Composite EEG biomarker modeling and energy–accuracy trade-off analysis for multi-patient seizure detection

Abstract

Abstract Reliable cross-patient seizure detection remains challenging because EEG characteristics vary considerably between patients. To address this problem, we proposed a compact composite feature termed the Seizure Intensity Index (SII) together with an extended representation incorporating additional theta and alpha band descriptors. The proposed feature representations were evaluated using Logistic Regression, Artificial Neural Network (ANN), and Spiking Neural Network (SNN) on the CHB-MIT and TUH EEG Epilepsy Corpus datasets. Spectral parameterization analysis showed that the proposed features primarily reflected localized oscillatory activity rather than global spectral shifts. The ANN achieved accuracies of 90.40% under random-split validation and 56.58% under CHB-MIT Leave-One-Patient-Out (LOPO) validation, indicating reduced cross-patient generalization. Logistic Regression achieved CHB-MIT LOPO accuracies of 72.83% using the baseline SII representation and 74.50% using the extended representation, while the SNN improved from 64.34% to 68.01%. During external validation on the TUH EEG Epilepsy Corpus, the SNN achieved the highest accuracy under the extended feature representation (90.78%). Although the extended representation improved performance, substantial cross-patient variability remained.