Supervised Learning in Automatic Channel Selection for Epileptic Seizure Detection

TL;DR

This paper introduces a supervised learning method with automatic channel selection for epileptic seizure detection using iEEG data, achieving higher computational efficiency and comparable accuracy to current methods.

Contribution

The novel automatic channel selection engine improves computational efficiency in seizure detection without sacrificing accuracy.

Findings

49.4% increase in computational efficiency

Seizure detection sensitivity of 91.95%

Mean detection delay of 2.77 seconds

Abstract

Detecting seizure using brain neuroactivations recorded by intracranial electroencephalogram (iEEG) has been widely used for monitoring, diagnosing, and closed-loop therapy of epileptic patients, however, computational efficiency gains are needed if state-of-the-art methods are to be implemented in implanted devices. We present a novel method for automatic seizure detection based on iEEG data that outperforms current state-of-the-art seizure detection methods in terms of computational efficiency while maintaining the accuracy. The proposed algorithm incorporates an automatic channel selection (ACS) engine as a pre-processing stage to the seizure detection procedure. The ACS engine consists of supervised classifiers which aim to find iEEGchannelswhich contribute the most to a seizure. Seizure detection stage involves feature extraction and classification. Feature extraction is performed in both frequency and time domains where spectral power and correlation between channel pairs are calculated. Random Forest is used in classification of interictal, ictal and early ictal periods of iEEG signals. Seizure detection in this paper is retrospective and patient-specific. iEEG data is accessed via Kaggle, provided by International Epilepsy Electro-physiology Portal. The dataset includes a training set of 6.5 hours of interictal data and 41 minin ictal data and a test set of 9.14 hours. Compared to the state-of-the-art on the same dataset, we achieve 49.4% increase in computational efficiency and 400 mins better in average for detection delay. The proposed model is able to detect a seizure onset at 91.95% sensitivity and 94.05% specificity with a mean detection delay of 2.77 s. The area under the curve (AUC) is 96.44%, that is comparable to the current state-of-the-art with AUC of 96.29%.

Tables

Table 1. Table 1: Summary of existing EEG-based seizure detection methods

Reference	EEG type	No. of patients	No. of seizures	Data duration		Patient -specific	Split data for training	Testing sensitivity	FDR∗	Mean detection delay
				ictal	interictal
Saab and Gotman [2005]	scalp	$44$	$195$	$1012$ h†		No	$64\%$	$76\%$	$0.34$/h	$9.8$s
Kuhlmann et al. [2009]	scalp	$21$	$88$	$525$ h†		No	$70\%$	$81\%$	$0.60$/h	$16.9$s
Wang et al. [2016]	scalp	$10$	$44$	$72$ min	$121$ h	Yes	$80\%$	$91.44\%$	$99.34\%$	n/a
Zabihi et al. [2016]	scalp	$24$	$161$	$2.55$ h	$169$ h	Yes	$25\%$	$88.27\%$	$93.21\%$	n/a
Fatichah et al. [2014]	intracranial‡	n/a	n/a	$39.3$ min	$2.62$ h	n/a	$90\%$	$94.55\%$	$98.41\%$	n/a
Hills [2014]	intracranial	$12$	$48$	$41$ min	$6.5$ h	Yes	$50\%$	$91.33\%$	$94.02\%$	$3.17$s
Parvez and Paul [2015]	intracranial	$21$	$87$	$58$ h	$490$ h	n/a	$80\%$	$100\%$	$97\%$	n/a

