Abstract: Existing gait datasets primarily target identity recognition, lacking specialized gait data resources for medical diagnosis that cover multiple musculoskeletal diseases. To address this, a musculoskeletal disease gait dataset GaitMed oriented towards medical scenarios was independently designed and collected, covering 6 common disease types and including multi-view and multi-modal data, providing a data foundation for research on gait-based disease identification, classification, and rehabilitation assessment. The dual-branch gait recognition model MSGait, designed based on this dataset, achieves 81.99% Rank-1 accuracy and 96.27% Rank-3 accuracy in disease classification tasks by fusing silhouette and human parsing features, validating the application potential of gait recognition in medical diagnosis.