Causal Inference and Evidence-Grounded Language Models for Trustworthy Personalized Clinical Decision Support

Clinical decision support systems (CDSS) are evolving from passive predictive tools into active collaborators in clinical reasoning. However, most machine learning approaches remain limited to risk prediction, lacking the causal reasoning, patient-specific personalization, and evidence-verifiable justification required for high-stakes medical decision-making. This thesis presents a principled framework for advancing CDSS from prediction to actionable clinical intelligence through three interconnected pillars: personalization, evidence grounding, and trustworthiness. First, we address personalization through causal machine learning, moving beyond population-level estimates toward individualized treatment effects. We introduce a treatment-response–aware phenotyping framework that integrates post-intervention outcomes into patient clustering, enabling the discovery of subgroups defined by differential treatment response. Building on this, we propose X-MultiTask, a meta-learning architecture for multi-valued treatment effect estimation that leverages shared representations and inverse probability weighting to model counterfactual outcomes across treatment options. We further develop FUSION, a unified framework combining latent phenotyping, calibrated risk prediction, and causal inference to forecast intraoperative events and their downstream physiological impact, supporting personalized surgical decision-making. Second, we establish evidence-grounded reasoning as a critical bridge between personalized predictions and actionable recommendations. We introduce Evidence-Guided Diagnostic Reasoning (EGDR), a structured pipeline that decomposes clinical reasoning into verifiable steps grounded in clinical knowledge graphs, improving diagnostic accuracy and ensuring traceability to medical evidence. To embed grounding directly into model training, we propose EvidenceRL, a reinforcement learning framework that jointly optimizes diagnostic correctness and evidence alignment using a novel entailment-based reward. This approach significantly reduces hallucinations while increasing the proportion of evidence-supported predictions, achieving strong performance across multiple large language model backbones. Finally, we address trustworthiness as a prerequisite for safe clinical deployment by targeting three key challenges: bias, overconfidence, and opacity. We develop FAIR-MTL, a fairness-aware multi-task learning framework that mitigates performance disparities across demographic groups under distributional shift. We introduce an Uncertainty-Aware Ensemble (UAE) to quantify predictive uncertainty and enable reliable deferral in high-risk scenarios. Additionally, we propose the Diagnostic Confidence Score (DCS), a model-agnostic metric that evaluates both factual grounding and logical consistency of clinical reasoning. Together, these contributions form a unified framework for personalized, evidence-grounded, and trustworthy CDSS. By integrating causal inference, verifiable reasoning, and reliability mechanisms, this work advances clinical AI from predictive analytics toward transparent, patient-specific decision support capable of safely augmenting clinical expertise.

Name thesis:degree_name

Machine Learning, PhD

Grantor

Georgia Institute of Technology

Year dc:date.issued

2026

Author dc:creator

• Ben Tamo, Junior

Advisor dc:contributor.advisor

• Wang, May Dongmei

Committee members dc:contributor.committeemember

• Cassie Mitchell
• Anderson, David
• Heck, Larry
• Brenn, Randall

Handle dc:identifier.uri

https://hdl.handle.net/1853/81788

OAI identifier oai:identifier

oai:repository.gatech.edu:1853/81788