Explainable machine learning reveals the challenges of predicting new-onset motor complications in Parkinson's disease

Predicting new-onset motor fluctuations and levodopa-induced dyskinesias (LID) is crucial for optimizing Parkinson's disease management. To establish a transparent prognostic framework, we applied explainable machine learning to real-world, multicentric clinical data from 247 patients to forecast the 3-year onset of these complications. Evaluated strictly on complication-free patients, the models achieved moderate predictive power (LID MCC=0.28; fluctuations MCC=0.32). SHAP-based interpretability confirmed predictions aligned accurately with established clinical knowledge, driven primarily by levodopa duration and Levodopa Equivalent Daily Dose, with risk increasing significantly above a 300-400 mg threshold. Crucially, an ablation study revealed that excluding patients with pre-existing complications from training caused model sensitivity to collapse, demonstrating that the full spectrum of disease severity is essential for robust risk stratification. Ultimately, this rigorous methodological stress-test demonstrates that baseline clinical features alone yield limited absolute sensitivity, highlighting the necessity of integrating dynamic, longitudinal data to achieve clinical-grade individualized prediction.