Integrating LLMs and Causal Inference: Comparing Oral Anticoagulant Effects on Thrombosis Recurrence, Bleeding Risk and Death Using MIMIC-IV Data

<p>Many clinically important variables are found only in medical notes, leaving gaps for prediction and confounding adjustment. We analyzed medical discharge notes for acute-thrombosis admissions in MIMIC-IV. Privacy was preserved by generating synthetic medical notes that mirror the content and structure of the original medical discharge notes, labelling them with reasoning steps and classifications using the DeepSeek-R1 API, and fine-tuning 8–14 billion parameter versions of Llama-3, Qwen-3 and Gemma-3 models with Group Relative Policy Optimization (GRPO). Qwen-3 achieved the highest accuracy on a hold-out set of synthetic notes and de-identified real medical notes.</p> <p>Variables extracted from medical notes, including family history of clot and provoked versus unprovoked status, were combined with structured fields and passed to a Super Learner formed from non-parametric, semi-parametric and tree learners. The ensemble attained the lowest negative log-likelihood (NLL) when predicting major bleeding and thrombosis recurrence within 3 and 6 months and mortality within 12 months; adding discharge medical note-level covariates reduced loss by 0.2% – 1.7% across outcomes.</p> <p>Causal effects of Vitamin K Antagonist (VKA) which was primarily Warfarin, and Factor Xa inhibitors (Apixaban, Edoxaban, Rivaroxaban) were estimated with targeted maximum likelihood estimation (TMLE), which uses initial Super Learner fits for the outcome and treatment mechanisms followed by a targeting step that achieves doubly robust, efficient inference. After adjustment for both structured and medical notes-text covariates, VKA increased the risk of major bleeding by 4.5% with 95% C.I of (3.4% – 5.7%) at 3 months and 5.8% with 95% C.I (4.6% – 7.0%) at 6 months and raised thrombosis recurrence by 3.2% with 95% C.I (1.8% – 4.6%) and 3.5% with 95% C.I (2.1% – 5.0%) over the same horizons. No significant difference in 12-month mortality was detected with 95% C.I (–1.7 % to 0.6%).</p> <p>These findings show that large language models (LLMs) extraction of discharge summaries can improve risk prediction and strengthen confounding control in targeted learning. Incorporating such variables reveals lower short-term risks of bleeding and recurrence with Factor Xa inhibitors compared with VKA while long-term survival remains similar.</p>

Name thesis:degree_name

Doctor of Philosophy (PhD)

Level thesis:degree_level

Dissertation

Discipline thesis:degree_discipline

Computational and Data Sciences

Year dc:date.available

2025

Author dc:creator

• Ofosu, Duncan K

Contributors dc:contributor

• Cyril Rakovski, Ph.D
• Adrian Vajiac, Ph.D
• Ehsan Yaghmaei, Ph.D
• Sheth Parthiv, MD

• targeted learning
• causal inference
• super learner
• llm
• grpo
• reinforcement learning
• Biostatistics
• Data Science
• Disease Modeling
• Diseases
• Statistical Models

Repository record dc:identifier

https://digitalcommons.chapman.edu/cads_dissertations/50

OAI identifier oai:identifier

oai:digitalcommons.chapman.edu:cads_dissertations-1051