Table 2. Table 2: Summary of the dataset

Subject	No. of electrodes	Ictal data length (s)	Interictal data length (s)	Unlabeled data length (s)	Train/Test ratio
Dog–1	$16$	$178$	$418$	$3181$	$0.19$
Dog–2	$16$	$172$	$1148$	$2997$	$0.44$
Dog–3	$16$	$480$	$4760$	$4450$	$1.18$
Dog–4	$16$	$257$	$2790$	$3013$	$1.01$
Patient–1	$68$	$70$	$104$	$2050$	$0.08$
Patient–2	$16$	$151$	$2990$	$3894$	$0.81$
Patient–3	$55$	$327$	$714$	$1281$	$0.81$
Patient–4	$72$	$20$	$190$	$543$	$0.39$
Patient–5	$64$	$135$	$2610$	$2986$	$0.92$
Patient–6	$30$	$225$	$2772$	$2997$	$1$
Patient–7	$36$	$282$	$3239$	$3601$	$0.98$
Patient–8	$16$	$180$	$1710$	$1922$	$0.98$
Total		$\mathrm{𝟐𝟒𝟕𝟕}$	$\mathrm{𝟐𝟑𝟒𝟒𝟓}$	$\mathrm{𝟑𝟐𝟗𝟏𝟓}$	$0.79$

Table 3. Table 3: Comparison between state-of-the-art and proposed method on computational efficiency. ACS engine with M 𝑀 M = 16 16 16 for all subjects.

			Hills [2014]		Proposed method
Subject	No. of electrodes	Data duration (min)	FA† (s)	Training (s)	ACS∗ (s)	FA† (s)	Training (s)	Processing time improvement
Dog–1	$16$	$9.9$	$2$	$23.7$	n/a	$1.7$	$23.7$	n/a
Dog–2	$16$	$22$	$3.8$	$62.7$	n/a	$3.5$	$62.5$	n/a
Dog–3	$16$	$87.3$	$15.4$	$341.8$	n/a	$13.8$	$348.8$	n/a
Dog–4	$16$	$50.8$	$8.8$	$144.9$	n/a	$8.5$	$149.2$	n/a
Patient–1	$68$	$2.9$	$2.7$	$14.1$	$6.5$	$0.7$	$13.1$	$17.9\%$
Patient–2	$16$	$52.4$	$21.3$	$105.8$	n/a	$22.3$	$106.3$	n/a
Patient–3	$55$	$17.4$	$25.5$	$80.7$	$48.4$	$6$	$57.6$	$40.1\%$
Patient–4	$72$	$3.5$	$7.1$	$13.3$	$10.6$	$1.5$	$11.6$	$35.8\%$
Patient–5	$64$	$45.8$	$79.1$	$519.3$	$180.3$	$29.4$	$162.4$	$67.9\%$
Patient–6	$30$	$50$	$38.5$	$146.2$	$93.5$	$15.2$	$98.1$	$38.7\%$
Patient–7	$36$	$58.7$	$53.7$	$435.3$	$171.9$	$19$	$244$	$46.2\%$
Patient–8	$16$	$31.5$	$13.3$	$59.7$	n/a	$13.4$	$60.1$	n/a
Average								$41.1\%$

Table 4. Table 4: A ​ U ​ C 𝐴 𝑈 𝐶 AUC Comparison between state-of-the-art and proposed method with M 𝑀 M = 16 16 16 for all subjects.

	Hills [2014]			Proposed method
Subject	$AU{C}_E$ (%)	$AU{C}_S$ (%)	$AUC$ (%)	$AU{C}_E$ (%)	$AU{C}_S$ (%)	$AUC$ (%)
Dog–1	$97.59$	$99.40$	$98.49$	$97.14$	$99.11$	$98.13$
Dog–2	$95.00$	$98.85$	$96.92$	$94.02$	$97.76$	$95.89$
Dog–3	$96.77$	$99.52$	$98.14$	$96.57$	$99.47$	$98.02$
Dog–4	$99.87$	$97.16$	$98.51$	$99.88$	$97.08$	$98.48$
Patient–1	$90.94$	$98.14$	$94.54$	$96.05$	$99.00$	$97.52$
Patient–2	$99.29$	$99.34$	$99.31$	$99.15$	$99.32$	$99.23$
Patient–3	$87.98$	$94.25$	$91.11$	$87.63$	$92.69$	$90.16$
Patient–4	$100$	$100$	$100$	$99.63$	$99.63$	$99.63$
Patient–5	$83.02$	$89.38$	$86.20$	$87.54$	$90.73$	$89.13$
Patient–6	$98.61$	$99.80$	$99.20$	$98.86$	$99.83$	$99.35$
Patient–7	$89.97$	$96.02$	$92.99$	$93.84$	$97.58$	$95.71$
Patient–8	$81.46$	$97.82$	$89.64$	$79.83$	$97.80$	$88.81$
Average	$93.37$	$97.47$	$95.42$	$94.18$	$97.50$	$95.84$

