Domain‐Specific Prediction of Clinical Progression in Parkinson's Disease Using the Mosaic Approach

ABSTRACT Purpose: Due to the highly individualized clinical manifestation of Parkinson's disease (PD), personalized patient care may require domain‐specific assessment of neurological disability. Evidence from magnetic resonance imaging (MRI) studies has proposed that heterogenous clinical manifestation corresponds to heterogeneous cortical disease burden, suggesting customized, high‐resolution assessment of cortical pathology as a candidate biomarker for domain‐specific assessment. Method: Herein, we investigate the potential of the recently proposed Mosaic Approach (MAP), a normative framework for quantifying individual cortical disease burden with respect to a population‐representative cohort, in predicting domain‐specific clinical progression. Using MRI and clinical data from 135 recently diagnosed PD patients from the Parkinson's Progression Markers Initiative, we first defined an extremity‐specific motor score. We then identified cortical regions corresponding to “extremity functions” and restricted MAP, respectively, and contrasted the explanatory power of the extremity‐specific MAP to unrestricted MAP. As control conditions, domain‐related but less specific general motor function and nondomain‐specific cognitive scores were considered. We also tested the predictive power of the restricted MAP in predicting disease progression over 1 and 3 years using support vector machines. The restricted, extremity‐specific MAP yielded higher explanatory power for extremity‐specific motor function at baseline as opposed to the unrestricted, whole‐brain MAP. On the contrary, for general motor function, the unrestricted, whole‐brain MAP yielded higher power. Finding: No associations were found for cognitive function. The extremity‐specific MAP predicted extremity‐specific motor progression over 1 and 3 years above chance level. The MAP framework allows for domain‐specific prediction of customized PD disease progression, which can inform machine learning, thereby contributing to personalized PD patient care.