Unlocking Insights from Postictal EEGs: Investigating Predictive Markers of Seizure Recurrence

Seizure recurrence, often presenting as clusters, is a major clinical concern linked to increased morbidity. The immediate postictal period is a critical yet understudied window where recurrence frequently arises. This study evaluates whether EEG features from postictal intervals can distinguish postictal-to-ictal (P-I) from postictal-to-interictal (P-Inter) transitions, enabling early recurrence prediction.

EEG data from the CHB-MIT database were analyzed, comprising 73 postictal episodes from seven patients (44 P-I, 29 P-Inter). Each episode was segmented into 10-second windows, yielding 876 segments. Fifty wavelet-based features were extracted from low-sample entropy channels and classified using Decision Tree (DT) and Long Short-Term Memory (LSTM) models. Performance was evaluated using nested cross-validation and, to test inter-patient generalization, per-patient stratified nested cross-validation.

In subject-independent nested CV, DT achieved accuracy 0.75 (95% CI ± 0.029), sensitivity 0.73 (±0.041), specificity 0.77 (±0.038), F1-score 0.70 (±0.032), AUC 0.75 (±0.028), and FPR 0.20 (±0.039). LSTM yielded accuracy 0.71 (±0.027), sensitivity 0.69 (±0.066), specificity 0.72 (±0.069), F1-score 0.65 (±0.027), AUC 0.73 (±0.035), and FPR 0.28 (±0.069). Under patient-stratified evaluation, accuracy decreased to 0.67 (±0.076) for DT and 0.65 (±0.093) for LSTM, reflecting inter-patient variability.

These proof-of-concept findings indicate that postictal EEG, particularly P-I transitions, may encode information relevant to seizure recurrence. While the observed performance remains moderate, these results provide preliminary evidence warranting further investigation rather than indicating immediate clinical applicability.