Interpretable Machine Learning Models for Motor Fluctuations in Parkinson’s Disease with Cross-Cohort Analysis#

Authors#

Rebecca Ting Jiin Loo, Lukas Pavelka, Enrico Glaab

Abstract#

Background:#

Motor fluctuations are significant complications of long-term levodopa therapy in Parkinson’s disease, impairing motor functions and quality of life. Effective management strategies are needed to improve patient outcomes.

Objectives:#

This study uses machine learning on single and multi-cohort analyses to identify key risk factors for motor fluctuations in Parkinson’s disease, explore their interactions, and understand their contributions to this complication variability.

Methods:#

We applied interpretable machine learning techniques to clinical data from three independent longitudinal Parkinson’s disease cohorts: LuxPARK (n=395), PPMI (n=485), and ICEBERG (n=116). Cross-cohort prognostic models were developed to identify potential predictors of motor fluctuations. We used cross-validation to evaluate cohort-specific and shared predictive factors, model performance, stability, clinical utility, and calibration.

Results:#

Single- and multi-cohort analyses demonstrated the effectiveness of predictive models, identifying significant baseline clinical predictors for motor fluctuations. Positively correlated predictors included Hoehn & Yahr stage, MDS-UPDRS Parts I and II, bradykinesia, freezing, axial symptoms, and rigidity. Motor fluctuations risk was inversely associated with tremors, late onset of PD, and visuospatial abilities. Cross-cohort analyses showed higher stability than single-cohort analyses, mitigating cohort-specific bias and enhancing model robustness. The prognostic models exhibited substantial clinical utility and calibration, with reliable predictions closely aligning with observed outcomes.

Conclusions:#

This study presents interpretable machine learning models for motor fluctuations prognosis, demonstrating significant predictive capabilities in cross-cohort analyses despite differences between cohorts. Optimized predictive models can improve Parkinson’s disease management and inform clinical decision-making, ultimately improving patient outcomes.

Keywords: Motor fluctuations; Machine learning; Cross-cohort analysis; Longitudinal cohorts; Predictive modeling