AI helps neurologists distinguish epilepsy from functional seizures in retrospective study

Neurologists became more accurate at distinguishing epilepsy from functional seizures when an AI system supplied predictions and highlighted clues in medical notes. A study published October 8 in npj Digital Medicine evaluated EpiScreen across two institutions. Its 300-case clinician assessment showed gains from AI assistance, but the retrospective findings still need testing in real-world care.

Reading notes before confirmatory monitoring

Epilepsy and functional seizures can look similar, but they have different causes and require different care. Epileptic seizures arise from abnormal electrical activity in the brain. Functional seizures, also called psychogenic non-epileptic seizures, resemble them without that epileptic activity. Prolonged video-electroencephalography, which records brain waves alongside behavior, is the diagnostic gold standard. Its cost and limited availability can delay diagnosis.

EpiScreen takes a different input: the clinical narratives already stored in electronic health records. Researchers adapted large language models using notes labeled with clinicians’ final documented diagnoses. The retrospective dataset contained 13,633 records: 3,451 from Beth Israel Deaconess Medical Center’s MIMIC-IV database and 10,182 from the University of Minnesota. Notes were processed to approximate the information available before diagnosis.

The team fine-tuned Qwen, DeepSeek, Llama and Med42 models using a method that reduces the memory and training demands of adaptation. To make predictions easier to inspect, it used sentence-level Integrated Gradients. This attribution technique highlights sentences that most influenced a model’s decision, organized across 11 predefined clinical feature categories.

The stronger result was clinician plus AI

EpiScreen outperformed comparison models, including ClinicalBERT. Its best within-institution score reached 0.992 on the University of Minnesota records. When tested at an institution that had not supplied training data, the best scores were 0.875 on MIMIC-IV and 0.980 on Minnesota’s cohort. These are AUC scores: measures of how well a model separates the conditions across classification thresholds, not percentages of correct diagnoses.

The clinician exercise offered a separate comparison. Six neurology residents and two senior attending physicians worked in four pairs, each reaching decisions by consensus on 300 MIMIC-IV test cases. News-Medical’s account of the assessment gives standalone EpiScreen an accuracy of 81.0%, versus 81.7% for unaided attendings and 76.3%–79.3% for unaided resident groups. Those differences were not statistically significant.

With AI predictions and sentence attributions, resident-group accuracy rose to 87.7%–89.0%, and the attending pair reached 89.3%. Error patterns also changed: one resident group’s missed epilepsy cases fell from 36 to 14, while the attending pair’s false positives fell from 37 to 19. The gains therefore included fewer errors in both directions across those groups.

A screening role, not a replacement diagnosis

The authors propose using EpiScreen during or after neurology referral, before confirmatory monitoring, to support assessment and help prioritize urgent testing. Shorter delays, lower costs and less unnecessary exposure to antiseizure medication are proposed benefits, not outcomes measured by this study. The intended role is to support clinical judgment alongside definitive testing.

Important boundaries of the test

• Patients with both epilepsy and functional seizures were excluded, so the study did not evaluate that overlapping group.
• Not every diagnostic label was confirmed by video-electroencephalography. Analyses restricted to confirmed cases supported the findings, but documentation bias remains possible.
• Other conditions that can resemble seizures, including syncope and sleep disorders, were not evaluated.

The authors call for prospective, real-time trials using minimally curated notes to establish how EpiScreen performs in actual clinical workflows. That next step would move the evaluation from reviewing historical cases to testing assistance as care unfolds.