This study by Ding et al. focused on developing a deep learning artificial intelligence (AI) model that utilizes video data to distinguish epileptic seizures (ES) from non-epileptic events (NEE) in children. The dataset consisted of 438 retrospective EEG monitoring videos used for training and 130 prospective EEG monitoring videos for validation collected from Beijing Children’s Hospital. The model was based on a Multi-scale Vision Transformer (MViT) framework, which processes video inputs through frame embedding and multi-scale attention pooling to classify ES versus NEE. AI performance was compared to clinicians with different experience levels (interns, attending physicians, and chief physicians). Each group reviewed all videos independently. Statistical analysis included McNemar’s test and Generalized Linear Mixed Model (GLMM) to identify factors influencing AI misclassification (e.g., motor vs. non-motor events). The model outperformed MViTv2 and SlowFast in most metrics, achieving 85.95% accuracy, 91.8% sensitivity, 74.2% specificity, and an AUC of 0.93. The GLMM Error analysis found that event type (motor vs. non-motor) significantly affected accuracy (p=0.020). The model performed better for motor seizures, but struggled with subtle non-motor ones. Regarding human versus AI accuracy, interns achieved 60% accuracy, attending physicians 81.5%, and chief physicians 95.4%, while the AI model achieved 80.8% accuracy. The AI frequently mislabeled NEE as seizures for non-motor events, while interns did the opposite. The authors highlighted the promise of video-based AI as a diagnostic support tool in the discussion, particularly in distinguishing observable seizure behaviors without EEG data. They emphasized the model's proof-of-concept success but also emphasized the limitations of dataset imbalance, dependence on annotated video segments, and restricted generability to non-hospital settings.