MyoFormer: Anatomy-guided transformer for muscle-level pathological gait assessment

Pathological gait is an early indicator of many disorders, and the ability to precisely analyze such gait plays a critical role in medical diagnosis and rehabilitation. However, existing wearable or vision-based approaches struggle to localize the disordered muscles because they are frequently constrained by limited data and inaccurate classification models. We present MyoFormer, a structure-aware, multi-representation transformer for muscle-level localization and classification of pathological gait. MyoFormer integrates both joint coordinates and bone vectors extracted from skeletal keypoints. Specifically, spatial and temporal attention mechanisms applied to coordinates are utilized to encode skeletal topology and model gait periodicity, whereas temporal attention applied to vectors serves to capture features across multiple scales. Therefore, MyoFormer captures both local kinematic cues and global gait cycle patterns, retaining sensitivity to subtle, dynamically evolving abnormalities. MyoFormer achieves 86.4% accuracy on synthetic data, and exploratory validation on a clinical cohort comprising 47 patients yields a 61.7% Top-1 accuracy, demonstrating its preliminary clinical feasibility as an auxiliary assessment tool.