Seizure Detection Using Seizure Probability Estimation: Comparison of Features Used to Detect Seizures LEVIN KUHLMANN,1 ANTHONY N. BURKITT,1,2 MARK J. COOK,2,3 KAREN FULLER,3 DAVID B. GRAYDEN,1,2 LINDA SEIDERER,3 and IVEN M. Y. MAREELS1 1Department of Electrical and Electronic Engineering, The University of Melbourne, Parkville, VIC 3010, Australia 2The Bionic Ear Institute, 384-388 Albert St., East Melbourne, VIC 3002, Australia and 3Department of Neurology, St. Vincent���s Hospital Melbourne, Fitzroy, VIC 3065, Australia (Received 6 February 2009 accepted 29 June 2009 published online 10 July 2009) Abstract���This paper analyses seizure detection features and their combinations using a probability-based scalp EEG seizure detection framework developed by Marc Saab and Jean Gotman. Our method was evaluated on 525 h of data, including 88 seizures in 21 patients. The individual performances of the three features used by Saab and Gotman were compared to six alternative features, and combinations of these nine features were analyzed in order to find a superior detector. On a testing set with the combination of their three features, Saab and Gotman reported a sensitivity of 0.78, a false positive rate of 0.86/h, and a median detection delay of 9.8 s. Based on 10-fold cross-validation the testing performance of our implemen- tation of their method achieved a sensitivity of 0.79, a false positive rate of 0.62/h, and a median detection delay of 21.3 s. A detector based on an alternative combination of features achieved sensitivity of 0.81, a false positive rate of 0.60/h, and a median detection delay of 16.9 s. By including filtering techniques, it was possible to achieve performance levels similar to Saab and Gotman using our implementa- tion of their method, although this involved increases in detection delays. Of the seizure detection measures investi- gated, relative average amplitude, relative power, relative derivative, and coefficent of variation of amplitude provided the best performing combinations. These better-performing features can be employed together to make robust and reliable seizure detectors. Keywords���EEG, Epilepsy, Seizure detection, Seizure onset. INTRODUCTION Automated epileptic seizure detection methods, involving the analysis of electroencephalographic (EEG) signals, have been developed with mixed results. For scalp EEG, the applications for seizure detection include automatic labeling of seizures in the EEG for faster reporting33,35,39 and activating radioactive tracer injection for improved seizure focus localization using SPECT.7,42 Application of these methods to scalp EEG has been a difficult task given the many artifacts found in scalp EEG,16,39 and this has meant that automated seizure detection is not widely used in standard clinical practice. Examples of artifact include line noise, elec- tromyographic (EMG) signals, phase reversals and impedance changes due to loose electrodes, and amplifier dropout. Future work will determine whether seizure detection can be reliably applied to scalp EEG in the clinical environment. Most seizure detectors are based on classification algorithms that vary in complexity. The simplest are direct feature-based detectors13���15,20,24,30,31,45 that involve calculation of a feature from the EEG and then thresholding the feature to make detections. Features can be calculated either directly from the EEG or from filtered EEG signals, where usually either finite- impulse response (FIR), infinite-impulse response (IIR), or wavelet filters are used.4,5,34,36 More complex seizure detectors1,10,11,18,39,42,43,46 involve classification algorithms6 that are trained on combinations of fea- tures calculated from the EEG or its filtered signals. The output of the classifier is thresholded to make detections. Classifiers commonly applied to seizure detection include neural networks,1,10,11,43 support vector machines,42 Bayesian classifiers,18,39 and genetic algorithms.8,9 Direct feature-based seizure detectors are typically more transparent and easier to understand than com- plex classification algorithms that analyze the joint space of several features. These complex methods tend to be more mathematically and computationally intensive, and can sometimes hide the ���rules��� they are using to detect seizures.6 The advantage of the more complex methods is that they can be trained to detect Address correspondence to Levin Kuhlmann, Department of Electrical and Electronic Engineering, The University of Melbourne, Parkville, VIC 3010, Australia. Electronic mail: levink@unimelb. edu.au Annals of Biomedical Engineering, Vol. 37, No. 10, October 2009 (�� 2009) pp. 2129���2145 DOI: 10.1007/s10439-009-9755-5 0090-6964/09/1000-2129/0 �� 2009 Biomedical Engineering Society 2129

