Low-Latency Seizure Onset Detection from Wearable EEG
Wearable EEG classifiers now detect seizure onset within a few seconds of true onset while holding false-alarm rates low enough for at-home use, though performance drops sharply outside the cohort a model was trained on.
Compact convolutional–recurrent models detect onset with sensitivity in the low nineties at latencies under six seconds. ✳ For example, a depthwise CNN with a two-layer GRU head reaches 93.4% sensitivity at 4.7 s median latency on a single-centre cohort, while a temporal-convolution ensemble reports 91.6–94.2% across focal and generalised events. ✳ Eight of eleven studies that reported alarm burden stayed under 0.6 false alarms per hour, with one behavioural-artefact-heavy deployment reporting 2.3 and another a 21% reduction over its own prior baseline. ✳ Across the seven studies reporting on-device inference, all ran within the power envelope of a wrist-worn unit. ✳ A range of architectures — dilated temporal convolutions, bidirectional recurrent stacks, attention pooling, and small transformer encoders — support onset detection under everyday motion and variable electrode contact. ✳
We reviewed 11 papers drawn from an initial pool of 63, applying 6 screening criteria. Each paper was assessed on 5 dimensions selected for their bearing on the research question. More on methods