SHAP-Based Interpretable Machine Learning for Parkinson's Disease Severity Prediction: Integrated Analysis of Clinical and Environmental Features

Yuting Jin¹Xiang Li^2*Xinsheng Han¹Yang Qiu¹Hongyang Zhang¹Jianke Xu¹Miao Han¹

• ¹Kaifeng Central Hospital, Kaifeng, China
• ²Yellow River Conservancy Technical University, Kaifeng, China

Parkinson's Disease (PD) is the second most common neurodegenerative disease after Alzheimer's disease. With over 10 million patients worldwide and projected to reach 25.2 million by 2050, the acceleration of population aging has made PD severity assessment a core issue in clinical treatment and resource allocation. Traditional assessment methods such as the Unified Parkinson's Disease Rating Scale (UPDRS) and Hoehn-Yahr staging, though widely used, have limitations including strong subjectivity and difficulty in quantifying interactive factors. To achieve more precise and transparent prediction, this study constructed a multi-dimensional machine learning model integrating clinical and environmental features, introducing SHAP (SHapley Additive exPlanations) methods to enhance model interpretability. Based on data from 500 Parkin-son's disease patients, the study integrated 7 standardized clinical phenotypes (excluding UPDRS to prevent data leakage) and 8 environmental exposure fac-tors (such as temperature, PM2.5, UV intensity, etc.), compared the performance of 10 machine learning algorithms, and selected the optimal model through 5-fold cross-validation with comprehensive sampling strategy evaluation. Following rigorous target variable reconstruction using independent clinical dimensions, 1 XGBoost with SMOTE sampling achieved realistic discriminative performance (AUC = 0.781, precision = 0.548, recall = 0.750), demonstrating clinically meaningful capability appropriate for screening applications. SHAP interpretability analysis revealed non-motor symptoms as the primary predictor (SHAP value = 2.76), followed by serum dopamine concentration (2.39) and age (2.16), while environmental factors such as PM2.5 concentration, wind speed, and sunshine duration demonstrated modest but statistically significant contributions. Cross-validation results confirmed model stability (AUC: 0.781 ± 0.016), with balanced sensitivity (75.0%) and specificity (83.8%) suitable for clinical decision support. This proof-of-concept study developed an interpretable PD severity prediction framework with methodological safeguards against data leakage, achieving inte-gration of disease phenotypes, biochemical markers, and environmental factors. While demonstrating promising screening potential with realistic performance expectations (AUC = 0.781), the cross-sectional, single-center design limits gen-eralizability, requiring external validation and longitudinal studies before clinical deployment. This preliminary framework provides foundation for evidence-based severity screening approaches pending broader validation studies.