2025 Theses Doctoral

Statistical and Machine Learning Methods for Precision Medicine

Fei, Wenbo

Precision medicine aims to tailor medical care and treatment plans based on an individual's characteristics. This dissertation develops machine learning methods to extract meaningful features from digital marker signatures and address the challenges of learning individualized treatment rules using clinical trials and observational studies.

The first part of this dissertation proposes a joint nonparametric Bayesian approach that extends the hierarchical Dirichlet process autoregressive hidden Markov model with subject-specific transitions. This model allows for simultaneous learning of latent states across multiple subjects and repeated intensive measurements, facilitating symptom monitoring through wearable device technologies as an objective, low-cost, real-time alternative in movement disorders.

The second part introduces a novel approach to integrate the intermediate outcomes from multiple domains through a modified restrictive Boltzmann machine (RBM) model, such that clinical or biological measures can be combined into a personalized composite outcome. This model facilitates the use of interim measures in learning individualized treatment rules for early detection of non-responders and early intervention to improve final outcomes in mental disorders.

In the third part, we develop a novel framework for effective and generalizable learning of the individualized treatment effect (ITE) to address the multifaceted nature of treatment responses to mental disorders. This model jointly evaluates multi-domain treatment outcomes and can ensure generalizability across a potentially infinite class of diverse yet clinically relevant outcomes by leveraging a distributionally robust framework and the generalized latent factor models.