Seizure Detection Using Dynamic Warping for Patients With Intellectual Disability Lei Wang∗ , Johan B. A. M. Arends∗a , Xi Long∗b , Yan Wu∗ , Pierre J. M. Cluitmans∗a ∗

a

Department of Electrical Engineering, Eindhoven University of Technology Department of Clinical Neurophysiology, Epilepsy Center Kempenhaeghe, The Netherlands b Philips Research, Eindhoven, The Netherlands Email: [email protected]

Abstract—Electroencephalography (EEG) is paramount for both retrospective analysis and real-time monitoring of epileptic seizures. Studies have shown that EEG-based seizure detection is very difficult for a specific epileptic population with intellectual disability due to the cerebral development disorders. In this work, a seizure detection method based on dynamic warping (DW) is proposed for patients with intellectual disability. It uses an EEG template of an individual subject’s dominant seizure type, to extract the morphological features from EEG signals. A linear discriminant analysis (LDA) classifier is used to perform the seizure detection. Results show that the DW-based feature in the frequency domain is superior than that in the time domain, and the features extracted using wavelet transform method.

I. INTRODUCTION Electroencephalography (EEG) is paramount for both retrospective analysis and real-time monitoring of epileptic seizures. In the previous studies, however there is no EEG feature set which can achieve desirable performance in seizure detection for the whole patients population with variant seizure types [1]. In fact, since EEG signals vary significantly with not only specific seizure types but also individuals, the performance of seizure detection on a specific dataset suffers [2]. Studies have shown that EEG-based seizure detection is very difficult for a specific epileptic population with intellectual disability due to the cerebral development disorders [3], [4]. Thus, the prior knowledge of seizure type is essential for improving seizure detection performance due to the stereotyped character of the individual seizures. It means that the known seizure patterns can be used for the patient-specific seizure detection. The mimic EEG features [5], the wavelet transform [6] and even the raw EEG signals [7] have been used for seizure detection. However, those studies did not consider the seizure types which may influence the performance of seizure detection. According to the definition of international league against epilepsy (ILAE) [8], there are numerous different clinical seizure types. From the perspective of morphology of EEG signals, there are mainly three dominant EEG seizure types which usually exist in EEG segments of clinical seizures [9]. These EEG seizure types include: 1) spike pattern with fast spike discharges or afterdischarge, existing in most tonic seizures, 2) spike-wave complex pattern with hyper-synchronous discharges, existing in most absence seizures, 3) wave pattern with rhythmic discharges, existing in most temporal lobe

seizures, and rhythmic delta/theta seizures. These EEG seizure types potentially influence the seizure detection performance. It has been shown in our previous study [4]. In this work, on each individual patient, the seizure type which lasts longest time is chosen as the dominant seizure type. Then an EEG epoch with two seconds of the dominant seizure type is used as the template. We extract features based on dynamic warping (DW) method [10]. They capture the morphological information of the EEG recordings by comparing similarity between the selected template and each epoch. To ddate these features, a linear discriminant analysis (LDA) classifier is applied to evaluate the seizure detection performance, and the classification performance is compared with the features extracted by using discrete wavelet transform (DWT) proposed in [6]. II. SUBJECTS AND DATA A. EEG Data For each patient, the continuous scalp EEG signals (with sampling rate of 100 Hz) of one or two days were acquired using 24 electrodes (or EEG channels) of Ag/AgCL in positions according to the 10-20 positioning system in the Epilepsy Center Kempenhaeghe. The continuous EEG recordings containing at least two seizure segments, which show visually identified EEG changes, were selected in this study. In total, ten epileptic patients with intellectual disability were selected as shown in Table I. The procedure of data collection has been approved by Kempenhaeghe’s ethical review board. TABLE I S UBJECT D EMOGRAPHICS Prameter Mean ± Std Range Gender 5 males and 5 females Intellectual disability level 4 moderate and 6 severe Age [yrs] 28.9 ± 13.7 12 - 51 Seizure number 4.1 ± 1.7 2-8 Accumulated seizure duration [sec] 183.1 ± 214.3 38 - 707 Selected recording*length [hrs] 11.2 ± 2.8 7.6 - 15.0 *

wake or sleep period containing at least 2 seizures are selected.

