Assessing Parkinson's Disease Severity Using Channel Attention and Graph Attention Networks on Kinect Skeleton Images

Parkinson's Disease (PD), a prevalent global neurological disorder, presents with motor symptoms like rigidity and tremors. Accurate severity assessment is vital for intervention evaluation. This study uses computer vision, specifically Kinect cameras and deep learning, to develop a system for assessing the severity of Parkinson's disease. Leveraging Kinect cameras' Precision in tracking key motion patterns, the system utilizes convolutional neural networks, namely Spatial-Temporal Graph Convolutional Network (STGCN) and Two-Stream Adaptive Graph Convolutional Network (2SAGCN), to detect movement anomalies and provide quantitative severity measures. A novel model is proposed, representing an enhanced version of 2SAGCN featuring a channel and graph attention module. Parkinson's severity classification comprises 11 classes based on 9 motion categories, totaling 99 classes. Three-dimensional transfer learning and data augmentation improve model Accuracy. This research advances non-invasive, quantitative Parkinson's assessment using innovative technology and advanced neural network models. The program is hosted on GitHub for implementation and result observation. The 10-fold cross-validation results of the best model, incorporating 2SAGCN structure, Graph, and Channel Attention, yield Accuracy, Precision, Recall, and F1-Score of Confidence Intervals 94.12± 0.14, 99.6 ± 0.20, 98.9 ± 0.10,99.2 ± 0.80 respectively.