Learning Phenotypic Associations for Parkinson’s Disease with Longitudinal Clinical Records

Abstract

Background. Parkinson's disease (PD) is associated with multiple clinical manifestations including motor and non-motor symptoms, and understanding of its etiologies has been informed by a growing number of genetic mutations, and various fluid-based and brain imaging biomarkers. However, the precise mechanisms by which these phenotypic features interact remain elusive. Therefore, we aimed to generate the phenotypic association graph of multiple heterogeneous features within PD to reveal pathological pathways of the complex disease. Methods. A data-driven approach was introduced to generate the phenotypic association graphs using data from the Parkinson's Progression Markers Initiative (PPMI) and Fox Investigation for New Discovery of Biomarkers (BioFIND) studies. We grouped features based on the structure of the learned graphs in both cohorts, and investigated their dynamic patterns in the longitudinal PPMI cohort. Findings. 424 patients with PD from the PPMI study and 126 patients with PD from the BioFIND study were available for analysis. For PPMI, the phenotypic association graphs were generated at different time points of the disease, including baseline (without any PD treatments), and 1-, 2-, 3-, 4-, and 5-year follow-up time points. Based on topological structure of the learned graph, clinical features were classified into homogeneous groups, that were densely intra-connected while sparsely inter-connected. Importantly, we observed both stable and longitudinally changing relations in the graphs generated, likely reflecting the dynamic pathologies of PD. By cross-cohort comparison, we observed very similar structure for graphs constructed from BioFIND (in which patients have a much longer duration of PD at enrollment than PPMI) and later-period (4- and 5-year follow-up) data from PPMI. This consistency demonstrates the effectiveness of our method. Interpretation. We analyzed the heterogeneous features of PD by generating the phenotypic association graphs. By analyzing the structural relationships among the features over time, our findings could improve the understanding of the pathologies of PD.