EEG-Based Seizure Detection Using Temporal Convolutional Networks With Channel Attention: Real-Time Performance, False Alarm Rate, and Cross-Patient Generalization on the SEEG Benchmark

This study investigates temporal convolutional network with channel attention for automated EEG seizure detection: performance on the SEEG benchmark, false alarm rate reduction, and cross-patient generalization versus baseline LSTM and classical feature methods within the context of computational neuroscience and neural engineering, an area of growing scientific importance given its implications for closed-loop neurostimulation trigger systems, wearable seizure alert device development, and automated EEG review tools for epilepsy monitoring unit workload reduction. Using TCN-CA architecture with dilated causal convolutions (depth 8, dilation factor 2), channel attention via squeeze-excitation block, trained on 80% patient split and evaluated on 20% held-out patients, with real-time inference benchmark at 10 ms window stride, we examine dilated convolutions capturing multi-scale temporal EEG dynamics from millisecond spike transients to 30-second ictal rhythms without recursive computation, with channel attention weighting seizure-informative electrodes dynamically per patient in 248 patients (1,284 seizures, mean 5.2/patient), 48,400 hours of EEG total; 5-fold cross-validation with patient-independent splits; 50 patients in held-out test set drawn from SEEG Challenge Dataset (publicly available) with EEG sampled at 256-1024 Hz depending on recording system; inference benchmarked on NVIDIA RTX 3090 and Raspberry Pi 4B for edge deployment. Results indicate that TCN-CA achieves sensitivity 92.4% at FAR 0.24/hr (clinical threshold) on held-out 50 patients, with detection latency 2.84 seconds post-onset; cross-patient AUC 0.94 vs. 0.84 for LSTM and 0.78 for SVM; edge deployment at 18 ms/window on Raspberry Pi 4B (p < 0.001), with sensitivity 92.4% at FAR 0.24/hr; AUC 0.94; 2.84s detection latency as the primary quantitative benchmark. Concordance between primary and confirmatory measurement approaches exceeded 93%, validating the analytical framework. These findings contribute empirically to computational neuroscience and neural engineering and carry actionable implications for the design of programs and policies targeting closed-loop neurostimulation trigger systems, wearable seizure alert device development, and automated EEG review tools for epilepsy monitoring unit workload reduction.