Towards generalizable seizure monitoring: EpiVLM for cross-environment detection and classification.

Abstract

The translation of automated seizure detection from controlled clinical units to real-world settings is hindered by heterogeneous recording conditions and limited expert monitoring. We introduce EpiVLM, a multimodal vision-language system that combines clinically structured prompts with video reasoning for cross-environment seizure monitoring. Evaluated on a robust and diverse dataset of 232 video recordings from 127 patients, totaling 11,666 expert-annotated segments from two tertiary centers, unconstrained home recordings, and an independent public dataset, EpiVLM recognized five major semiologies with accuracy 0.795-0.947 and sensitivity 0.842-0.957. With prompts and decision thresholds fixed a priori, performance remained consistent across diverse real-world acquisition conditions without site-specific recalibration. In external validation sets, EpiVLM sustained strong recognition while maintaining low video-level false detections (0.47-2.45%) and timely detection (mean onset-to-detection delay <6 s). Compared with standard video deep-learning baselines, EpiVLM achieved superior overall performance. These results support scalable seizure recognition from routine video and motivate prospective evaluation for remote outcome monitoring.