EpiScreen: Language Models Support Epilepsy Screening from Clinical Notes
Published in Artificial Intelligence, Brain/Neurology, Journals, Tools.
October 8, 2026
A peer-reviewed study published today in npj Digital Medicine describes EpiScreen, an approach that fine-tunes large language models on labeled clinical notes to help distinguish epilepsy from psychogenic non-epileptic seizures.
The researchers reported an area under the receiver operating characteristic curve of up to 0.875 on MIMIC-IV and 0.980 on a University of Minnesota cohort. AUC measures discrimination across thresholds; it should not be read as the percentage of patients diagnosed correctly.
In the study’s clinician–AI collaboration assessment, assisted neurologists outperformed unaided experts by up to 14.3%, according to the authors. The approach could support screening using information already collected in electronic health records.
These findings concern performance in the evaluated datasets and assessment setting. They do not establish improved outcomes after routine deployment or replace specialist evaluation and prolonged video-EEG. The journal has released an accepted manuscript that remains subject to editorial changes.
Source: Zhou and colleagues, npj Digital Medicine, published October 8, 2026. DOI: 10.1038/s41746-026-03346-2.