Epileptic Seizure Detection on Patients with Mental Retardation based on EEG Features: A Pilot Study Lei Wang∗ , Pierre J. M. Cluitmans∗a , Johan B. A. M. Arends∗a , Yan Wu∗ , Andrei V. Sazonov∗b ∗

a

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

Abstract—Mental retardation (MR) is one of the most common secondary disabilities in people with Epilepsy. However, to our knowledge there are no reliable seizure detection methods specified for MR-patients. In this paper we performed a pilot study on a group of six patients with mental retardation to assess what EEG features potentially work well on this group. A group of EEG features on the time, frequency and spatio-temporal domain were extracted, the modified wrapper approach was then employed as an improved feature subset selection method. Results show high variance on obtained features subset across this group, meanwhile there exist some common features which characterize the high-frequency components of epileptic EEG signals.

I. I NTRODUCTION Mental retardation (MR) is one of the most common secondary disabilities in people with epilepsy, and the prevalence of epilepsy increases with the severity of the intellectual disability. About 50% of those with profound learning disability and between 10% and 20% of those with mild disability have suffered from seizures at some time in life [1]. However, compared with other epilepsy patients with normal mental condition, the EEG-based research on MR patients is insufficient. One preliminary result of the seizure detection program on MR-patients shows that only one of every six seizures in average is recognized by specialized epilepsy nurses [2]. Epileptic seizure onset detection aims at differentiating between normal, preictal and ictal stages. Many automated detection methods based on EEG features have been introduced in literature [3], [4]. These seizure detection methods were designed, optimized and evaluated for non-MR patients. However, the EEG signals of non-MR patients are potentially different to the signals of MR-patients [1]. To our knowledge, this is the first paper describing the EEG-based epileptic seizure detection on MR-patients. In this paper, we performed a pilot study on a group of six patients with mental retardation. The individuals in this group have different levels of MR, seizure types and the demographics. Thus to assess what EEG features potentially work well on individuals, the classification of normal and seizure epochs was conducted on each individual separately. Because of the limited sample size in this study, to reduce the risk of ’overfitting’ and to obtain some features with desirable generalization power, we need to select a small feature subset with a good class-separation ability. Therefore the wrapper

978-1-4244-9270-1/15/$31.00 ©2015 IEEE

approach [5] with a modified stop rule was employed as a feature subset selection method. The linear discriminant analysis (LDA) served as the classifier in the wrapper approach, and its classification performance was evaluated by an improved criterion, the area under curve (AUC) of the precision and recall (P-R) curve [6]. II. MATERIALS AND METHODS A. EEG Database A scalp EEG signals database was provided by Epilepsy Center Kempenhaeghe. This database includes nine adults MR-patients. For each patient, 24 hours continuous EEG signals were acquired using 24 electrodes of Ag/AgCL in positions according to the 10-20 positioning system with a sampling rate of 100 Hz and the montage of common average reference (CAVG). Six subjects with good quality, i.e., with limited movement artifacts, of recording EEG signals were selected for this study. Annotations (sleep duration, seizure onset and offset time, seizure types) for epileptic EEG patterns were conducted by the visual inspection on the EEG signals by experienced electroencephalographers and EEG technician. The seizures types are classified according to [7]. For each of the six subjects, the recording EEG includes at least two seizures. B. Preprocessing on Raw EEG Given the artifacts account for a significant proportion in the raw EEG recordings both during a seizure and normal background activities [8], there exists a trade-off between using reliable information and discarding artifacts in the EEG preprocessing, because the artifacts segments also could contain useful information for seizure detection. However, we constrain this study to use only sufficiently clean EEG segments for further analysis because the EEG recordings of MR group tend to have higher proportion of artifacts. Thus we set a relatively ’strict’ rule for the preprocessing. Firstly, for each EEG channel, the raw EEG signals are digitally filtered by using a 10th order Butterworth bandpass filter with pass band between 0.5 and 45Hz. Secondly, the filtered EEG recordings then are split into non-overlap sliding windows of two seconds. In each EEG segment two features, amplitude range (ra = (max(x) − min(x))/2, where x is the amplitude sequence of an EEG segment) and frequency