B. Hierarchical Annotations The hierarchical annotations were conducted for all the subjects in this dataset. It includes wake/sleep time duration and seizure annotation. The wake and sleep time durations

were identified based on nurses’ video-based observation and EEG rhythm analysis by EEG technician. The seizure annotations include the clinical seizures types according to ILAE [8], the onset and offset time of seizures, and the EEG seizure types. The EEG seizure types include 1) type-A, fast spike pattern, 2) type-B, spike-wave complex pattern, 3) type-C, wave pattern, 4) type-D, the electromyography (EMG) artifacts existing in most tonic, tonic-clonic and myoclonic seizures, and other mixed seizure types and non-classifiable types. In this work, the EEG seizure type is the main concern, thus the seizure types will refer to the EEG seizure types in the rest of this paper. Note that more than one seizure types can exist in a single clinical seizure segment, e.g., the seizure types ‘A-C-B-A’ sequentially exist in a continuous epileptic EEG signal as shown in Fig.1. III. METHODS A. Preprocessing on Raw EEG EEG recordings contain a mixture of the electrical activity of the brain along with potentials arising from eyes, face muscles, movements and other non-physical sources, known as artifacts. The interesting portion of the EEG recordings is the superposition of the potentials generated from the electrochemical activity in the neocortex. The preprocessing therefore is employed to suppress these artifacts. However, the artifacts can be still useful for seizure detection, e.g., the EMG artifacts are often associated with the tonic, tonic-clonic seizures. Based on this assumption, in this work, we set a preprocessing rule that allows not only EEG signals but also certain amount of artifacts to be kept for the seizure detection. The raw signal on each electrode is used as an EEG channel instead of using the montages, common average reference (CAR) or bipolar [11], that could change the synchrony among EEG channels thereby corrupting the EEG quality [12]. Three EEG electrodes above the eyes, Fp1, Fpz, and Fp2 are excluded because where the signals are contaminated by the successive eye blink and movement artifacts, and visually distinct from that on other electrodes. On each EEG channel, the signal is filtered by using a 10th-order Butterworth bandpass filter with the lower and the higher cutoff frequency of 0.5 Hz and 45 Hz, respectively. This aims to remove the movement artifacts in the lower frequency band, and also to suppress the interference of 50 Hz power line that usually exists in EEG recordings. The filtered EEG signals are then split into nonoverlapping windows of two seconds. In each window, the amplitude range, ra is computed as ra = (max(x) − min(x))/2, where x is the amplitude sequence of an EEG segment. We set a range of 10µV ≤ ra ≤ 200µV. The lower threshold is to reject the artifacts caused by loose electrode-skin collection or sweating. The higher threshold is to reject the certain amount of artifacts caused by electrocardiogram (EKG or ECG), movements, and excessive EMG activities [13]. B. Feature Extraction In this work, the features are extracted by using DW distance, Euclidean (ED) distance and DWT.

1) DW: DW method allows nonlinear aligning of two signals, known as the warping path, thus it can be used to characterize the morphological information of EEG signals. One common DW variant is to impose global constraint on the admissible warping paths. These constraints not only speed up computations but also prevent pathological alignments by globally controlling the route of a warping path. Here we use the warping band named Sakoe-Chiba band, which states that the warping path cannot deviate more than r cells (horizontal and vertical) from the diagonal [14]. The DW distances in the time domain, known as dynamic time warping, DWt , between the EEG segment in a 2-second sliding window and the EEG template is computed [15]. The r is chosen to constrain that only the elements within 0.5 second shift can be aligned. The 0.5 second is the largest period of seizure spike/wave in our dataset. Similarly, the DW distances in the frequency domain, DWf , is computed based on the signals’ power spectral density (PSD) estimates [14]. Here we use the frequency resolution of 0.5 Hz. The aligning constraint is 2 Hz, which is empirically chosen to avoid pathological alignments. An example (as in Fig.1) shows that the proposed feature, DWf can separate EEG segments properly, and the separation is consistent with the visual classification of EEG experts. It is obtained by using the build-in function ‘linkage’ with the customized distance metric, DWf , in MATLAB. The frequency spectrums range used here is from 0.5 to 45 Hz. However, we find DWt on both the raw and normalized [10] EEG signals tends to get the meaningless classification results as suggested in [16].