Table 5. Table 5: Comparison between state-of-the-art and proposed method with M 𝑀 M = 16 16 16 for all subjects on sensitivity (SEN), specificity (SEP) and onset detection delay with corresponding threshold (Thres.) for binary classification of seizure and non-seizure states.

	Hills [2014]				Proposed method
Subject	Delay (s)	SEN (%)	SPE (%)	Thres.	Delay (s)	SEN (%)	SPE (%)	Thres.
Dog–1	$1.92$	$95.83$	$97.13$	$0.34$	$1.92$	$94.46$	$98.09$	$0.42$
Dog–2	$2.25$	$92.86$	$89.29$	$0.11$	$2.25$	$92.44$	$91.82$	$0.13$
Dog–3	$1.83$	$95.84$	$97.33$	$0.14$	$2.17$	$94.80$	$98.28$	$0.21$
Dog–4	$1$	$89.55$	$89.90$	$0.09$	$1$	$89.55$	$90.26$	$0.09$
Patient–1	$4$	$92.31$	$96.16$	$0.36$	$3$	$94.87$	$96.16$	$0.28$
Patient–2	$2$	$95.03$	$99.20$	$0.29$	$2$	$95.03$	$99.20$	$0.28$
Patient–3	$1.75$	$85.42$	$76.61$	$0.17$	$1.38$	$86.67$	$77.45$	$0.13$
Patient–4	$1$	$100$	$98.95$	$0.25$	$1$	$100$	$97.37$	$0.23$
Patient–5	$10$	$71.67$	$87.78$	$0.05$	$3.75$	$77.78$	$83.49$	$0.03$
Patient–6	$1.75$	$98.24$	$98.85$	$0.19$	$2.25$	$97.35$	$99.21$	$0.26$
Patient–7	$6$	$86.51$	$99.14$	$0.14$	$2$	$99.59$	$99.79$	$0.20$
Patient–8	$4.5$	$92.78$	$97.90$	$0.22$	$4.5$	$92.78$	$97.90$	$0.20$
Average	$3.17$	$91.33$	$94.02$		$2.27$	$92.94$	$94.08$

Table 6. Table 6: Comparison between state-of-the-art and proposed method on computational efficiency. ACS engine with M 𝑀 M was optimized on the training data per subject.

				Hills [2014]		Proposed method
Subject	No. of electrodes	$M$	Data duration (min)	FA† (s)	Training (s)	ACS∗ (s)	FA† (s)	Training (s)	Processing time improvement
Dog–1	$16$	$9$	$9.9$	$2$	$23.7$	$6$	$1.5$	$19.8$	$17.1\%$
Dog–2	$16$	$10$	$22$	$3.8$	$62.7$	$15$	$2.8$	$49.9$	$20.8\%$
Dog–3	$16$	$8$	$87.3$	$15.4$	$341.8$	$77$	$9.7$	$224.5$	$34.4\%$
Dog–4	$16$	$13$	$50.8$	$8.8$	$144.9$	$36$	$7.1$	$122.2$	$15.9\%$
Patient–1	$68$	$16$	$2.9$	$2.7$	$14.1$	$5$	$0.7$	$11.7$	$26.2\%$
Patient–2	$16$	$11$	$52.4$	$21.3$	$105.8$	$39$	$11.2$	$82.8$	$26.0\%$
Patient–3	$55$	$8$	$17.4$	$25.5$	$80.7$	$31$	$4$	$45.1$	$53.8\%$
Patient–4	$72$	$4$	$3.5$	$7.1$	$13.3$	$8$	$0.8$	$11.1$	$41.7\%$
Patient–5	$64$	$16$	$45.8$	$79.1$	$519.3$	$115$	$28.2$	$158.1$	$68.9\%$
Patient–6	$30$	$8$	$50$	$38.5$	$146.2$	$56$	$9.6$	$60.6$	$62.0\%$
Patient–7	$36$	$13$	$58.7$	$53.7$	$435.3$	$118$	$15.6$	$198.6$	$56.2\%$
Patient–8	$16$	$8$	$31.5$	$13.3$	$59.7$	$23$	$5.4$	$41.6$	$35.6\%$
Average									$49.4\%$