578

centroid, fc (refer to section C), are extracted. If they meet the following two criteria simultaneously, this EEG segment is labeled as ’reliable’ and kept for further analysis. Otherwise it is treated as artifacts and discarded. The value of the thresholds are based on the EEG experts’ domain knowledge and statistical analysis on this particular EEG database. Criterion1 : 10µV ≤ ra ≤ 200µV . Criterion2 : 0.5Hz ≤ fc ≤ 20Hz

(1)

C. Spectral Analysis The standard Periodogram and Welch Periodogram are commonly-used nonparametric methods for spectrum estimation. Study [9] shows that Welch Periodogram has better performance than standard Periodogram on classification of different EEG segments. In this study, the power spectral density (PSD) spectrum of EEG is estimated by the Welch Periodogram using a Hamming window of length 100 with 50% overlap, and frequency resolution of 0.5Hz. It is implemented in each EEG segment of two seconds. According to the definition in [10], the power of δ band (0.54 Hz), θ band (4-7.5 Hz), α band (8-13 Hz), β band (14-29 Hz), γ band (30-45 Hz) are calculated in terms of integral on the corresponding frequency band on estimated power spectral density (PSD) spectrum. Ptotal = Pδ + Pθ + PRα + Pβ + Pγ , where the power of each frequency band Px = P SD(f )df , x

and the corresponding power ratio on the total power is computed. The Spectral Centroids are calculated using formula: fP max

fc =

(f SˆW (f ))

f =fmin fP max

.

(2)

SˆW (f )

f =fmin

where SˆW (f ) is the estimated PSD by Welch Periodogram [9], fmin is 0.5Hz, fmax is 45Hz. D. Nonlinear Features in the Time Domain 1) Approximate Entropy (ApEn) : ApEn was proposed by Pincus and it is a measure of the regularity of data [11]. An irregular time series results in a higher non-negative value for ApEn while a regular and predictable time series signal results in low ApEn value [12]. E.g., for typical tonic seizure, ApEn tends to fall from normal EEG to interictal and increase in ictal stage [3]. 2) Lempel-Ziv Complexity: Lempel-Ziv (L-Z) complexity is the metric of complexity proposed by Lempel and Ziv to evaluate the randomness of finite sequences [13]. L-Z complexity tends to have high value at the beginning and the end of a seizure, and decrease during mid-seizure [14]. 3) Hurst Exponent: Hurst exponent is proposed by Hurst [15]. It is a measure of self-similarity, predictability and the degree of long-range dependence in a time-series and it tends to have lower value in ictal seizure stage [3]. A study on continuous EEG/ECoG suggested that an increase on Hurst exponent can precede the onset of seizures by several minutes [16].

E. Features in Time-frequency Domain Wavelet transform is particularly effective for representing information in time-frequency domain of non-stationary signals especially for EEG signals analysis. Study [17] shows that the epileptiform discharges of the patients with absence seizure can be characterized by using the discrete wavelet transform (DWT). In this study, the six levels DWT with Daubechies order 4 wavelet is performed on the EEG segment in each sliding window. Then the standard deviation sdi , i = 1, 2, ..., 6 of the wavelet coefficient in each level of decomposition is calculated as a feature. F. Mean Phase Lock Index The mean phase coherence was used as a statistical measure for synchronization related to pathological activity of epilepsy patients [18]. The instantaneous phase of a signal can be calculated by using Hilbert transform [19]. The difference of instantaneous phases of two signals sa (t) and sb (t) are defined as ∆φ(t) = φa (t) − φb (t), then the phase lock index (PLI) between the two signals is defined as: D E P LI = ej∆φ(t) , (3) t

where h·it is the operator of averaging over time. In a period time of synchronized signals, if ∆φ(t) is a constant, P LI = 1. If the signals are unsynchronized, then ∆φ(t) follows a uniform distribution and P LI = 0. Note that instantaneous amplitude and phase have a clear physical meaning only if s(t) is a narrow-band signal [20]. Therefore, in this study, filtering is required in order to separate the broad-band raw EEG signals into sub-band with limited frequency band width as shown in table I. The PLI between arbitrary pair of EEG channels is defined as P LI (n) , which is a bivariate feature, where n is the No. of arbitrary trail. Then the mean PLI is defined: 2

CN 1 X P LI (n) , mP LI = 2 CN n=1

(4)

2 where N is the number of EEG channels, CN is the combination number extracting two from N.

