2026 Theses Doctoral

Machine Learning Methods for Optimal Treatment Policy for Chronic Disorders

Yang, Bin

Chronic psychiatric disorders such as major depressive disorder (MDD) have complex disease mechanisms, yielding significant challenges for developing a universal treatment strategy. The current response rate of antidepressant treatment remains low and unpredictable due to substantial subject heterogeneity. To facilitate treatment selection for MDD, it is imperative to consider a variety of biosignatures that could match patients with optimal treatment options. Electroencephalogram (EEG) provides noninvasive measures of brain activity and is found to be valuable for the diagnosis of some chronic disorders. Specifically, pre-treatment EEG signals in the alpha and theta frequency bands have demonstrated some association with antidepressant response.

However, obtaining high-quality EEG is expensive, prone to artifact contamination, and subject to substantial subject-specific variation in data quality. In addition, even with proper EEG preprocessing, it is methodologically challenging to estimate heterogeneous treatment effects (HTE) due to the high-dimensionality of EEG biomarkers. Depression also manifests through behavioral dimensions, with anhedonia—reduced ability to experience pleasure—being a core symptom. The probabilistic reward task (PRT) provides a behavioral measure of reward learning that captures individual differences in how patients respond to reinforcement schedules. Modeling and integrating multi-modal data including EEG and PRT remain methodologically challenging.In this work, we develop three projects to address these challenges.

In the first project, we develop an automatic site-specific EEG preprocessing pipeline that accommodates equipment differences across multi-site clinical trials. We extract relative theta and alpha band power using the multitaper method, employ causal forests to estimate the conditional average treatment effect, and learn an optimal depth-2 treatment assignment decision tree outperforming Q-learning and outcome-weighted learning in simulation studies and application to the EMBARC trial.

In the second project, we propose MEHE (Multihead EEG-variational autoencoder for Heterogeneous treatment Effect), a unified deep learning framework that jointly optimizes EEG representations and treatment effect estimation through a generative, outcome-guided latent space. Through simulation studies and application to the EMBARC trial, we demonstrate superior performance in predicting optimal individualized treatment rules compared to existing approaches.

In the third project, we introduce personalized state-partitioned gated recurrent units (GRUs), a minimal recurrent architecture for modeling individual differences in PRT behavior. The proposed model bridges the interpretability of asymmetric Q-learning and the flexibility of recurrent neural networks, outperforming asymmetric Q-learning on choice prediction and the reinforcement learning drift-diffusion model on response-time prediction across four independent PRT cohorts. The learned hidden states recover Q-value-like signals without being given the update rule, and subject-specific deviations align with established MDD-related behavioral biomarkers.