Fig. 1. Classification on the subsequences of one channel seizure EEG signal. A continuous seizure EEG signal (top) is divided into 7 segments (down left) sequentially, then based on their frequency spectrums, a DW distance-based separation (down right) is obtained.

2) ED: Euclidean alignment can be considered as an extreme case of DW with warping path fixed at the diagonal line. The ED distance between the EEG segment in a 2-second sliding window and the EEG template, EDt and EDf in the time and the frequency domains, are computed as comparison. Note that EDt is computed on the normalized EEG amplitude sequences while EDf is computed on the raw PSD spectrums without normalization.

3) DWT: DWT has been widely-applied in seizure detection tasks [6]. We use it as a comparison in this study. The Daubechies order 4 wavelet is proposed for seizure detection in [17] for its similarity with the seizure spikes. Thus the six levels DWT with Daubechies order 4 is performed in this study. The relative energy (or ratio of coefficients), ei of each DWT level, with i = 1, 2, ..., 6, are computed as in [18]. It is termed as the feature set, DW Tratio . Furthermore, the energy fluctuation in a certain DWT level often experience significant change during seizures, thus the standard deviation, sdi with i = 1, 2, ..., 6, of the coefficients on each DWT level are calculated as another feature set, DW Tsd . These EEG features described above are computed on each EEG channel. Taking into account the seizures to be detected are all generalized seizures in this dataset, i.e., all EEG channels show similar seizure patterns in a seizure. Moreover, the generalized seizures account more than 97% of the total seizures and focal seizures only 3% in real life [3], thus the average value of these EEG features, DWf , DWt , EDf and EDt , together with DW Tratio and DW Tsd across all valid EEG channels are computed as inputs of the classifier. C. Classification LDA is a well-known method for dimensionality reduction and classification. For the two-class problem, the Bayesian solution provides a decision boundary consisting of a weight vector and a threshold [19]. Furthermore, to avoid the effect of different baselines in feature space between wake and sleep EEG signals, the classifications in wake and sleep EEG segments were performed separately. The notes ‘w’ and ‘s’ (as in Fig.2) will be used to show the background of the seizures in wake and sleep status respectively. D. Seizure Detection Performance Evaluation 1) Evaluation Criterion: Although the Bayesian solution can compute a determined threshold for LDA, it may not be necessarily the optimal one for specific data distribution. In this work, instead of using the default output of LDA, (i.e., one determined threshold), we evaluate the performance on the whole solution space by varying the value of threshold in the valid interval. We name the method as thresholding. The output of thresholding corresponds to a metric curve in the whole solution space, e.g., the receiver operating characteristic (ROC) curve (sensitivity vs (1- specificity) plot) or precision and recall (P-R) curve (precision vs recall plot). Precision is of all detected positives the fraction which is the true positive, It is also termed as positive predictive value (PPV), and (1PPV) is the well-known indicator, false positive rate. Recall is equal to the sensitivity in the ROC. Numerous studies have shown that P-R curve is more suitable for the imbalanced class distributions [20], and its area under curve (AUC) [4] is used as the criterion of classification performance [14]. 2) Cross Validation (CV): For each subject, the 5-fold CV is used to evaluate the classification performance. In the EEG feature space, the samples of both positive (seizure) and negative (non-seizure) classes are split into five equal parts

