Scientists have created an artificial-intelligence platform named “EpiScreen” that leverages electronic health records to assist neurologists in separating epilepsy from functional seizures. By scanning clinicians’ notes, the system uncovers distinguishing patterns prior to patients undergoing expensive video-electroencephalography (vEEG) studies. Development and testing employed a combined collection of 3,451 entries from the Beth Israel Deaconess Medical Center’s Medical Information Mart for Intensive Care IV (MIMIC-IV) database together with 10,182 entries from the University of Minnesota (UMN) Clinical Data Repository. Reported in the journal npj Digital Medicine, the work applied large language models to interpret unstructured information drawn from routine medical documentation for clinical decision support and research purposes globally.
Distinguishing Two Seizure Types
Epilepsy involves sudden, abnormal electrical discharges in the brain, affecting about 50 million people globally. Functional seizures, also known as psychogenic non-epileptic seizures, resemble epileptic events but lack the underlying electrical abnormalities. These events are psychological rather than neurological in origin. Because the symptoms overlap significantly, clinicians often struggle to tell them apart without specialized testing.
Studies show that 20% to 30% of patients referred for drug-resistant epilepsy actually have functional seizures. Accurate identification is critical because the treatments for each condition are opposite. The standard method for confirming a diagnosis is prolonged inpatient video-electroencephalography (vEEG), which records brain waves while a patient undergoes a seizure. This process is expensive, often exceeding US$15,000 per admission, and patients may wait eight to 12 weeks for an appointment.
Training the Model on Clinical Notes
The research team built EpiScreen by adapting several open-source large language models. EpiScreen was developed by instruction fine-tuning open-source LLMs, including Qwen-2.5-32B, DeepSeek-R1-32B, Llama-3.3-70B, and Med42-70B, using Quantized Low-Rank Adaptation (QLoRA). The training set included 13,633 patient records. To reduce the computational demands of fine-tuning, the researchers used a method called Quantized Low-Rank Adaptation (QLoRA). This technique allows the models to learn from clinical text while requiring less memory and processing power.
The study evaluated the system’s ability to distinguish epilepsy from functional seizures using two main metrics. The primary measure was the area under the receiver operating characteristic curve (AUC), which indicates how well the model separates the two groups. In tests within a single institution, EpiScreen achieved an AUC of 0.992. When tested on data from a different institution, the score dropped to 0.980. These scores far exceeded those of conventional supervised baselines, such as ClinicalBERT, which typically score around 0.875 to 0.891 on similar datasets. EpiScreen consistently outperformed conventional supervised baselines, achieving AUCs of 0.891 on MIMIC-IV and 0.976 on UMN for its Med42-based variant.
Human-AI Collaboration Improves Accuracy
To test how the system performs with actual clinicians, the researchers conducted a retrospective assessment using 300 test cases. Four medical teams, including six neurology residents and two senior attending physicians, reviewed the cases. Each team reached a diagnosis by consensus, with or without the assistance of EpiScreen. The AI system provided a diagnosis along with sentence-level explanations highlighting which parts of the clinical note influenced its decision. To explain the model, the study used sentence-level Integrated Gradients (IG), a mathematical technique that computes attribution scores to highlight the sentences most influential in the algorithm’s decisions across 11 predefined clinical feature categories.
Without assistance, the residents achieved diagnostic accuracy between 76.3% and 79.3%. Senior attendings scored 81.7%. The researchers note that these differences were not statistically significant, meaning random chance could explain the variation. When residents used the AI, their accuracy improved to 87.7% to 89.0%. The attending physicians saw an increase to 89.3%. The largest improvement, roughly 12.7 percentage points, occurred in the resident group. The AI also reduced errors by filtering out false negatives in epilepsy cases and false positives in functional seizures. Finally, error-pattern analyses showed that AI assistance reduced resident Group A’s false negatives (missed epilepsy cases) from 36 to 14, while the attending pair reduced false positives from 37 to 19.