Development and validation of a video-based deep learning model for distinguishing epileptic seizures from...

This study aimed to develop and validate a video-based deep learning system for distinguishing epileptic seizures (ES) from non-epileptic events (NEE) in a pediatric cohort. Using a prospective validation cohort, we further assessed the diagnostic performance and clinical applicability of the artificial intelligence (AI) model, investigated potential factors contributing to its diagnostic errors, and benchmarked its clinical utility against clinicians grouped by different levels of expertise.

An enhanced multiscale vision transformer was trained on 438 retrospectively collected videos, with benchmark comparisons against MViTv2 and SlowFast architectures. Prospective validation was performed using 130 consecutive videos to assess the diagnostic performance of the AI system against tiered clinician groups (interns, attending physicians, and chief physicians). A generalized linear mixed model (GLMM) was employed to identify factors associated with AI misdiagnosis, with further comparative analysis of diagnostic performance between AI and human clinicians.

Our model demonstrated significantly higher accuracy (p = 0.001) and sensitivity (p = 0.004) compared to the MViTv2 model. Although all performance metrics were numerically higher than those of the SlowFast model, these differences did not reach statistical significance. GLMM analysis indicated that event type (motor vs. non-motor) was a significant factor influencing model misclassification (p = 0.020). The model achieved substantially higher diagnostic accuracy for motor events compared to non-motor events (p < 0.001).

The video-based AI classifier shows promise as an assistive tool for clinicians in differentiating ES from NEE based on video evidence in a pediatric cohort. Our AI model demonstrated notably effective diagnostic performance for motor events, while its accuracy was more limited for non-motor events.