Machine-based classification of epileptiform activity in selected EEG excerpts from genetic generalised epilepsy

To evaluate whether the deep learning model IGENet-TS, a time-domain convolutional neural network (CNN), can classify expert-selected EEG excerpts containing visible generalised epileptiform activity from genetic generalised epilepsy (GGE) versus normal healthy-control excerpts, and whether this selected-excerpt task generalises across centres.

We performed a retrospective selected-excerpt classification study of 455 routine 32-channel EEGs: 237 from Cuenca (107 GGE, 130 controls) for development and 218 from Bremen (106 GGE, 112 controls) for external testing. Each GGE recording contributed a 120-s resting-state segment containing at least one visible generalised discharge; controls contributed representative normal resting background. IGENet-TS analysed thirty non-overlapping 4-s windows per segment and averaged window probabilities to obtain a segment-derived label.

Repeated 70:30 internal testing yielded sensitivity 98.15% (95% CI 97.52-98.78), specificity against healthy-control excerpts 97.95% (97.40-98.50), accuracy 98.04% (97.45-98.63), F1 98.05% (97.50-98.60) and AUC 0.98 (0.96-1.00). Externally, sensitivity was 97.02% (96.39-97.65), specificity against healthy-control excerpts 96.80% (96.19-97.41), accuracy 96.91% (96.24-97.58), F1 96.62% (95.97-97.27) and AUC 0.97 (0.95-0.99).

The IGENet-TS model distinguished curated discharge-containing GGE excerpts from normal healthy-control excerpts with stable centre-separated performance. These results support technical feasibility for the selected-excerpt classification task, but do not validate unattended full-recording EEG interpretation or clinical workflow use.