respectively. One part of them is chosen as a test set, with the remaining four parts as training sets. Then one-fold testing performance is obtained by classifying the test set using the trained classifier (on the training sets). Repeating the process five times, we compute the average classification performance on the 5-fold CV tests. IV. RESULTS AND DISCUSSION The feature extraction and classification were performed on the ten subjects with the dominant seizure types A, B, C and D. The classification performances measured by AUC of P-R curve are shown in Fig.2. The features DWf and EDf achieve generally the best classification performance across all subjects, while DWt and EDt have worse performance even though DWt is slightly better than EDt in most cases. The reasons might be explained as follows. In the classification of EEG segments based on its morphological patterns, the phase information became less important due to the significant time-varying character of EEG signals. While PSD only takes into account the power distribution on certain frequency components, as a result, the features based on that, DWf and EDf , are inherently suitable for recognizing EEG waveform (as shown in Fig.1). On the other hand, the features in the time domain, DWt and EDt on the raw EEG sequences tend to get meaningless classification results, since time-varying phase shifts in the normalized amplitude sequences affected them. We get similar results when we add the amplitude range, ra , into the classifier. Therefore, it suggests that sliding window-based DW in the time domain may not be appropriate for extracting morphological feature. In addition, the feature, DWf achieves the similar performance with EDf . It might be because the warping band we empirically chosen is not the optimal one. The classification performance of DWf may be improved by further validating the optimal warping constraint. The seizure types have potential influence on the classification performance. Fig.2 shows that the best detection performance is achieved on the type-D seizures while the worst on the type-B seizures. It is because the type-D seizures involve a lot EMG artifacts which show high amplitude and high frequency spikes thereby being easiest to detect. This seizure type in fact is very common in this population, e.g., tonic, tonic-clonic and myoclonic seizures in real life. In the opposite, type-B seizures, spike-wave complex patterns, are the most difficult to detect. It is because most patients with intellectual disability tend to have abnormal (‘slow’) background EEG signals which also show the similar spikewave complex patterns with that in seizures. To improve the seizure performance for this specific seizure type, other EEG features, even other physical signals, e.g., ECG, might need to be combined in the future research. Interestingly, the feature set we proposed, DW Tsd , is better than the classic feature set DW Tratio across all subjects. It suggests for seizures detection on this patients group, the fluctuation levels of power on certain frequency components are more robust properties than its distributions, since it

Average classification performance on CV tests 0.8 DWTratio DWTsd

AUCPR

0.6

DWf EDf

0.4

DWt EDt

0.2

0

#1s [A]

#2w [A]

#3w [B]

#4w [B]

#5s [AC]

#6s [C]

#7s [C]

#8s [D]

#9s [D]

#10w [D]

subjects

Fig. 2. Average classification performance on 10 subjects. The notes ‘w’ and ‘s’ stand for wake and sleep status respectively. The letters in brackets stand for the dominant seizure types, i.e., A: fast spike pattern, B: spike-wave complex pattern, C: wave pattern, D: the EMG artifacts.

characterizes the dynamic process of seizure evolving. Type-C seizures (rhythmic discharges) tend to cover a wide range of frequency bands and its rhythm is more flexible in real life. Thus the classification performance of the feature set, DW Tsd is significantly better than DW Tratio as shown on the subjects #5, #6 and #7. Overall, the proposed feature, DWf show more robust classification performance. It suggests that the use of prior information of seizure patterns can improve the seizure detection performance compared with the DWT feature sets without using priors. On the other hand, the variation of classification performance across these subjects is still significant. One reason is due to the effect of different seizure types. Another reason is the different proportion of the dominant seizures in the whole seizure segment. The subject #2 has larger proportion of the dominant seizure than that on the subject #1, accordingly it shows better classification performance. Such performance variation could be reduced by adaptively selecting a more general template set in the further study. From the clinical perspective, the results emphasize the important role of the individual EEG ’signature’ during seizures. This property could be further explored in patient-specific analytic schemes that should result in a better detection performance. V. CONCLUSION In this work, a seizure detection method based on DW is proposed for patients with intellectual disability. The DW approach is used to extract the EEG features in the time and the frequency domains. Its seizure detection performance is compared with the EEG features based on ED distance, and those extracted by using DWT method. Results show that, by using LDA, the DW-based feature in the frequency domain is superior than those in time domain, and also better than those extracted using wavelet transform method. Moreover, the standard deviation of wavelet coefficients, as a feature set, is superior than the commonly-used wavelet coefficients ratio in our dataset. The future work will focus on adaptively obtaining a more representative seizure template set on an extended dataset to improve the seizure detection performance. R EFERENCES [1] U. R. Acharya, S. V. Sree, G. Swapna, R. J. Martis, and J. S. Suri, “Automated EEG analysis of epilepsy: A review,” Knowledge-Based Systems, vol. 45, no. 0, pp. 147 – 165, 2013.

