Unraveling Parkinson's Progression with AI: Multi-Model Predictive Analysis and Explainable Diagnostic Biomarker Insight

Abstract

Parkinson's Disease (PD) is a progressive neurodegenerative disorder whose clinical progression remains difficult to model due to its heterogeneous symptomatology and limitations in existing tracking methods. This study evaluates the effectiveness of supervised machine learning (ML) techniques for forecasting PD progression over a 3-year time window using high-dimensional, longitudinal biomarker data from the Parkinson's Progression Markers Initiative (PPMI) dataset. The nine predictive models are trained and compared, including linear regression, ridge regression, random forests, gradient boosting, and neural networks on curated input features and clinically validated progression scores. Ensemble models such as Random Forest and Gradient Boosting achieve the highest performance, yielding prediction accuracies between 94.8% and 96.4%. In contrast, deeper models exhibit no performance benefit, highlighting the suitability of simpler architectures for structured clinical data. To enable interpretability, SHAP and LIME are applied, which identify baseline clinical scores, age, REM Sleep Behavior Disorder metrics, and Body Mass Index as the most influential features across both global and patient-specific contexts. These results demonstrate that accurate PD progression prediction is feasible using standard ML methods trained on well-structured biomarker data. Moreover, the integration of model-agnostic explanation techniques facilitates biomarker-level insight, providing a basis for early intervention and individualized treatment planning.

Ähnliche Arbeiten