Researchers developed and validated EpiScreen, a tool designed to distinguish epilepsy from psychogenic non-epileptic seizures by drawing on clinical notes already captured in routine electronic health records. Findings from the research appeared in npj Digital Medicine, a peer-reviewed journal.
Drawing on health records from 13,633 patients spread across two institutions, the retrospective study evaluated fine-tuned large language models to determine whether those AI tools could reliably differentiate epilepsy from PNES. When tested both within a single institution and across multiple institutions, EpiScreen posted AUC scores of up to 0.992 and 0.980 respectively, placing it ahead of every comparison model evaluated. The AUC metric captures how well a model separates one condition from the other across the full range of classification thresholds.
In a retrospective review of clinical cases, EpiScreen allowed doctors to categorise patients more accurately, pointing toward a potential role for the technology in epilepsy screening. Epilepsy is described by doctors as a long-term neurological condition in which seizures recur without any identifiable external trigger. Triggered by abrupt, irregular electrical activity within the brain, the condition affects an estimated 50 million people worldwide, according to public health data.
Distinguishing epilepsy from psychogenic non-epileptic seizures — a condition also referred to as functional or dissociative seizures — presents a recognised challenge in the diagnostic process. In PNES, sudden episodes that resemble epileptic seizures occur without the underlying electrical disturbances in the brain that drive epilepsy. Because epilepsy and PNES operate through distinct biological pathways and call for different treatment approaches, prompt differentiation between the two conditions forms the foundation of effective patient care, according to the research.
Earlier research has placed the proportion of patients referred to tertiary centres on the assumption of drug-resistant epilepsy who ultimately receive a PNES diagnosis at somewhere between one in five and nearly one in three, a pattern that points to how often the condition goes unrecognised at earlier stages of care. Extended inpatient monitoring with video-electroencephalography tends to be the approach clinicians rely on most heavily for a definitive answer when differentiating epileptic seizures from psychogenic non-epileptic events. Neural activity and behavioural data are captured together in real time through this approach.
Admission costs for vEEG in the United States have been documented at above US$15,000, and scheduling delays of eight to twelve weeks have been recorded at various facilities. Prolonged waits for a diagnosis can leave patients undergoing procedures they did not require, while interventions that may address their condition arrive only after delays.
The research team retooled pre-existing large language models — rather than purpose-built tools — to separate epilepsy cases from those involving PNES, drawing solely on free-text clinical records. The EpiScreen model underwent both its training phase and its performance evaluation using patient records from two separate institutional repositories: 3,451 entries sourced from the MIMIC-IV database at Beth Israel Deaconess Medical Center and a further 10,182 drawn from the Clinical Data Repository maintained by the University of Minnesota. Diagnostic labels drew from clinicians' final recorded assessments, while the clinical notes underwent processing to capture only the information that would have been available prior to any confirmed diagnosis. Patients whose conditions involved both epilepsy and PNES together were not included in the study group.
Four open-source large language models — Qwen-2.5-32B, DeepSeek-R1-32B, Llama-3.3-70B, and Med42-70B — underwent instruction fine-tuning via Quantized Low-Rank Adaptation (QLoRA) to produce the EpiScreen tool. This approach to adapting large-scale models places notably lighter demands on both available memory and the computational resources required for training. Attribution scores generated by Integrated Gradients at the sentence level allowed researchers to identify which passages carried the most weight in the algorithm's outputs, spanning 11 predefined clinical feature categories.
To evaluate 300 test cases drawn from the MIMIC-IV dataset, four medical teams were assembled — a group comprising six neurology residents alongside two senior attending physicians — conducting the assessment through a retrospective clinician-AI framework. On the two benchmark datasets — MIMIC-IV and UMN — the Med42-based variant of EpiScreen posted area-under-the-curve scores of 0.891 and 0.976 respectively, placing it above the results recorded by conventional supervised models such as ClinicalBERT across both evaluations. Models built on Qwen and DeepSeek architectures each posted a score of 0.992 against the UMN benchmark. When AUC values were tallied across institutions, the leading models recorded 0.875 against the MIMIC-IV dataset and 0.980 against UMN, placing them above the comparison benchmarks on both measures.
In the clinician assessment, the tool operating independently posted a diagnostic accuracy rate of 81.0%, placing it just below senior attendings working without assistance at 81.7%, while neurology residents in the same unaided condition ranged from 76.3% to 79.3%. None of the observed differences cleared the threshold for statistical reliability. When clinicians received both AI-generated predictions and the sentence-level attributions behind them, diagnostic accuracy among residents climbed to a range of 87.7% to 89.0%, while attending physicians reached 89.3%. Tracking errors across both pairings, AI-assisted work in resident Group A brought missed diagnoses down from 36 to 14, while the attending pair reduced overcalled findings from 37 to 19.
Where vEEG confirmation was absent from a portion of the diagnoses, supplementary analyses restricted to verified cases continued to uphold the overall pattern of results. The study's reliance on recorded data leaves room for inaccuracy, and conditions that can resemble seizures — among them syncope and sleep disorders — fell outside the scope of evaluation.
The research team put forward placement of EpiScreen within the neurology referral pathway — whether concurrent with or following that process, and ahead of confirmatory video-EEG — as a way to bring assessment forward in time and give clinicians a clearer basis for prioritising monitoring resources. Prospective, real-time trials conducted on minimally curated clinical notes were identified as the most direct path toward understanding how this collaborative AI system performs in everyday healthcare environments.