Table 7. Table 7: A ​ U ​ C 𝐴 𝑈 𝐶 AUC Comparison between state-of-the-art and proposed method with M 𝑀 M was optimized on the training data per subject.

	Hills [2014]			Proposed method
Subject	$AU{C}_E$ (%)	$AU{C}_S$ (%)	$AUC$ (%)	$AU{C}_E$ (%)	$AU{C}_S$ (%)	$AUC$ (%)
Dog–1	$97.59$	$99.40$	$98.49$	$96.93$	$99.11$	$98.02$
Dog–2	$95.00$	$98.85$	$96.92$	$93.90$	$97.56$	$95.73$
Dog–3	$96.77$	$99.52$	$98.14$	$95.67$	$99.17$	$97.42$
Dog–4	$99.87$	$97.16$	$98.51$	$99.83$	$97.45$	$98.64$
Patient–1	$90.94$	$98.14$	$94.54$	$96.05$	$99.00$	$97.52$
Patient–2	$99.29$	$99.34$	$99.31$	$99.21$	$99.30$	$99.25$
Patient–3	$87.98$	$94.25$	$91.11$	$86.29$	$93.31$	$89.80$
Patient–4	$100$	$100$	$100$	$100$	$100$	$100$
Patient–5	$83.02$	$89.38$	$86.20$	$87.54$	$90.73$	$89.13$
Patient–6	$98.61$	$99.80$	$99.20$	$99.04$	$99.89$	$99.46$
Patient–7	$89.97$	$96.02$	$92.99$	$95.01$	$97.92$	$96.47$
Patient–8	$81.46$	$97.82$	$89.64$	$80.42$	$98.23$	$89.33$
Average	$93.37$	$97.47$	$95.42$	$94.16$	$97.64$	$95.90$

Table 8. Table 8: Comparison between state-of-the-art and proposed method with M 𝑀 M was optimized on the training data per subject on sensitivity (SEN), specificity (SEP) and onset detection delay with corresponding threshold (Thres.) for binary classification of seizure and non-seizure states.

