Pediatric EEG Seizure Detection Using a Lightweight Temporal Convolutional Network: Recording-Case-Level Spectral Insights from the CHB-MIT Dataset

This is the authors' abstract. We don't add key points for this paper.

Accurate detection of epileptic seizures from scalp electroencephalography (EEG) remains difficult because clinical recordings are typically noisy, non-stationary, and highly variable across patients. We present a lightweight temporal convolutional network (TCN) for automated pediatric seizure detection using the CHB-MIT dataset, comprising 24 EEG recording cases from 23 children. The model learns directly from raw 18-channel EEG without heavy preprocessing and integrates dilated depthwise-separable residual blocks with squeeze-and-excitation modules to capture short- and long-range temporal dependencies within a compact architecture. Performance was rigorously evaluated using leave-one-recording-case-out cross-validation (LORCO-CV), case-grouped 5-fold cross-validation, seizure-event-grouped 5-fold sensitivity analysis, and continuous recording-level testing. Under LORCO-CV, the model achieved 82.3±16.5% accuracy, 80.1±25.6% sensitivity, and 0.866±0.194 AUC. The separate post hoc restricted-cohort analysis of 15 good-responder cases showed 91.5±8.6% accuracy and 0.986±0.022 AUC. Case-grouped 5-fold cross-validation achieved 98.2±0.8% accuracy, 98.3±1.6% sensitivity, and 0.999±0.001 AUC, while the seizure-event-grouped analysis retained high sensitivity (0.958±0.035; AUC=0.952±0.044). Continuous evaluation showed 0.910±0.103 event sensitivity, with 3.21±1.79 false alarms/hour and 15.9±4.6 s latency, highlighting performance with deployment-level considerations. Subject-session-level AUC was significantly associated with ΔTheta (ρ=0.5779, FDR-adjusted p=0.0083) and normalized ΔDelta–Theta index (ρ=0.5486, FDR-adjusted p=0.0083). Poorer performance reflected weak ictal–interictal separability and elevated baseline low-frequency activity. A frontal delta–theta pattern also emerged which should be interpreted cautiously as a possible signal-level observation rather than a definitive physiological marker, given the annotation and artifact-control limitations of CHB-MIT. These findings show that compact temporal learning, coupled with spectral interpretability, can reveal case-specific signal factors shaping pediatric EEG seizure-detection performance and deployment-related limitations more transparently.