G. Feature Extraction (Feature Vector) All the annotated seizures in this study are generalized seizures (all EEG channels are involved in seizure stage), as a result, to reduce the false alarms by the abnormal EEG patterns (artifacts and epileptiform discharges which do not lead to a clinical seizure) in focal EEG channels, we use the mean value of each feature on all EEG channels. Thus in each two seconds sliding window, a feature vector (Fv) including the 30 extracted features in table I is obtained. H. Feature Subset Selection (Wrapper Approach) The feature subset selection is to select k of m available features, with the goal of maximizing class separation. When m is large, the brutal-force search method although is optimal but is not computationally practical. Thus an improved-efficiency

579

TABLE I F EATURE V ECTOR

0.9 0.8

Feature Index and Name

0.7

*

[0.5-45Hz],

∗∗

11. 12. 13. 14. 15. 16. 17. 18. 19. 20.

β band power δ band power θ band power γ band power power in f*range sd1 of DWT** sd2 of DWT sd3 of DWT sd4 of DWT sd5 of DWT

21. 22. 23. 24. 25. 26. 27. 28. 29. 30.

sd6 of DWT ApEn L-Z complexity Hurst Exponent mPLI (0.5-4Hz) mPLI (4-8Hz) mPLI (8-12Hz) mPLI (12-16Hz) mPLI (16-30Hz) mPLI (30-45Hz)

AUC of P−R Curve

1. standard deviation 2. zero-cross times 3. amplitude range 4. α band ratio 5. β band ratio 6. δ band ratio 7. θ band ratio 8. γ band ratio 9. spectral centroid 10. α band power

0.6 0.5 0.4 0.3 0.2 0.1 0

1 2 3 4 5 6 7 8 9 101112131415161718192021222324252627282930

discrete wavelet transform.

step

Fig. 1. The forward stepwise selection in wrapper approach. P−R Curve on 5 fold, average AUC=0.7298 1 0.9 0.8

Precision (PPV)

search method, forward stepwise selection, is used to find a suboptimal feature subset alternatively. Forward stepwise selection starts with the intercept, zero in this study, and sequentially adds into the classifier a predictor or feature that most improves the performance. In this study, the classifier is Linear Discriminant Analysis (LDA), which determines a high-dimensional line in feature space on which to project all samples, such that the samples are maximally separated. Since the discriminative power of features are evaluated using classifiers thus this feature subset selection method is termed the wrapper approach [5]. The criterion for the classification performance is the area under curve (AUC) of precision and recall (P-R) curve. Its advantage is discussed in section III. To select a limited number of features which have the desirable classification performance, we define a modified stop rule for the forward stepwise selection method. It stops when crit ≥ Tc or ∆crit ≤ T∆ , where crit is the AUC of P-R curve in each step, and ∆crit is the criterion increment of each step. The threshold value Tc and T∆ depend on specific subjects, and the principle is to choose the least number of features which reach the desirable performance. To ensure at least two features are chosen, the stop rule is performed from the second step. It computes km − k(k−1) times of feature combination 2 to choose k features from m feature set. We use subject #1 to illustrate the process of the wrapper approach. All the m steps were computed as shown in the Fig.1. Each bubble shows the classification performance of a feature subset. In this demo, it stops at step three with Tc = 0.8, T∆ = 0.02, and the top bubble in step three corresponds to an obtained subset with three features. The classification performance of obtained feature subset is shown in Fig.2. It shows the average of the P-R curve over five-fold tests with the standard deviation bar. The AUC of the average P-R curve serves as the criterion value in the forward stepwise selection.

