Patient-level EEG seizure classification with AHCNet: Hybrid CNN and handcrafted feature fusion

Automated seizure classification from electroencephalography (EEG) necessitates rigorous validation, particularly concerning patient independence and potential dataset biases. This study evaluated hybrid feature fusion and confidence-based deferral for patient-level EEG classification. We assembled a class-balanced cohort of 198 patients from the Temple University Hospital seizure and epilepsy corpora, including individuals with focal non-specific seizures, generalized non-specific seizures, or no epilepsy. Our approach combined a multiscale one-dimensional convolutional neural network with handcrafted spectro-temporal and inter-channel spatial features. We assessed its performance using nested patient-level cross-validation and metadata-only baselines. Stacking and feature-fusion models achieved accuracies of 78.8% (bootstrap 95% confidence interval, 73.2–84.3%) and 78.3% (72.2–83.8%), respectively, showing no statistically significant difference. Notably, segment-level validation inflated accuracy by 20–23 percentage points. Metadata alone yielded 72.2% accuracy, and complete confounding between corpus origin and control-versus-seizure status precluded attributing this distinction solely to EEG content. Among 132 seizure patients from the same corpus, a two-class focal-versus-generalized classification achieved 81.1% accuracy, significantly outperforming metadata alone (at most 59.8%). Spatial features enhanced accuracy by 2.0–8.1 percentage points across models. While confidence weighting did not improve accuracy, deferring the 20% least-confident patients increased accuracy among the retained patients from 78.8% to 84.8%. These findings underscore the critical importance of patient-level validation and thorough confounding assessment when evaluating hybrid EEG classifiers. This framework is still investigational. For clinical application, it requires controls drawn from the same corpus as seizure cases and external validation.