[2] C. E. F. Mormann, R. Andrzejak and K. Lenhnertz, “Seizure prediction: The long and the winding road,” Brain, vol. 130, pp. 314–333, 2007. [3] T. M. Nijsen, J. B. Arends, P. A. Griep, and P. J. Cluitmans, “The potential value of three-dimensional accelerometry for detection of motor seizures in severe epilepsy,” Epilepsy & Behavior, vol. 7, no. 1, pp. 74 – 84, 2005. [4] L. Wang, P. Cluitmans, J. Arends, Y. Wu, and A. Sazonov, “Epileptic seizure detection on patients with mental retardation based on eeg features: A pilot study,” in EMBC, Aug 2015, pp. 578–581. [5] H. Qu and J. Gotman, “A patient-specific algorithm for the detection of seizure onset in long-term eeg monitoring: possible use as a warning device,” Biomedical Engineering, IEEE Transactions on, vol. 44, no. 2, pp. 115–122, Feb 1997. [6] A. Shoeb, H. Edwards, J. Connolly, B. Bourgeois, S. T. Treves, and J. Guttag, “Patient-specific seizure onset detection,” Epilepsy & Behavior, vol. 5, no. 4, pp. 483 – 498, 2004. [7] D. F. Wulsin, J. R. Gupta, R. Mani, J. A. Blanco, and B. Litt, “Modeling electroencephalography waveforms with semi-supervised deep belief nets: fast classification and anomaly measurement,” J Neural Eng, vol. 8, no. 3, p. 036015, Jun 2011. [8] R. S. Fisher, W. van Emde Boas, and W. Blume, “Epileptic seizures and epilepsy: definitions proposed by the ILAE and the International Bureau for Epilepsy (IBE),” Epilepsia, vol. 46, no. 4, pp. 470–472, Apr 2005. [9] Automated Epileptic Seizure Detection Methods: A Review Study, Epilepsy - Histological, Electroencephalographic and Psychological Aspects. InTech, 2012. [10] E. Keogh and S. Kasetty, “On the need for time series data mining benchmarks: A survey and empirical demonstration,” Data Min. Knowl. Discov., vol. 7, no. 4, pp. 349–371, Oct. 2003. [11] B. J. Fisch, Epilepsy and Intensive Care Monitoring: Principles and Practice. Demos Medical Publishing, 2010. [12] S. Schiff, “Dangerous phase,” Neuroinformatics, vol. 3, no. 4, pp. 315– 317, 2005. [13] G. L. Krauss, The Johns Hopkins Atlas of Digital EEG: An Interactive Training Guide. Johns Hopkins University Press, 2011. [14] X. Long, P. Fonseca, J. Foussier, R. Haakma, and R. M. Aarts, “Sleep and wake classification with actigraphy and respiratory effort using dynamic warping,” IEEE J BHI, vol. 18, 2014. [15] A. Aarabi, K. Kazemi, R. Grebe, H. A. Moghaddam, and F. Wallois, “Detection of {EEG} transients in neonates and older children using a system based on dynamic time-warping template matching and spatial dipole clustering,” NeuroImage, vol. 48, no. 1, pp. 50 – 62, 2009. [16] E. Keogh, J. Lin, and W. Truppel, “Clustering of time series subsequences is meaningless: implications for previous and future research,” in Data Mining, 2003. ICDM 2003. Third IEEE International Conference on. IEEE, Nov. 2003, pp. 115–122. [17] H. Adeli, Z. Zhou, and N. Dadmehr, “Analysis of {EEG} records in an epileptic patient using wavelet transform,” Journal of Neuroscience Methods, vol. 123, no. 1, pp. 69 – 87, 2003. [18] Y. Khan and J. Gotman, “Wavelet based automatic seizure detection in intracerebral electroencephalogram,” Clinical Neurophysiology, vol. 114, no. 5, pp. 898 – 908, 2003. [19] T. J. Hastie, R. J. Tibshirani, and J. H. Friedman, The elements of statistical learning : data mining, inference, and prediction, ser. Springer series in statistics. New York: Springer, 2009. [20] H. He and E. Garcia, “Learning from imbalanced data,” Knowledge and Data Engineering, IEEE Trans., vol. 21, pp. 1263–1284, 2009.