0.7 0.6 0.5 0.4 0.3 0.2 0.1 0

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

Recall (sensitivity)

Fig. 2. The average P-R curve on the 5-fold test sets.

sensitivity in receiver operating characteristic (ROC), is the true positive rate of all real positives. Compared with the commonly-used ROC curve, the P-R curve can show a more informative representation of performance assessment under the imbalanced class distribution [21]. A particular example is the seizure class less than 2% of whole recordings, while non-seizure class larger than 98% in this study. The difference of P-R curve is to use PPV insteads of false positive rate (FPR) or (1-specificity) in ROC. FPR is the false positive rate of all real negatives. This is simply because given an arbitrary value of sensitivity, the corresponding FPR tends to be very small due to the large number of the negative class. Consequently it shows an overly optimistic performance and less power on differentiating classification performances. Conversely PPV in the P-R curve can correctly capture the classifier’s performance in highly skewed data [6]. B. Classification Performance Validation

III. EVALUATION A. Precision Recall (P-R) Curve The P-R curve is the precision vs recall plot. Precision or positive predictive value (PPV) is the ratio of true positives (TP) to the summation of TP and false positives (FP). TP is the number of the annotated seizure EEG segments which are correctly classified as seizures. Recall, also termed as

The cross validation (CV) is used to evaluate the classification performance. Because of the limited seizure samples on each subject, five-fold CV is performed on each subject. For each subject, the feature samples of normal and seizure are split into five equal parts respectively, one part of normal and seizure samples is chosen as the test set, with the remaining four parts of both normal and seizure samples as training

580

set. Repeating the process five times on different test set, we compute the average classification performance which serves as the criterion in the wrapper approach. IV. RESULTS Taking into account the different EEG rhythm characteristics in wake and sleep EEG signals suggests different baselines in feature space, each 24 hours EEG signals are split into wake and sleep segments according to the nurses’ reports. Seven EEG segments (wake or sleep) containing seizure epochs are selected from six subjects. We define a unified desirable performance Tc = 0.8, and choose the proper value of T∆ to ensure the number of feature subset smaller than five. Then the feature subset and the corresponding AUC of the average P-R curve on five-fold tests are obtained (table II). The index of the feature in table II corresponds to table I. TABLE II R ESULT S TATISTICS Subject*

AUC

Feature subset**

Tc , T∆

Seizure type

#1[s] #2[w] #2[s] #3[s] #4[s] #5[w] #6[w]

0.73 0.61 0.51 0.59 0.76 0.25 0.10

[16, 25, 8] [5, 20, 4, 22] [5, 29, 16, 18] [14, 16, 22, 18] [16, 7, 29, 20, 5] [9, 5, 30, 27] [16, 17, 29]

0.8, 0.8, 0.8, 0.8, 0.8, 0.8, 0.8,

tonic tonic, others tonic, others tonic, myoclonic tonic-clonic myoclonic, absence myoclonic, absence

*

0.02 0.018 0.018 0.018 0.018 0.018 0.01

[w]:wake, [s]:sleep, ∗∗ the order of features in the subset corresponds to the step in forward selection. The subjects #1-4 belong to one subgroup.

The first feature in the feature subset is selected in the first step of wrapper approach. It means they work better than other single features, e.g. feature No.16 potentially be the best single feature for subjects #1, #4 and #6. Furthermore, the results in table II show that the classification performance (according to AUC) depends on seizure types. The tonic-seizure subjects (#1- #4) have significant better classification performance than the myoclonic and absence-like seizures type subjects (#5, #6). In the subgroup (#1- #4) of tonic-seizure subjects, the common features whose number of occurrence is larger than 2 across the subjects are shown in table III. TABLE III C OMMON F EATURE IN S UBGROUP Feature in subset DWT*

sd1 of β band ratio sd3, sd5, ApEn, mPLI (16-30Hz) *

No. of occurrence 4 3 2

the pseudo-frequency of 1st level of DWT is 35.7 Hz, 3rd level 17.9 Hz and 5th level 11.9 Hz.

The pseudo-frequency of DWT is the center frequency of the scaled wavelet. The feature with the largest number of occurrence, sd1 of DWT, suggests that tonic-type seizures EEG signals might contain a large component of muscle activities (or EMG) on scalp which is characterized by high frequency. The other features including sd3, sd5, mPLI (1630Hz) correspond to the visually observed generalized tonic components of high frequency between 10 and 20 Hz.

