Evaluating an MRI-Based Machine Learning Classifier for Parkinson’s Progression Using Real-World Clinical Measures
Discuss this preprint
Start a discussion What are Sciety discussions?Listed in
This article is not in any list yet, why not save it to one of your lists.Abstract
Background: Parkinson’s disease (PD) shows marked variability in disease progression, and predicting individual trajectories remains challenging. We previously developed a structural MRI–based machine learning classifier that distinguished faster from slower motor progressors using OFF-medication Movement Disorder Society–Unified Parkinson’s Disease Rating Scale Part III (MDS-UPDRS-III) scores with 89% accuracy. As OFF assessments are rarely performed in clinical practice, we evaluated whether this classifier predicts outcomes routinely used in care. Methods: Eighty-eight early PD patients from the Parkinson’s Progression Markers Initiative were previously classified as faster (n=42) or slower (n=46) motor progressors using a support vector machine model incorporating patient-specific multivariate gray matter volumetric distance and baseline clinical features. Primary outcomes were 48-month changes (Δ) in MDS-UPDRS Part II (experiences of daily living), Schwab & England Activities of Daily Living (S&E ADL), and levodopa equivalent daily dose (LEDD). Secondary analyses examined clinically meaningful thresholds: MDS-UPDRS-II worsening (≥2.51 points), ≥10% S&E ADL decline, ≥100 mg/day/year LEDD slope, and ≥1-stage Hoehn & Yahr (HY) progression. Results: Faster progressors showed significantly greater functional decline (ΔMDS-UPDRS-II: 5.31±4.77 vs. 2.76±4.56, p=0.01; ΔS&E ADL: 9.29±8.12% vs. 4.89±6.18%, p=0.009) and higher medication requirements (LEDD: 423.45±274.16 vs. 278.37±203.84 mg/day, p=0.006). Clinically meaningful deterioration was more frequent among faster progressors for MDS-UPDRS-II (81% vs. 52%, OR=3.90, p=0.009) and HY staging (62% vs. 28%, OR=4.13, p=0.003). Conclusions: An MRI-based classifier trained on OFF-medication motor assessments successfully predicts clinically meaningful deterioration across multiple real-world outcomes, supporting its potential utility for prognostic stratification in early PD.
