Machine learning-based non-invasive Parkinson's disease diagnostic model using clinical blood biomarkers

Parkinson's Disease (PD) diagnosis lacks effective non-invasive markers, complicating early detection and timely intervention. Machine learning (ML) combined with clinical blood biomarkers may provide a feasible approach for early diagnosis and monitoring.

This study aims to construct and validate a non-invasive diagnostic model for PD using machine learning and routine clinical blood biomarkers, and identify key biomarkers linked to disease severity.

A total of 920 participants (428 PD and 492 non-PD) from two medical centers were included as training and validation sets. Biomarker selection was performed via least absolute shrinkage and selection operator (LASSO) and stepwise regression. Five machine learning models-logistic regression (LR), support vector machine (SVM), decision tree (DT), Naive Bayes (NB) and K-Nearest Neighbor (KNN)-were constructed and compared. The optimal model was interpreted using Shapley values (SHAP), and correlation with PD severity (Hoehn-Yahr stage) was assessed.

The SVM model demonstrated the best external validation performance (AUC = 0.916, recall = 0.949, F1-score = 0.843). SHAP analysis revealed superoxide dismutase (SOD) contributed the most to the model prediction, followed by gender and uric acid (UA). Furthermore, albumin (ALB) and SOD showed significant negative correlations with PD severity.

The SVM-based diagnostic model effectively differentiates PD from controls using readily obtainable clinical biomarkers, offering promising clinical utility for PD screening, diagnosis, and progression monitoring.