Patient-Independent Seizure Detection Based on Channel-Perturbation Convolutional Neural Network and Bidirectional Long Short-Term Memory

Segment-based average accuracies of 97.51% and 93.70% were achieved on the CHB-MIT and SH-SDU scalp EEG databases, respectively.

• A novel patient-independent method for seizure detection was developed to improve generalization across different patients.
• Multi-channel EEG recordings were processed using wavelet decomposition before feature extraction.
• A Convolutional Neural Network (CNN) was employed to extract features from the EEG data.
• Temporal variations in the extracted features were captured using a Bidirectional Long Short-Term Memory (BiLSTM) network.
• Postprocessing techniques, including smoothing and collar, were applied to reduce false detection rates and enhance sensitivity.
• The method showed average event-based sensitivities of 86.51% on the CHB-MIT database and 89.89% on the SH-SDU database.

AI simplified