Explainable machine learning for the prediction of motor fluctuations and Levodopa-induced dyskinesias in Parkinson's disease

Background: Motor complications, such as motor fluctuations and Levodopa-induced dyskinesias (LID), significantly impair quality of life in persons with Parkinson's disease (PD) on long-term Levodopa treatment. Predicting their onset is crucial for tailored patient care. Objectives: To develop and evaluate machine learning (ML) models to forecast the onset of new motor fluctuations and LID in PD patients within three years from baseline assessment, and to assess how training cohort composition influences performance. Methods: A comprehensive ML workflow with repeated Nested Grid Search Cross-Validation was applied to real-world clinical data from a multicentric cohort of 247 PD patients. ML models were rigorously evaluated on the clinically relevant subgroup free of motor complications at baseline. SHAP analysis provided model explainability. Results: Models achieved moderate predictive power for both LID (SVC: MCC 0.28 {+/-} 0.14) and motor fluctuations (Voting MCC = 0.32 {+/-} 0.18). For LID prediction, the strongest predictors were the Levodopa Equivalent Daily Dose (LEDD), baseline motor fluctuations, and duration of Levodopa therapy, with risk increasing significantly above a LEDD threshold of 300-400 mg. A critical ablation study revealed that excluding patients with pre-existing complications caused a collapse in model sensitivity, highlighting their essential role in defining the upper bound of predicted risk. Conclusions: The model-based risk assessment is consistent with established clinical factors. Inclusion of the full spectrum of disease severity, including patients with pre-existing motor complications, in the training set is essential for achieving a robust probabilistic risk scale and reliable model calibration for new-onset prediction.