EpiScreen: Early Screening of Epilepsy from Medical Records Using Large Language Models

The University of Minnesota team developed the EpiScreen system, which enables early detection of epilepsy by fine-tuning large language models to analyze electronic medical records. The system achieved an AUC of 0.875 on the MIMIC-IV dataset, and the diagnostic accuracy of doctors increased by 10.9% in the human-AI collaboration mode, providing a low-cost and high-efficiency auxiliary solution for early epilepsy screening.

Epilepsy is a neurological disorder affecting approximately 50 million patients worldwide, but its diagnosis faces challenges—psychogenic non-epileptic seizures have similar symptoms but require completely different treatments, and misdiagnosis can lead to delays or side effects. The traditional gold standard, long-term video electroencephalogram (EEG) monitoring, is expensive, complex to operate, time-consuming, and has limited accessibility, especially in resource-constrained areas.

EpiScreen is based on fine-tuned large language models, with the core function of automatically identifying suspected epilepsy patients from electronic medical records. The model uses supervised learning with labeled medical records of epilepsy and psychogenic seizures, and employs an end-to-end approach to map medical record text to epilepsy risk scores without the need for manually defined feature rules.

The system performed well in dual dataset validation: AUC of 0.875 on the MIMIC-IV database and 0.980 on the University of Minnesota's private cohort. In the human-AI collaboration mode, the diagnostic accuracy of neurologists increased by 10.9% compared to independent diagnosis, combining AI processing capabilities with doctors' clinical judgment to improve diagnostic efficiency.

EpiScreen can serve as a low-cost screening tool to improve epilepsy detection rates (especially in resource-constrained areas); shorten diagnostic delays by providing risk assessment at the first visit; and prevent patients with psychogenic seizures from being misdiagnosed as epilepsy and receiving unnecessary medication, thereby reducing costs and side effects.

Limitations include reliance on the quality and completeness of medical record data, the current status as a research prototype requiring large-scale clinical validation, and inability to replace doctor's diagnosis. Future directions: expand to the identification of different epilepsy syndromes, integrate multiple data sources such as imaging, and develop a more user-friendly interface.