the consistency of the measurements and more impor- tion errors and their standard deviations of the hold-out tantly the position of a subject in the feature space. Only validation test.
the data of the basic representation are taken into The best set here is [18 16 15], it had an average error account here.
5%. shows the best 10 sets.
and b show the evoked responses of subjects A Investigation of the best feature sets has shown that they and B. The blue and red line indicate the result of two dis- all rely on features 15 (m) and 18 (second order D). Note tinct measurements. Especially the two measurements of that other tests already indicated the usefulness of f.15 subject B look very similar to the human eye, although this is less obvious for subject A. More relevant though are thepositions of the two measurements of these two subjects in 3.5. Most relevant features the relevant feature space. It can be seen in the 2D and 3Dfeature spaces of c and d that the data sets of the As described in Section , we focus here on the occur- same subject are positioned not far from each other, espe- rence of different features, in those feature sets that pro- cially in the case of B.
duce an error smaller than 25% in the cross-validationtests.
3.5.1. Best feature sets of the leave-one-out test The results of this study indicate that it is possible, using 89 feature combinations produced an error <25%. In non-provocative visual stimulation, to predict from the the error fractions of the best fea- visual evoked potentials whether or not a subject is ture combinations are shown. In these 89 best combina- tions feature f.15 is used 63 times (see Different approaches to represent the evoked response Other frequently occurring features are f.8, were tested. The good performance of the basic and PCA f.17 and f.18. They are used 25, 27 and 25 times, representations suggests that the alignment destroys dis- criminatory information. The basic representation givesthe best results.
3.5.2. Best feature sets of the hold-out out test Combinations of features were found that could be used Seventy-six feature combinations produced an error to separate the sensitive and non-sensitive subjects without <25%. In the error fractions of provoking a seizure. The most important feature of the the best feature combinations are shown. The results are response appears to be f.15 (m, the first order frequency similar to the results of Section . In these 76 best com- from the response model).

J. Vermeulen et al. / Clinical Neurophysiology 119 (2008) 842–852 As mentioned before, appropriate linear combination of features f.15 and f.18 can separate EPs from sensitive fromthose from non-sensitive subjects and therefore might be agood candidate for practical, non-provocative diagnostics.
This is illustrated in and is also shown in In conclusion the method described here can be applied in clinical practice in order to test photosensitivity inpatients without using relatively high frequencies of inter-mittent photic stimulation, and thus avoiding discomfortfor the patient and increasing the safety of the stimulationprotocol.
We thank the Epilepsy Institutes of The Netherlands Foundation (SEIN) for providing the EEG facilities and Fig. 4. This illustration shows the usefulness of features 15 and 18. A clinical data. The assistance and helpful comments of green diagonal line is drawn to demonstrate that the two classes can be Demetrios Velis and Wouter Blanes were much appreciated.
separated perfectly. Additionally two blue lines are drawn to demonstratethe thresholds of the voting scheme. The data from the basic represen- We are thankful for the help and assistance of the col- tation are used here. The data are not scaled. Blue crosses represent leagues at the Quantitative Imaging group of the TU Delft.
control subjects, red asterisks represent photosensitive patients. (For The helpful comments, suggestions and feedback during interpretation of the references to colour in this figure legend, the reader is progress meetings and research discussions were of great referred to the web version of this article.) benefit to this project.
These results, although based on a relatively small sam- Appendix A. Supplementary data ple, show that it is possible to discriminate between photo- Supplementary data associated with this article can be sensitive and non-photosensitive subjects. Nevertheless additional data would be necessary to make meaningful learning curves.
The main result was obtained using a weighted voting system what is very basic, since no relation between classi- fiers is assumed. According to this approach differentparameters are taken individually and a classification is Gokcay A, Celebisoy N, Gokcay F, Ekmekci O, Ulku A. Visual evoked obtained based only on that information.
Additionally we tested a linear classifier. The scatter plot Jirsch JD, Urrestarazu E, LeVan P, Olivier A, Dubeau F, Gotman J.
in clearly shows that a combina- High-frequency oscillations during human focal seizures. Brain tion of f.15 and f.18 (can be used to separate the two classes with a linear classifier.
Harding GFA, Jeavons PM. Photosensitive epilepsy. London: Mac Keith Press; 1994.
Although additional features may improve the perfor- Kalitzin S, Parra J, Velis D, Lopes da Silva F. Enhancement of phase mance, it is clear that the two features f.15 and f.18 are clustering in the EEG/MEG gamma frequency band anticipates mainly responsible for good classification. Of these two, transitions to paroxismal epileptiform activity in epileptic patients f.15 appears to be more important. The interpretation is that the discrimination of photosensitive subjects has a Kalitzin S, Zbijewski W, Parra J, Velis D, Manshanden L, Lopes da Silva F.
strong frequency dependency. Indeed the importance of Correlation-based alignment of multichannel signals and application to f.15 in this respect is in accordance with the hypothesis that paroxysmal events. IEEE Trans Biomed Eng 2002b;49(9):1068–70.
high frequencies are associated with the transition between Kasteleijn-Nolst Trenite´ DGA. Photosensitivity in epilepsy. Electrophys- normal and seizure activity Massey Jr FJ. The Kolmogorov–Smirnov test of goodness of fit. J Am Stat The interpretation of feature f.18 is less straightforward Ass 1951;46(253):68–78 (The Matlab implementation was used for the since it represents the second order distance between the response model and the measured evoked response. The Parra J, Kalitzin S, Iriarte J, Blanes W, Velis D, Lopes da Silva F.
value of f.18 tends to be lower for control subjects. From Gamma-band phase clustering and photosensitivity: is there an this we can conclude that response model function underlying mechanism common to photosensitive epilepsy and visualperception? Brain 2003;126:1164–72.
describes more closely the responses of control subjects Parra J, Kalitzin S, Lopes da Silva F. Photosensitivity and visually than responses of photosensitive subjects.
induced seizures. Curr Opin Neurol 2005;18(2):155–9.
J. Vermeulen et al. / Clinical Neurophysiology 119 (2008) 842–852 Porciatti V, Bonanni P, Fiorentini A, Guerrini R. Lack of cortical contrast Rubboli G, Parra J, Seri S, Takahashi T, Thomas P. EEG diagnostic gain control in human photosensitive epilepsy. Nat Neurosci procedures and special investigations in the assessment of photosen- sitivity. Epilepsia 2004;45(s1):35–9.
Le Van Quyen M, Khalilov I, Ben-Ari Y. The dark side of high-frequency Wilkins A, Bonanni P, Porciatti V, Guerrini R. Physiology of human oscillations in the developing brain. Trends Neurosci 2006;29(7):419–27.

Source: http://qi.tnw.tudelft.nl/fileadmin/Faculteit/TNW/Over_de_faculteit/Afdelingen/Imaging_Science_and_Technology/Research/Research_Groups/Quantitative_Imaging/Publications/List_Publications/doc/Non_provocative_diagnostics.pdf