Multivariate Prediction of Parkinson's Disease Clinical Progression (2016 PPMI Data Challenge Winner)
PPMI Data Challenge, 2016
Parkinson's disease (PD) is heterogeneous in both clinical representation and prognosis, as indicated by a large diversity of rates of progression in motor as well as non-motor symptoms. It could therefore be helpful to have well-characterized and distinct subtypes of Parkinson's disease slow and fast clinical progression and have early indicators of the clinical progression rate for better predicting individually the course of the disease and designing efficient management strategies.
Summary of Results:
Among baseline factors considered from the Parkinson's Progression Markers Initiative study, diffusion magnetic resonance imaging (MRI) -- a non-invasive neuroimaging technique to map the diffusion process of molecules, mainly water, in biological tissues -- and motor exam score from the Unified Parkinson Disease Rating Scale (UPDRS) -- a clinical rating tool to follow the longitudinal course of Parkinson's disease -- jointly were the best factors to distinguish slow-progressing versus fast-progressing PD patient early in the disease process.
Description of Methods Used:
An unbiased definition of "clinical PD progression" is crucial in determining the value of baseline factors in predicting such progression. We assessed each individual UPDRS item, rather than composite scores within different UPDRS domains, from patients both on and off medication at clinical visits to identify groups of patients with similar clinical progression characteristics. Brain patterns or networks highly associated with slow-progression versus fast-progression state were identified from baseline diffusion MRIs and fed to a data-driven approach together with other multidisciplinary baseline observations including patient demographics, genetic risk score and cerebrospinal fluid biomarkers to model individual predictions of slow and fast progression.
Impact of Results:
The prediction model could improve patient management in clinical practice as well as patient selection in clinical trials. Furthermore, machine learning of PD progression can lead to a biological meaningful pattern of brain damage. Brain imaging data, relative to clinical and cognitive measures, could potentially provide more objective measures with lower measurement errors, therefore, providing relative objective means to better predict disease progression.
Next Steps for Further Analysis/Development:
While we ignored a continuous spectrum of disease progression, modern machine learning techniques are in principle applicable to predicting a continuous outcome. Repeated and iterative analysis can lead to better stratification of patient-specific disease progression.
Assistant Professor of Radiology and Biomedical Imaging and Co-Director of the Center for Imaging of Neurodegenerative Diseases at University of California, San Francisco
Location: San Francisco, California, United States