New AI model improves seizure detection across different patients by focusing on body movement patterns rather than visual background

Researchers developed a joint-centric attention model that extracts and analyzes body joint movements from clinical videos to detect seizures more accurately

The method uses Video Vision Transformer (ViViT) technology to tokenize joint-centered video clips while suppressing distracting background information

Cross-joint attention mechanism learns spatial and temporal interactions between body parts to capture characteristic seizure movement patterns

Cross-subject experiments demonstrate the approach outperforms existing CNN-, graph-, and transformer-based methods in generalizing to unseen patients

Automated detection from long-term clinical videos could significantly reduce manual review time and enable real-time seizure monitoring