Abstract
<title>Abstract</title> <p>Large Language Models have shown remarkable promise in clinical text generation but continue to suffer from hallucination, factually incorrect or misleading outputs posing serious risks in medical decision-making. This paper introduces ChainTrust-LLM, a novel framework that integrates human-in-the-loop trust calibration and blockchain-based auditability to detect and mitigate hallucinations in medical LLMs. Focusing on hypothyroidism related symptoms, i.e., fatigue, weight gain, and menstrual health. Our approach leverages expert feedback to iteratively update trust scores through a time-weighted decay model and logs interactions securely using a DAG-based blockchain ledger. We formalize hallucination as a binary classification problem and propose a risk detection classifier built on semantic contradiction scoring. Extensive evaluation using GPT-4, MedPaLM, and Claude across 300 curated medical prompts demonstrates that ChainTrust-LLM reduces hallucination rates by up to 64.8%, improves trust calibration accuracy by 23.4%, and enhances inter-rater agreement (Fleiss’ k) from 0.65 to 0.87. Additionally, we show minimal blockchain overhead with an average latency of 211 ms per transaction. These results confirm ChainTrust-LLM as a robust and auditable framework for improving the trustworthiness of LLMs in clinical applications.</p>