Bayesian and Deep Learning Latent Class Methodology for Ecological Momentary Assessment and Imaging Data

Author

• Tianyi Wang, University of KentuckyFollow

Author ORCID Identifier

https://orcid.org/0009-0001-7673-4196

Date Available

6-26-2028

Year of Publication

2026

Document Type

Doctoral Dissertation

Degree Name

Doctor of Philosophy (PhD)

College

Arts and Sciences

Department/School/Program

Statistics

Faculty

Anna L. Smith

Faculty

Derek S. Young

Abstract

Latent class models are widely used across many research areas, including psychology, social sciences, ecology, and bioinformatics. By identifying latent subgroups within heterogeneous populations, these models provide meaningful class-level features that can support downstream analyses such as classification, regression modeling, and survival prediction. This dissertation investigates methodological approaches for identifying latent class structures from both continuous and categorical data and demonstrates their practical applications in multiple real-world settings. In Chapter 2, we conduct a comprehensive model comparison using ecological momentary assessment (EMA) data. Latent class analysis (LCA) and latent transition analysis (LTA) have been commonly used for categorical longitudinal data; however, their performance and practical feasibility may be limited when the sample size is large and individuals have unequal observation durations or irregular measurement times. To adresses these challenges, we compare two widely used approaches: LCA and LTA, with a proposed PCA-based K-means clustering method under various simulation settings. The best-performing approach is then applied to real EMA data to identify meaningful subgroup patterns in stress responses.

Chapter 3 focuses on a latent-structured Bayesian framework for longitudinal categorical data. Several extensions for categorical or ordinal outcomes have been proposed, but ordinal latent growth curve analysis models remain underexplored in this setting. Specifically, we evaluate the performance of Bayesian latent growth curve analysis (BLGCA) and Bayesian latent class analysis (BLCA). Through extensive simulation studies, we assess classification performance, robustness under missing data mechanisms, and true parameter coverage. BLGCA is then applied to real world EMA data.

In Chapter 4, we extend latent structure discovery to digital pathology by learning tumor-specific molecular feature representations from histopathology whole slide images. Although recent approaches have significantly advanced representation learning in computational pathology, effectively integrating biologically meaningful local morphological features with global contextual tissue information remains challenging, which may limit the ability of learned WSI features to support downstream analyses such as molecular prediction. To address this limitation, we propose LoGo-H2R, a deep learning histology-to-RNA framework that combines pathology-specific local tile representations with global slide-level information. By integrating tile-level morphological features with inter-tile relationships through a correlated multiple instance learning strategy, LoGo-H2R captures both fine-grained cellular and tissue morphology and broader spatial organization across the whole slide. The framework is evaluated using breast cancer data through gene expression regression analyses, with predictive performance assessed by Pearson correlation. Across these studies, this dissertation demonstrates how bridging statistical modeling and machine learning methods can uncover critical latent structures in complex behavioral and biomedical data.

Digital Object Identifier (DOI)

https://doi.org/10.13023/etd.2026.334

Archival?

Archival

Recommended Citation

Wang, Tianyi, "Bayesian and Deep Learning Latent Class Methodology for Ecological Momentary Assessment and Imaging Data" (2026). Theses and Dissertations--Statistics. 86.
https://uknowledge.uky.edu/statistics_etds/86