MODWT-MRA-based activity-aware epilepsy discrimination using single-channel wearable EEG during activities of daily living

Automated EEG-based epilepsy discrimination has been studied extensively, but most approaches rely on multichannel recordings acquired under controlled conditions, limiting applicability to wearable monitoring during daily activities. This study presents an activity-aware discrimination framework using portable single-channel EEG acquired from the frontopolar (FP1) position during activities of daily living (ADL). The pipeline combines short-window segmentation, maximal overlap discrete wavelet transform multiresolution analysis (MODWT-MRA), and interpretable handcrafted amplitude, temporal-variation, nonlinear-energy, entropy, and spectral features. The framework is evaluated on 90 recordings (60 healthy, 30 epileptic) under four settings: combined, stationary, light ambulatory, and intense ambulatory. The selected configuration uses 5 s windows with 50% overlap and an ensemble classifier. Under leave-one-recording-out (LORO) validation with nested Bayesian hyperparameter selection, the primary evaluation, the ensemble achieves accuracies of 80.00%, 79.15%, 84.79%, and 85.71%, with balanced accuracies of 73.43–83.87% and sensitivities of 57.14–79.76%. Nested segment-level 10-fold cross-validation, a within-recording estimate, yields 89.37%, 88.42%, 94.47%, and 93.51%. Under LORO, MODWT-MRA features give the ensemble higher accuracy than raw-segment features in every setting and than decimated-wavelet features in the stationary and intense ambulatory settings, and the complete pipeline requires approximately 20–115 ms per 5 s segment on a laptop processor. These results establish recording-level generalization only and provide a proof of concept for epilepsy/healthy cohort discrimination using wearable single-channel EEG.