Prediction of Parkinson's Disease Using Machine Learning Methods.

Area of Science:

• Neuroscience
• Computational Biology
• Medical Informatics

Background:

• Early detection of Parkinson's disease (PD) is vital for effective management.
• Current methods for PD risk prediction lack consensus on necessary data and optimal models.

Purpose of the Study:

• To develop and assess machine learning models for predicting Parkinson's disease risk.
• To identify the most impactful and accessible factors for PD risk assessment.
• To compare the performance of commonly used machine learning algorithms in PD risk prediction.

Main Methods:

• Grouped PD-associated factors by cost and accessibility.
• Developed risk prediction models using eight machine learning algorithms.
• Incorporated data incrementally, from demographic variables to invasive biomarkers.
• Utilized the Shapley Additive Explanations (SHAP) method to determine factor contributions.

Main Results:

• Models incorporating demographic variables, hospital examinations, clinical assessment, and polygenic risk scores demonstrated superior prediction performance.
• Invasive biomarkers did not significantly improve prediction accuracy.
• Penalized logistic regression (AUC 0.94, Brier score 0.08) and XGBoost were the most accurate models.
• Olfactory function and polygenic risk scores emerged as the most significant predictors.

Conclusions:

• A practical framework for PD risk assessment using accessible data and machine learning was established.
• The study highlights the efficacy of non-invasive factors and specific machine learning models for PD risk prediction.
• Further research can refine these models for earlier and more accurate PD diagnosis.