Boston University
Cluster analysis of multivariate longitudinal trajectories with random effects and applications
Abstract
dc:description.abstractElectronic Health Record (EHR) systems have been accumulating longitudinal medical data for over twenty-five years, creating the opportunity to generate longitudinal phenotypes. Longitudinal clustering of trajectories is one approach to accomplish this by identifying individuals with similar changes over time. While a handful of methods for longitudinal clustering have been compared, there is no clear preferred method, and there is a lack of testing on data that simulates realistic EHR data with irregularly spaced time points and non-polynomial curves. This limitation hinders our ability to effectively leverage valuable information within EHRs. Furthermore, many existing methods fail to leverage the potential of clustering multiple clinical variables simultaneously, and those that do, do not evaluate where the multivariate methods excel. These limitations were addressed in this dissertation by extending clustra, a k-means clustering method with thin-plate splines (TPS), in two ways. First, we evaluated the prediction of cluster centers using generalized additive mixed models (GAMM). We simulated data with within subject correlated data and assessed if there was an improvement in the percent of individuals correctly classified and in recovery. We showed that although there was an improvement in classification in 23% of simulations, the improvement was trivial at 1.4% and it was not related to the proportion of variance from random effects. Furthermore, the computation time for generalized additive models (GAM) was a fraction of that for GAMM. We concluded that using GAM in the TPS was sufficient for correctly clustering individuals. GAMM can be applied to the final clusters to estimate the standard errors. Second, we extended clustra to use more than one outcome (clustraMV). We simulated data with small to large datasets with a range of complexity and irregular data patterns to inform researchers of the types of data for which this method can be used with high confidence. Across almost all scenarios, using two outcomes outperformed using one. Finally, we applied clustraMV to eating disorder (ED) data from the Department of Veterans Affairs (VA) EHR. We showed how using multiple outcomes created clinically useful trajectories in multiple cohorts of veterans with eating disorder diagnoses, and how using derivations of variables (e.g. change in BMI) is a possible way to gain more clinically relevant trajectories. In summary, we highlighted the flexibility of clustra by extending it in multiple ways and we presented an example of how it can be applied to complex real data to create clinically useful longitudinal phenotypes. We concluded that clustra was effective at classifying individuals even when there were moderate levels of within-subject correlation. The extension to clustraMV also showed promising results for clustering two or more variables and we were able to uncover clinically useful trajectories when clustraMV was applied to ED data in the VA healthcare system.
Degree
thesis:*- Name thesis:degree_name
- Doctor of Philosophy
- Level thesis:degree_level
- doctoral
- Discipline thesis:degree_discipline
- Biostatistics
- Grantor
- Boston University
- Year dc:date.issued
- 2026
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Ehrbar, Rachel Quaden
- Advisor dc:contributor.advisor
-
- Gagnon, David R.
- ORCID dc:identifier.orcid
Subjects
dc:subject × 1Rights
- Language dc:language.iso
- en_US
Identifiers
dc:identifier.*- Handle dc:identifier.uri
- https://hdl.handle.net/2144/53758