Information-theoretical channel selection to enable efficient single-channel EEG-based epileptic seizure detection in ambulatory monitoring

Wearable single-channel EEG devices represent a promising solution for the continued detection of epileptic seizures outside clinical settings, but selecting the optimal electrode location for each patient remains a key challenge. In this work, we investigate the feasibility of using a single-channel EEG device for the continued detection of epileptic seizures, focusing on the selection of the optimal channel to detect ictal activity for each patient. The brute-force approach requires training a classifier from scratch for every channel and then selecting the best one. Instead, we propose a more efficient alternative, based on using information-theoretical measures (ITMs) to select the best channel from the input features and then training only one classifier. In particular, we show that both the Jensen-Shannon Divergence (JSD) and the Jensen-Tsallis Divergence (JTD), a generalization of the JSD, offer a computationally efficient alternative to the brute-force approach that enables accurate epileptic seizure detection using a single electrode. We validate our approach on the well-known CHB-MIT scalp EEG database, comparing the performance of the brute-force approach with the proposed ITMs and other statistical measures used in the literature. A patient-specific Support Vector Machine (SVM) is used to classify seizure events, with feature extraction performed both in the temporal and frequency domains. The experiments performed show that using the JSD/JTD for channel selection, instead of the brute-force approach, can lead to a reduced number of false alarms per hour (0.266 vs. 0.360 FA/h) and a slightly lower detection delay (7.373 vs. 7.486 seconds), with a substantial (10– \\(27.5\\times \\) ) increase in processing speed. The price to pay for the use of ITM-based channel selection is a moderate reduction in sensitivity with respect to the brute-force approach (0.933 vs. 0.960). These results highlight the benefits of ITM-based channel selection in substantially reducing computational requirements while attaining a good performance in terms of sensitivity, detection delay, and especially in the number of false alarms per hour.