Modeling progression of Parkinson's disease

Abstract

Chronic diseases progress slowly over years and impose a significant burden on patients. To help alleviate this burden, we propose tackling three clinical questions: predicting progression events, summarizing patient state, and identifying prognosis-driven subtypes. These questions are challenging because progression is highly heterogenous across patients. In this thesis, we address these challenges for Parkinson's disease (PD), the second-most common neurodegenerative disorder, using various machine learning approaches. First, we process data from the Parkinson's Progression Markers Initiative to convert it into a format that is easier to use for downstream machine learning analyses. Utilizing this data, we design novel data-driven outcomes that capture impairment in motor, cognitive, autonomic, psychiatric, and sleep symptoms and allow for heterogeneity in the patient population. Then, we build survival analysis models to predict these outcomes from baseline. Using our motor and hybrid outcomes can reduce the sample sizes and enrollment time for early PD clinical trials. We can provide further reductions by identifying more severe patients for enrollment via survival analysis and binary classification methods. For summarizing patient state, we seek better representations of disease burden by learning trajectories of disease progression. Lastly, we consider ways to use these patient representations and outcomes for discovering subtypes that capture differing rates of progression. We hope this thesis starts to answer the three clinical questions for PD and sparks more machine learning research in this area.