A Parkinson's disease pose estimation and assisted diagnosis model by integrating DeST-GCN and multi-scale graph convolutional networks

Parkinson’s disease (PD) is a progressive neurodegenerative disorder often associated with gait abnormalities, making gait analysis valuable for early diagnosis. However, current approaches are hindered by inaccurate pose estimation and limited model interpretability, reducing their clinical usefulness. To address these issues, we propose a two-stage framework that decouples 3D pose estimation from disease classification.In the first stage, we develop a Deformable Strided Transformer Graph Convolutional Network (DeST-GCN), which integrates deformable convolutions and graph-based modeling to extract accurate 3D skeletal sequences from gait videos. In the second stage, a Multi-Scale Deformable Graph Convolutional Network (MSDeGCN) leverages adaptive multi-scale kernels to learn hierarchical spatial-temporal features for PD classification.The framework was evaluated on both public and custom gait datasets. DeST-GCN achieves a 3D pose estimation accuracy of 0.9923 with a test loss of 0.0011. The MSDeGCN module delivers strong diagnostic performance with an F1-score of 0.905, accuracy of 0.942, precision of 0.934, and recall of 0.893. Gradient-based saliency and perturbation analyses highlight the spine, knees, and ankles as key regions influencing classification outcomes.This work introduces a novel diagnostic approach that combines deformable attention mechanisms and graph-based learning, enhancing both model accuracy and interpretability. By separating pose estimation and classification, the model improves robustness and offers clinically meaningful insights, demonstrating significant promise for early PD screening and decision support.