Title: Divisive hierarchical bayesian clustering of longitudinal data
Authors: Elliot Burghardt - University of Iowa (United States)
Joseph Cavanaugh - University of Iowa (United States)
Daniel Sewell - University of Iowa (United States) [presenting]
Abstract: Discovering hidden subgroups of a population with differing temporal trends is critical in clinical medicine, allowing researchers to design more powerful studies, personalize treatment options, and provide more information to patients about their condition. Clustering longitudinal data to find these subgroups is challenging due to temporal dependence, differing number of observations per subject, and the need to model multivariate trajectories. We propose a model-based clustering algorithm that takes advantage of the relationships between observations across variables and across time. We use a divisive hierarchical method which provides guidance on the plausible number of clusters in a principled way while still allowing for scientific insight to guide the selection process, and yields at each level of the hierarchy a valid estimate of the partition of the data. Our methods were inspired by the need to find subgroups in patients with Parkinson's Disease based on disease progression. We applied our method to the Parkinson's Progression Markers Initiative and discovered meaningful subgroups with varying rapidity of progression which is able to be predicted using baseline data.