EpiScreen: Early Epilepsy Detection from Electronic Health Records with Large Language Models

Dimensions overall score 7.0

GitHub Code Pulse

No public code linked for this paper yet.

• EpiScreen, a method using fine-tuned large language models on clinical notes, achieved an AUC of 0.875 on the MIMIC-IV dataset for epilepsy detection.result 95%
• EpiScreen achieved an AUC of 0.980 on a private cohort from the University of Minnesota for epilepsy detection.result 95%
• Fine-tuning large language models substantially improved performance relative to off-the-shelf LLMs, yielding average absolute AUC gains of 13.2% on MIMIC-IV and 11.6% on UMN.technical 90%
• EpiScreen demonstrated improved cross-institutional robustness, with a model trained on UMN data achieving an AUC of 0.867 on MIMIC-IV, representing only a 12.6% relative decrease compared to its in-domain performance.result 90%
• In clinician-AI collaboration settings, EpiScreen-assisted clinicians outperformed independent clinicians by up to 10.9%.result 85%
• EpiScreen's cross-institutional performance surpassed established comparative methods by over 14.2% and 24.2% on two external evaluations.result 85%
• Seizure semiology descriptions, such as prolonged, stereotyped, and lateralized motor activity, are among the highest-attribution features for epilepsy identification in EpiScreen's explainability analysis.technical 80%
• EpiScreen provides a low-cost, scalable approach for epilepsy detection that could facilitate earlier diagnosis, particularly in resource-limited regions.market 75%

Competitive landscape

Competitor map is still being generated for this paper. Enable generation or check back soon.

Startup potential card