Identification of Parkinson's Disease Subtypes with Divisive Hierarchical Bayesian Clustering for Longitudinal and Time-to-Event Data

TL;DR

This paper introduces a novel Bayesian clustering method that effectively identifies Parkinson's disease subtypes by analyzing complex longitudinal and survival data, overcoming traditional static data limitations.

Contribution

The paper presents a new divisive hierarchical Bayesian clustering approach tailored for multivariate longitudinal and time-to-event data in PD.

Findings

Successfully identified distinct PD subgroups with different progression patterns

Improved clustering accuracy over traditional methods

Demonstrated applicability on PPMI dataset

Abstract

In heterogeneous disorders like Parkinson's disease (PD), differentiating the affected population into subgroups plays a key role in future research. Discovering subgroups can lead to improved treatments through more powerful enrichment of clinical trials, elucidating pathogenic mechanisms, and identifying biomarkers of progression and prognosis. Cluster analysis is a commonly used method to identify subgroups; however, cluster analysis methods are typically restricted to static data or temporal data of a single variable. Progression of a complex disease process may be more appropriately represented by several longitudinal and/or time-to-event variables. Clustering with longitudinal and time-to-event data presents challenges, such as correlations between clustering variables, temporal dependencies, missing data, and censoring. To address these challenges, we present Divisive Hierarchical Bayesian Clustering methods featuring models for multivariate longitudinal trajectories and semi-parametric models for survival data to identify subgroups in idiopathic PD with differing progression patterns from the Parkinson's Progression Markers Initiative (PPMI) database.