Using vision transformers for electrographic seizure classification to aid physician review of intracranial electroencephalography recordings

We introduce a vision transformer (ViT)-based approach for automated electrographic seizure classification using time-frequency spectrogram representations of intracranial EEG (iEEG) recordings collected from patients implanted with the NeuroPace^® RNS^® System. The ViT model was trained and evaluated using 5-fold cross-validation on a large-scale dataset of 136,878 iEEG recordings from 113 patients with drug-resistant focal epilepsy, achieving an average test accuracy of 96.8%. Clinical validation was performed on an independent expert-labeled dataset of 3,010 iEEG recordings from 241 patients, where the model achieved 95.8% accuracy and 94.8% F1 score on recordings with unanimous expert agreement, outperforming both ResNet-50 and standard 2D CNN baselines. To evaluate generalizability, the model was tested on a separate out-of-distribution dataset of 136 recordings from 44 patients with idiopathic generalized epilepsy (IGE), achieving over 75% accuracy and F1 scores across all expert comparisons. Explainability analysis revealed focused attention on characteristic electrographic seizure patterns within iEEG time-frequency spectrograms during high-confidence seizure predictions, while more diffuse attention was observed in non-seizure classifications, providing insight into the underlying decision process. By enabling reliable electrographic seizure classification, this approach may assist physicians in the manual review of large volumes of iEEG recordings.