Spike-wave detection in ultra-long-term EEG recordings from patients with absence seizures

Seizure diaries are unreliable for tracking absence seizures, given their brief duration and subtle presentation. Ultra-long-term EEG monitoring provides a more objective alternative by capturing their characteristic spike-wave discharges. Here, we present a CNN-based approach for detecting spike-wave discharges in ultra-long-term EEG.

We obtained ultra-long-term two-channel EEG recordings from patients with pharmacoresistant epilepsy participating in our ongoing PREDYct study using a subcutaneous electrode. For each patient with absence seizures, approximately 96 h of continuous EEG data was annotated. We trained and evaluated patient-specific and cross-patient CNN-based spike-wave detectors using five-fold cross-validation within the study cohort. Patient-specific detectors were applied to the full dataset to estimate probability density functions of discharge durations. We applied the UNEEG EpiSight Analyzer to the annotated data and calculated its sensitivity and false positive rate.

We included five patients with a median of 2,175 h of EEG data per patient (range: 279-4,827). We annotated a median of 1,050 (695-3,332) spike-wave discharges per patient. Patient-specific detectors achieved a median AUC-ROC of 0.99 (0.98-1.00) and AUC-PR of 0.93 (0.81-0.95). Cross-patient models showed comparable performance. We detected a median of 19,249 spike-wave discharges per patient (2,763-45,442), with more short- than long-duration discharges. The UNEEG EpiSight Analyzer demonstrated high sensitivity for long discharges (>10 s), but sensitivity declined for shorter events.

Patient-specific CNN-based detectors demonstrated excellent performance, and cross-patient detectors performed comparably.

This pilot study supports the integration of automated spike-wave detection into long-term EEG monitoring to improve objective assessment of absence seizures.