seizures based on combinations of several feature values. This is useful when single features cannot be used to detect all seizures from all patients, which is often the case. It is important to note that while clas- sifiers can be trained to detect large proportions of training seizures, there can be over-fitting of the training data and the classifier can perform poorly on out-of-sample testing data.6 This over-fitting is more likely to be a problem for more complex classifiers. For direct single-feature-based seizure detectors to work in all cases, the features have to be made more compli- cated and patient-specific.20,31,30 This paper provides an analysis of seizure detection features by employing the classifier-based framework developed by Saab and Gotman,39 which involves the detection of seizures in scalp EEG via the estimation of seizure probability. By incorporating Bayes��� law2 to determine the probability of having a seizure given feature values calculated from the EEG, the Saab and Gotman39 framework provides an objective way of comparing the performance of different features. To estimate seizure probability, Saab and Gotman39 determined the conditional probability of having a seizure given the values of three features: relative average amplitude (RAA), relative scale energy (RSE), and co-efficient of variation of amplitude (CVA). In this paper, their approach was implemented on epi- leptic scalp EEG data collected by our group. The necessary modifications required to achieve similar performance to that given in their paper are described. Furthermore, six alternative features for estimating seizure probability were considered: relative power (RP), bounded variation (BV), mean of the average cross-correlation function (MXC), relative derivative (RD), relative bounded variation (RBV), and relative mean of the average cross-correlation function (RXC). The detection performance of each of the nine features mentioned was evaluated separately by estimating the conditional probability of having a seizure given a single feature value. Different combinations of three features, selected from the nine features, were incor- porated into the Saab and Gotman39 framework. The performance of all 84 possible combinations was analyzed to find a detector superior to that provided by our implementation of the original Saab and Gotman39 method. METHODS This section describes the method of Saab and Gotman39 and how our implementation of the method deviated from theirs. The methods for comparing the three original Saab and Gotman39 measures with six alternative measures are also described. Data Selection This subsection describes how both Saab and Gotman39 and our group obtained data. Saab and Gotman39 recorded their data using the Stellate Har- monie system for EEG monitoring (Stellate, Montreal, Canada). Their data was sampled at 200 Hz after bandpass filtering between 0.5 and 70 Hz, and they applied a variant of the longitudinal bipolar montage involving either 24 or 32 recording channels. Our data was recorded using a Compumedics E-series EEG system (Compumedics, Melbourne, Australia) and sampled at 512 Hz after bandpass filtering between 0.15 and 105 Hz. EEG data were recorded concurrently with video of the patient. For the sake of comparison, the signal was resampled down to 200 Hz, using the MATLAB Version 7.1 (Mathworks, Natick, USA) ���resample��� function which applied a 66.6 Hz low pass anti-aliasing filter. A longitudinal bipolar montage was also used to reduce 21 recording channels down to 16. Saab and Gotman39 used scalp EEG data, with the training set lasting 652 h in duration, and containing 126 clinical seizures from 28 patients. The method was then validated on a non-overlapping testing set involving 360 h of data, including 69 clinical seizures from 16 patients. To determine the probability distri- butions required for their method (see section ������Esti- mation of Conditional Seizure Probability������), Saab and Gotman39 used the following data from each patient in the training set: at least two seizures present in two 4���6 h segments each containing at least one seizure, and two 4 h non-seizure segments, one during the waking state, the other during the sleeping state. Our data consisted of 21 patients, was 525 h in duration, and included 88 clinical seizures. An initial analysis was performed on a specific training and testing set, then the data was reshuffled to perform a more definitive 10-fold cross-validation. The training set for the initial analysis was 367 h in duration, and contained 58 clinical seizures from 14 patients. The non-overlapping testing set involved 158 h of data, including 30 clinical seizures from 7 patients. These initial training and testing sets were selected such that they both had a similar distribution of number of seizures per patient (see Table S1 of the Supple- mentary Material). For the remaining 9 of the 10 cross-validation sets, each set was generated by ran- domly selecting 14 of the 21 patients for each train- ing set and the rest were used for testing. Data from a given patient was never in both the training and testing set. All clinical seizures from all patients were included as seizure data, and the rest of the data was labeled as non-seizure. Clinical seizure onsets and offsets were defined by a trained clinical neurophysiologist based KUHLMANN et al. 2130

