Capsule Networks with a Self-Attention Mechanism for Classification of Parkinson's Disase from Tremor

The traditional diagnosis of Parkinson's disease (PD)primarily relies on the Unified Parkinson's Disease Rating Scale and the clinical experience of physicians, leading to significant subjectivity in diagnosis. To address this issue, we employ sensor systems to collect tremor parameters from PD patients and analyzed these collected data using traditional machine learning algorithms, providing an objective and quantitative diagnosis for PD. To obtain more comprehensive features, this study adopts the PKG wearable device to extract 12 tremor features from 100 PD patients, along with recording 9 scale-based features and 3 demographic characteristics. Furthermore, to overcome the issues of low accuracy and poor generalization ability of traditional machine learning methods, we propose a capsule network integrated with a self-attention mechanism (SA-CapsNet) for classifying and predicting early stage and moderate to advanced stage stage PD patients. We also utilize the Synthetic Minority Over-sampling Technique (SMOTE) to improve the model's ability to generalize. Lastly, The Shapley Additive exPlanations (SHAP) model is introduced to improve the interpretability of the predictive model, and feature analysis is conducted to identify the main factors influencing classification and prediction. Experimental results demonstrate that the SA-CapsNet method for PD can achieve an accuracy of 92.31%, precision of 82.19%, recall of 91.23%, specificity of 80.00%, and an F1 score of 85.25%. Therefore, the proposed SACapsNet method not only exhibits good generalization ability for small sample sizes and imbalanced class data but also effectively assists physicians in staging the severity of PD.