On Dynamic Treatment Regimes: Collaborative Search and LLM-Driven Decision Trees

Abstract

This thesis evaluates the validity of current dynamic treatment regime algorithms and presents a novel data structure for extracting treatment decisions from unstructured clinical notes. The main contribution is the Clinical Decision Tree (CDT) which uses large language models (LLMs) to extract key decisions in chronic disease treatment. This addresses the main pain points in dynamic treatment regimes of low interpretability and reliance on poorly collected data for traditional machine learning methods. This work contains extensive experiments on mortality prediction, time series forecasting, and synthetic patient modeling. Experiments show that vital-based representations do not capture enough meaningful data about a patient to accurately predict and evaluate new treatment methods. By utilizing latent embeddings and vector search, experiments show that the collected vitals of patients fail to differentiate the outcomes of the related patients. Conversely, the clinical notes contain complex and substantial information about clinical decision making. LLMs enable the valuable knowledge extraction from unstructured data. Utilizing LLMs, experimental results and expert evaluation indicates that CDTs can extract and distill interpretable treatment decisions. Thus, CDTs are a valuable tool that can be refined to increase confidence in treatment decisions and identifying rare and uncommon medical practices.