BCIwiki (bciwiki.com) — Seizure-prediction models lost much of their accuracy when tested on continuous intracranial EEG, with mean AUROC falling from 90% offline to 57.9% online, but an online evolutionary machine-learning framework recovered complete seizure prediction in 75% of patients, researchers at the Hospital del Mar Research Institute (IMIM) reported in a preprint posted to medRxiv on September 14, 2026. The work has not been peer reviewed.
The team released what it describes as the first publicly available continuous long-term stereo-EEG (sEEG) dataset for seizure prediction, covering 16 patients with drug-resistant epilepsy, 664.9 hours of recordings and 121 seizures.
The authors argue that segmented offline testing artificially rebalances seizure and non-seizure data and hides gradual neural drift, inflating performance. An offline combinatorial analysis compared referencing schemes, frequency bands, functional connectivity metrics and classifiers. The best elements, monopolar referencing, cross-correlation connectivity matrices and Random Forest classifiers, were then run under realistic class imbalance in a pseudo-prospective online framework, where 1,000 candidate models per patient evolved across continuous data streams.
Models identified by the framework predicted every seizure in 75% of patients and all but one in 87.5%. Restricting inputs to the epileptogenic network estimated by the group's earlier method cut implanted contact requirements by 79.12% and outperformed inputs from clinically resected areas. The researchers say the approach could underpin self-adapting closed-loop neuromodulation devices that retrain after seizures.