	Hills [2014]				Proposed method
Subject	Delay (s)	SEN (%)	SPE (%)	Thres.	Delay (s)	SEN (%)	SPE (%)	Thres.
Dog–1	$1.92$	$95.83$	$97.13$	$0.34$	$2.5$	$93.99$	$96.41$	$0.34$
Dog–2	$2.25$	$92.86$	$89.29$	$0.11$	$2.25$	$92.44$	$89.20$	$0.11$
Dog–3	$1.83$	$95.84$	$97.33$	$0.14$	$2.5$	$94.17$	$97.25$	$0.16$
Dog–4	$1$	$89.55$	$89.90$	$0.09$	$1$	$92.16$	$90.61$	$0.09$
Patient–1	$4$	$92.31$	$96.16$	$0.36$	$3.5$	$93.59$	$97.12$	$0.38$
Patient–2	$2$	$95.03$	$99.20$	$0.29$	$2$	$95.03$	$99.20$	$0.26$
Patient–3	$1.75$	$85.42$	$76.61$	$0.17$	$1$	$90.45$	$76.19$	$0.12$
Patient–4	$1$	$100$	$98.95$	$0.25$	$1$	$100$	$100$	$0.49$
Patient–5	$10$	$71.67$	$87.78$	$0.05$	$4.75$	$73.89$	$85.94$	$0.04$
Patient–6	$1.75$	$98.24$	$98.85$	$0.19$	$2.75$	$96.48$	$99.39$	$0.33$
Patient–7	$6$	$86.51$	$99.14$	$0.14$	$5.5$	$87.35$	$99.14$	$0.15$
Patient–8	$4.5$	$92.78$	$97.90$	$0.22$	$4.5$	$93.89$	$98.19$	$0.26$
Average	$3.17$	$91.33$	$94.02$		$2.77$	$91.95$	$94.05$

Equations

$$AUC=\frac{1}{2}(AU{C}_{\mathrm{S}}+AU{C}_{\mathrm{E}}),$$

Full text

††journal: Expert Systems with Applications

1 Introduction

Epileptic seizure affects nearly $1\%$ of global population but only two thirds can be treated by medicine and approximately $7-8\%$ can be cured by surgery [Litt and Echauz, 2002]. Therefore, seizure onset detection and subsequent seizure suppression becomes important for the patients that cannot be cured by neither drug nor surgery. Early detection can allow early electrical stimulation to suppress the seizure [Echauz et al., 2007]. In this paper, we focus on how to effectively and reliably detect seizure onset based on iEEG patterns. Causes and treatment of seizure is beyond the scope of this paper.

EEG has been commonly used in brain-computer interface thanks to the convenient real-time readings and high temporal resolution of EEG signals [Zeng and Song, 2015, Zhang et al., 2013]. In recent years, EEG has provided a promising possibility to detect and even predict an epileptic seizure [Tieng et al., 2016, Fatichah et al., 2014, Parvez and Paul, 2015, Saab and Gotman, 2005, Osorio and Frei, 2009, Kuhlmann et al., 2009]. For seizure detection, Fatichah et al. [2014] used a combination of principle component analysis (PCA) and neural network with fuzzy membership function that can achieve accuracy rate up to $97.64\%$ . Tieng et al. [2016] combined wavelet de-noising with adapted Continuous Wavelet Transform in their algorithm and were able to achieve sensitivity of $96.72\%$ and specificity of $94.69\%$ with EEG data from mice. Another remarkable method is to transform EEG signals into images so as to leverage image processing techniques [Parvez and Paul, 2015]. This approach was able to obtain $98.91\%$ sensitivity and $94.35\%$ specificity. Zabihi et al. [2016] reconstructed EEG phase spaces using time-delay embedding method and PoinCare section. The phase spaces were then reduced by PCA before being fed to linear discriminant analysis (LDA) and Naive Bayesian classifiers. This approach achieved $88.27\%$ sensitivity and $93.21\%$ specificity in seizure detection.

Shoeb [2009] deployed $8$ filters spanning the frequency range of $0.5$–$24$ Hz for each $2$-s EEG epoch of all channels, then concatenated $3$ epochs to form a feature set to be fed to a SVM classifier. This approach was tested with the CHB-MIT EEG dataset and was able to detect $96\%$ of $163$ test seizures with a mean detection delay of $4.6$ seconds. Using the same CHB-MIT dataset, EEG signal was transformed into an image representation using $2$-D projection of the patient electrodes and the magnitude of $3$ different frequency bands spanning the range of [math]–$49$ Hz of each $1$ s block of EEG signal [Thodoroff et al., 2016]. The recurrent convolutional neural network took $30$ consecutive blocks as inputs to perform feature extraction and classification. The patient-specific detectors in this method have comparable performance compared to the proposed method by Shoeb [2009].