Predicting Parkinson's Disease Progression Rate Using Causal Measures of Functional MRI with Deep Learning Predictive Models

点击翻译按钮获取中文摘要

Project abstract Parkinson's disease (PD) is the second most common neurodegenerative disease. A critical gap in the treatment of PD patients is that there is no clinically adopted method to predict an individual's progression rate. A predictor would enable the enrichment of disease modifying drug trials with fast progressors likely to show changes in the short duration of a clinical trial and enable a more informed discussion with patients about their prognosis. This proposal develops a composite biomarker of progression rate using the connectivity information provided by resting-state functional Magnetic Resonance Imaging (rs-fMRI) and deep learning. Deep learning (DL) is well suited to form predictive models because it learns both an optimal hierarchy of features and how to combine them for accurate prediction. In rs-fMRI the blood-oxygen level dependent signal can be analyzed to infer connectivity throughout the brain. Traditionally, connectivity has been computed as the correlation between average regional activation time courses. However correlation based connectivity is prone to inferring spurious connections due to its inability to distinguish indirect from direct connectivity and inability to distinguish bidirectional from unidirectional connectivity. A causal connectivity approach can discern these differences and thereby provide a more faithful characterization of the true neurobiological connectivity. The existing literature suggests connectivity, particularly causal connectivity, from rs-fMRI can inform the estimation of PD progression, but the attempt to predict progression rate with causal connectivity in a DL model is unique to this project. This research develops several distinct approaches for building a progression rate predictor and apply them to three datasets including: the Parkinson's Progression Markers Initiative dataset, the NINDS Parkinson's Disease Biomarkers Program (PDBP) dataset, and the University of Texas Southwestern Medical Center's prospective imaging extension to the NINDS PBDP. In these studies, individual progression rates have been tracked over multiple years using multiple clinical measures. First, causal and correlative measures will be generated regionally and used with a DL model to create a baseline predictor of progression rate. Second, voxel- level causal measures will be generated as the increased granularity is expected to improve prediction accuracy. Third, since purely data-driven DL methods can be sensitive to dataset limitations, such as insufficient subjects and noise, these limitations will be addressed by developing a new structural connectivity regularization approach that constrains causal connectivity by the subject's own diffusion MRI. This regularization method will be general and likely applicable for building predictors for other neurological disorders such as stroke and Alzheimer's disease. This proposal will yield both DL models for predicting progression rate and a novel method to calculate constrained causal connectivity. All predictive models, composite neuroimaging biomarkers of progression rate and software will be publicly disseminated for ready incorporation by the scientific and clinical communities.