ECG-based Seizure Warning: A Developmental Comparison on Public Data

Abstract

ECG-based seizure warning requires advance notice, manageable false alarms and sufficient signal availability. We compared a causal waveform convolutional–recurrent network with a heart-rate and candidate-RR logistic baseline on identical SeizeIT2 validation endpoints and alarm rules. A custom partition assigned 50 ECG participants to development, 24 to validation and 50 to reserved testing; one additional participant lacked ECG. The primary protocol used 300-second inputs, 30-second decisions, a five-minute minimum lead, a 20-minute occurrence period and a 25-minute refractory period. Of 72 metadata-eligible validation seizures, 51 had a quality-passing warning opportunity. At threshold 0.5, CNN–GRU warned 22/72 seizures with 1,117 normal-period false alarms over 727.49 scored normal hours; logistic regression warned 21/72 with 1,430 false alarms. Mean patient window AUROC was 0.5273 versus 0.4392 across 16 participants with both classes; the descriptive paired interval included zero. Exploratory short-lead SeizeIT2 models warned 4–8/72 seizures with 4.164–6.825 false alarms per scored normal hour. A separate clean ECG-only Siena experiment reached internal checkpoint selection, but final evaluation and continuous replay were not completed. These developmental results do not establish reliable advance warning, deep-learning superiority or independent cross-dataset generalization. The contribution is a matched-support comparison with explicit alert burden, unavailable opportunities and evaluation provenance.