on staff and patient records of seizure times, and visual inspection of combined video and EEG data. Seizure onset times were determined by visual identification when there was a clear change in the background activity with epileptiform or lateralising abnormalities. Seizure offset times were marked when epileptiform or lateralising abnormalities end and traces return to previous background activity. Often EEG seizure tra- ces slowed and did not return to normal background for a period of time, in these cases offsets were labeled at the end of lateralising abnormalities. There was no pre-selection of patient seizures for analysis. The EEG recordings for each patient were long-term continuous. In general, information about the periods during which the patients were awake or asleep were not available. Therefore, all data from the training set was used to create the probability distri- butions required for the detection method. Recording durations for each patient ranged from 4.5 to 65.5 h. The shorter recordings correspond to awake periods, hence there is some training bias towards awake data. However, only 4 out of 14 patients in the initial training set had recordings less than 15.8 h in duration. Some clustering of seizures21 was present for 4 of the 21 patients investigated (see Table S1 of the Supple- mentary Material). However, clustering effects were not considered as part of this analysis. Bayesian Probability of Seizure Occurrence: Scalp Data Saab and Gotman39 provide a seizure detection method that determines the probability of having a seizure by applying Bayes��� Law2 to scalp EEG. Vali- dation on a testing set gave sensitivity of 0.78, false positive rate (FPR) of 0.86/h, and a median detection delay of 9.8 s. Missed seizures were mainly character- ized by subtle or focal activity, mixed frequencies, short duration, or some combination of these traits. False detections were mainly caused by short bursts of rhythmic activity, rapid eye blinking, and EMG arti- fact caused by chewing. The Saab and Gotman39 method begins with a 5-level wavelet transform, using a Daubechies-4 wavelet,4,5,38 computed separately on each non- overlapping 2 s epoch of data in each channel. Fre- quency bands of 50���100, 25���50, 12���25, 6���12, and 3���6 Hz were created (corresponding to decomposi- tion scales 1���5). Three characterizing measures for the EEG were derived from the wavelet coefficients: RAA, RSE, and CVA,24,39 described below. Relative Average Amplitude (RAA) The RAA is the ratio of the mean peak-to-peak amplitudes in the current 2 s epoch to the mean peak-to-peak amplitudes in the background. The background is defined as a 30 s block ending 60 s before the last epoch. The EEG waveforms are first wavelet transformed, then for each wavelet band sig- nal, the segment decomposition method of Gotman and Gloor15 breaks the waveform into segments, where a segment is defined as a single line connection between two local extrema in the waveform. These segments are then used to determine the peak-to-peak amplitudes. To compute the mean peak-to-peak ampli- tudes in the background, the mean peak-to-peak amplitudes computed over the last 45 windows (i.e., last 90 s) were buffered. The average of the first 15 elements of the buffer, corresponding to the data block of 30 s ending 60 s before the previous epoch, was then taken. Relative Scale Energy (RSE) RSE is defined as the ratio of the energy in the coe���cients in a given scale to the energy of the wavelet coe���cients in all scales. It serves as a measure of rhythmicity as a sustained elevated value in one scale indicates a somewhat constant frequency in the signal. Energy for the discrete wavelet transform band (or scale) i is given as24,39: e��i�� �� X Ni k��1 DikNi 2 Dt ��1�� where Ni is the number of wavelet coefficients present in band i, Dik are the coefficient values in band i, and Dt is the 2 s epoch length. The RSE is then given by24 er��i�� �� e��i�� PM j��1 e��j�� ��2�� where M is the number of wavelet bands, and e(i) is the energy of the ith band. Coe���cient of Variation of Amplitude (CVA) The CVA is defined as the square of the ratio of the standard deviation, r, to the mean, l, of the peak-to-peak amplitudes (i.e., CVA = r2/l2). The waveform segment decomposition method15 described in section ������Relative Average Amplitude (RAA)������, is used to compute the peak-to-peak amplitudes of a given window for each wavelet band. The CVA serves as a measure of the variability of the signal amplitude. A low value indicates little variation, which should coincide with seizures that are more ���periodic���, and hence less variable, than normal EEG. Probabilistic Seizure Detector 2131