PhD Thesis Defense: Vincent Chen

ZOOM LINK

"Fair Decision Making in Healthcare"

Abstract

Fair healthcare decision-making seeks outcomes close to those achieved by an oracle policy that maximizes expected benefit without fairness constraints. It also aims to reduce unjustified variation in decisions and outcomes. Existing optimization and machine learning approaches often lack explicit mechanisms to promote comparable decisions for patients with similar clinical needs. This dissertation examines fairness across healthcare decision-making, from conceptual frameworks to treatment planning and patient routing.

First, a systematic review examines fairness in healthcare optimization and machine learning. It screens 801 references and identifies 114 relevant studies, categorizing sources of bias, fairness metrics, and mitigation strategies. The findings establish a foundation for choosing fairness criteria according to clinical and operational contexts.

Second, we develop fair distributional reinforcement learning for individualized hypertension treatment planning. The algorithm models distributions of long-term outcomes and applies Cram.r-distance regularization to patients with comparable clinical characteristics. An additional post-update correction uses constrained interpolation or fallback mixing to move estimated return distributions toward high performing subgroup references. Independent simulations evaluate learned policies using quality-adjusted life years (QALYs). Compared with standard distributional reinforcement learning, QALYs for representative vulnerable patients increase from 29.75 to 31.32 in the low-risk group. Corresponding increases are from 24.46 to 25.86 and from 17.06 to 18.69 in intermediate- and high-risk groups, respectively.

Third, we develop a fairness-aware contextual bandit for assigning triage priorities in a medical guidance application. The application recommends appropriate care levels. The framework selects low, medium, or high priority while balancing clinical need, resource use, and local fairness constraints. Its policy adapts patient representations and subgroup estimates using feedback, and the selected priority guides response generation. Across 83,604 held-out emergency-department visits, priority conditioning improves evaluated response quality by 8.9% to 9.7% across three model families. The largest improvement is 15.1% among medium-priority patients.

Together, these studies show how clinical comparability can inform fairness criteria, improve simulated treatment outcomes, and guide urgency-sensitive patient communication. They connect fairness in healthcare to both downstream treatment planning and upstream patient routing.

Thesis Committee

• Wesley Marrero (chair)
• Eugene Santos Jr
• Peter Chin
• Shuhang Wang (Walmart Global Tech)

Contact

For more information, contact Thayer Registrar at thayer.registrar@dartmouth.edu .