Faithful Rule Extraction from Large Language Models for High-Stakes Decision Support: A Pilot Study on Clinical Trial Eligibility

This preprint presents RULEDISTILL, a post-hoc explainability framework for improving the transparency and trustworthiness of Large Language Model (LLM)-based clinical trial eligibility assessment. The proposed framework transforms free-text Chain-of-Thought (CoT) reasoning generated by LLMs into structured symbolic IF–THEN rules, enabling more interpretable, auditable, and clinically meaningful decision support for high-stakes healthcare applications. To evaluate the quality of the extracted rules, this work introduces the Faithfulness Score (F), a novel metric designed to quantify the semantic alignment between symbolic rules and the underlying LLM reasoning through semantic similarity and logical grounding heuristics. The framework is evaluated in a pilot study using 50 manually annotated eligibility triplets derived from 92 real clinical trials retrieved from ClinicalTrials.gov. Experiments were conducted using Llama-3-8B-Instruct, deployed locally via Ollama, together with synthetic patient profiles for eligibility assessment. Experimental results demonstrate that RULEDISTILL achieves a mean Faithfulness Score of 0.714 while preserving approximately 93% of the predictive performance of the original LLM reasoning process. These findings indicate that symbolic rule extraction can substantially improve the interpretability of LLM-generated clinical decisions without significantly compromising predictive capability. This work represents an initial proof-of-concept toward trustworthy and explainable AI for clinical decision support. The methodology has been designed with reproducibility in mind, and the accompanying data collection scripts, preprocessing pipeline, and evaluation framework are made publicly available to facilitate future research, independent validation, and extension by the broader scientific community.