V. C ONCLUSION There are significant variance among the obtained feature subsets across the subjects. It suggests the differences in seizure patterns and the background EEG activities among individuals. On the other hand, there exist some common features in the tonic seizure group, and they characterize the high-frequency component of EEG. This confirms the clinical knowledge that generalized high frequency components (fast EEG) are contained in tonic seizures. Further research with a large database will be needed to find the most significant features from this analysis to physiological meaning process. R EFERENCES [1] S. Deb, “Epilepsy in people with mental retardation,” in Handbook of Intellectual and Developmental Disabilities. Springer US, 2007. [2] 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. [3] 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. [4] C. E. F. Mormann, R. Andrzejak and K. Lenhnertz, “Seizure prediction: The long and the winding road,” Brain, vol. 130, pp. 314–333, 2007. [5] S. Theodoridis and K. Koutroumbas, Pattern Recognition, Fourth Edition, 4th ed. Academic Press, 2008. [6] J. Davis and M. Goadrich, “The relationship between precision-recall and roc curves,” in Proceedings of the 23rd ICML. NY, USA: ACM, 2006, pp. 233–240. [7] 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. [8] P. Cluitmans and M. Van De Velde, “Outlier detection to identify artefacts in eeg signals,” vol. 4, 2000, pp. 2825–2826 vol.4. [9] P. Diez, E. Laciar, V. Mut, and E. Avila, “A comparative study of the performance of different spectral estimation methods for classification of mental tasks,” in EMBS 2008, Aug 2008, pp. 1155–1158. [10] S. Sanei and J. Chambers, EEG Signal Processing. Wiley, 2007. [11] S. M. Pincus, “Approximate entropy as a measure of system complexity,” Proc. Natl. Acad. Sci. U.S.A., vol. 88, no. 6, pp. 2297–2301, Mar 1991. [12] S. M. Pincus, A. L. Goldberger, and A. L. Goldberger, “Physiological time-series analysis: what does regularity quantify?” Am. J. Physiol., vol. 266, no. 4 Pt 2, pp. H1643–1656, Apr 1994. [13] A. Lempel and J. Ziv, “On the complexity of finite sequences,” vol. 22, no. 1, pp. 75–81, 1976. [14] N. Radhakrishnan and B. Gangadhar, “Estimating regularity in epileptic seizure time-series data,” Engineering in Medicine and Biology Magazine, IEEE, vol. 17, no. 3, pp. 89–94, May 1998. [15] H. E. Hurst, “Long-term capacity storage of reservoirs,” Trans Amer Soc Civil Engineers, vol. 116, pp. 77–99, 1951. [16] I. Osorio and M. G. Frei, “Hurst parameter estimation for epleptic seizure detection,” CIS, vol. 7, no. 2, pp. 167–176, 2007. [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] F. Mormann, K. Lehnertz, P. David, and C. E. Elger, “Mean phase coherence as a measure for phase synchronization and its application to the EEG of epilepsy patients,” Physica D: Nonlinear Phenomena, vol. 144, no. 3C4, pp. 358 – 369, 2000. [19] A. Sazonov, C. Ho, J. Bergmans, J. Arends, P. Griep, E. Verbitskiy, P. Cluitmans, and P. Boon, “An investigation of the phase locking index for measuring of interdependency of cortical source signals recorded in the EEG,” Biological Cybernetics, vol. 100, no. 2, pp. 129–146, 2009. [20] M. L. V. Quyen, J. Foucher, J.-P. Lachaux, E. Rodriguez, A. Lutz, J. Martinerie, and F. J. Varela, “Comparison of hilbert transform and wavelet methods for the analysis of neuronal synchrony,” Journal of Neuroscience Methods, vol. 111, no. 2, pp. 83 – 98, 2001. [21] H. He and E. Garcia, “Learning from imbalanced data,” Knowledge and Data Engineering, IEEE Transactions on, vol. 21, no. 9, pp. 1263–1284, Sept 2009.

581

Epileptic seizure detection on patients with mental retardation based on EEG features: A pilot study.

Mental retardation (MR) is one of the most common secondary disabilities in people with Epilepsy. However, to our knowledge there are no reliable seiz...
566B Sizes 0 Downloads 7 Views