Low-latency six-channel Bi-LSTM for near-real-time epileptic seizure detection

This paper proposes a deep learning framework for near-real-time epileptic seizure detection using EEG data, designed to address two critical limitations of existing approaches: high detection latency and full-channel EEG dependence. The model employs a Bi-directional Long Short-Term Memory architecture to capture temporal dependencies in EEG signals, processing overlapping 0.5 s windows to enable low-latency sequence-to-sequence classification. Experiments are conducted on the publicly available CHB-MIT Scalp EEG dataset, evaluating different feature inputs including raw signals, channel means, and continuous wavelet transform representations. To address the well-known class imbalance problem in seizure detection, a dataset balancing strategy is applied to improve recall and minimize undetected seizures. A composite scoring metric is introduced to evaluate models jointly on accuracy, detection delay, and false negatives under clinically motivated weightings, providing a more meaningful basis for model selection than accuracy alone. Through systematic architectural search and correlation-based channel selection, the final model operates using only six EEG channels, reducing electrode requirements by 74% compared to the standard 23-channel setup. The system achieves an accuracy of 87.75% with a recall of 77.3% and an average detection delay of 1.33 s, representing a clinically meaningful improvement over prior studies reporting delays of 5–10 s on the same dataset. Performance is further reported in terms of specificity, F1-score, and false alarm rate per hour to provide a complete characterization under class imbalance. The reduced six-channel configuration is validated across all patients using Monte Carlo cross-validation and demonstrates strong suitability for wearable devices and edge computing platforms, where computational efficiency, patient comfort, and real-time responsiveness are critical design constraints.