💻 computer science

ChainTrust-LLM: Secure and Auditable Hallucination Mitigation in Medical LLMs Using Blockchain and Expert Feedback

This paper introduces ChainTrust-LLM, a novel framework that integrates expert feedback with a DAG-based blockchain ledger to detect and mitigate hallucinations in medical Large Language Models, achieving a 64.8% reduction in error rates while ensuring secure, auditable interactions with minimal latency.

Original authors: Muhammad Ismail Mohamand

Published 2026-07-29

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Muhammad Ismail Mohamand

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a world where you can talk to a super-smart robot librarian that has read every medical textbook ever written. This robot, known as a Large Language Model (LLM), is amazing at chatting, answering questions, and even helping doctors explain things to patients. But here's the catch: sometimes, this robot gets a little too confident and makes things up. It might tell you that a certain medicine cures a disease it doesn't, or that a symptom means something it doesn't. In the world of medicine, these made-up facts are called "hallucinations," and they are dangerous because they can lead to bad decisions.

To fix this, scientists are trying to build a system that acts like a strict editor. They want a way to check the robot's work, keep a permanent record of every correction, and make sure the robot learns from its mistakes without forgetting them. This is where the idea of "trust" comes in. If the robot gets a question right, we trust it more; if it gets it wrong, we trust it less. But keeping track of all this trust and making sure no one can secretly change the records is a huge challenge. That's why researchers are looking at combining human experts with a special kind of digital ledger called "blockchain," which is like a public notebook that no one can erase or alter once something is written in it.

This is exactly what the paper "ChainTrust-LLM" is all about. The researchers, led by Muhammad Ismail Mohmand and his team, wanted to solve the problem of medical hallucinations, specifically for a condition called hypothyroidism (a problem with the thyroid gland). They noticed that when robots talk about symptoms like feeling tired, gaining weight, or having trouble with menstrual cycles, they often get the details wrong. To fix this, they built a new framework called ChainTrust-LLM. Think of it as a high-tech safety net that logs the robot's mistakes and corrections securely, creating an auditable trail to improve future performance.

Here is how their system works, using a simple story: Imagine the medical robot is a student taking a test. Every time it answers a question, three real-life doctors (experts) check the answer. If the student is right, they give a thumbs-up; if they are wrong, they give a thumbs-down and write down the correct answer. But instead of just keeping these notes in a messy pile of paper, ChainTrust-LLM writes every single grade and correction into a "digital diary" that uses blockchain. This diary is unchangeable, meaning once a mistake is recorded, it stays there forever as proof.

The system also has a special rule about time. If the robot gets a question right today, its "trust score" goes up. But if it doesn't get checked for a long time, that trust score slowly fades away, like a battery running low. This ensures the system stays fresh and doesn't rely on old, potentially outdated information. When the robot makes a mistake, the system uses a "risk detector" to spot the lie based on how different the answer is from the truth, and then it logs the whole event securely.

The team tested this idea on three famous medical robots: GPT-4, MedPaLM, and Claude. They asked them 300 different questions about hypothyroidism symptoms, treatments, and lab tests. Before using ChainTrust-LLM, these robots were making mistakes quite often. For example, GPT-4 was hallucinating (making things up) about 23.1% of the time. After the new system was applied, that number dropped dramatically to 8.5%. That is a reduction of about 63%. For the other robots, the improvements were just as big, with mistake rates dropping by up to 64.8%.

Not only did the robots make fewer mistakes, but the system also got much better at knowing when it was right or wrong. The "trust calibration accuracy"—which is basically how well the robot's confidence matches reality—jumped from around 75% to over 91% for GPT-4. The experts who checked the answers also agreed with each other much more often, with their agreement score rising from 0.65 to 0.87. This means the system wasn't just fixing errors; it was creating a reliable, shared understanding of the truth.

One of the coolest parts of this research is how fast it works. Even with all the extra checking and the blockchain recording happening, the system only added a tiny delay. On average, it took just 211 milliseconds (that's less than a quarter of a second) to log a transaction. This suggests that such a system could actually be used in real hospitals without slowing things down.

The paper concludes that ChainTrust-LLM is a strong, secure way to make medical AI safer. By combining human experts, a time-based trust system, and an unchangeable blockchain record, they created a framework that significantly reduces dangerous lies in medical advice. While the study focused specifically on hypothyroidism symptoms like fatigue and weight gain, the method suggests a path forward for making AI trustworthy in other medical areas too. It's not a magic wand that fixes everything instantly, but it is a very promising step toward a future where we can trust our digital medical helpers.

Technical Summary: ChainTrust-LLM

1. Problem Statement

Large Language Models (LLMs) such as GPT-4, Med-PaLM, and Claude have demonstrated significant potential in clinical text generation, diagnostic assistance, and patient education. However, their integration into real-world medical settings is hindered by the phenomenon of hallucination—the generation of factually incorrect, fabricated, or misleading outputs presented with high confidence. In the medical domain, these errors pose severe risks to patient safety and clinical reliability.

Existing mitigation techniques, including Retrieval-Augmented Generation (RAG), structured knowledge injection, and Reinforcement Learning from Human Feedback (RLHF), often lack transparency, traceability, and domain-specific robustness. Furthermore, current literature frequently treats hallucination as a universal problem rather than addressing its specific implications for particular diseases. This paper identifies a critical gap in detecting and mitigating hallucinations related to hypothyroidism, specifically focusing on subtle and easily confused symptoms such as fatigue, weight gain, and menstrual health disruptions. Misinformation in these areas can lead to misdiagnosis or missed treatment, yet few systems address domain-specific symptom detection with secure auditability.

2. Methodology: The ChainTrust-LLM Framework

The authors propose ChainTrust-LLM, a novel framework that integrates human-in-the-loop trust calibration with blockchain-based auditability to detect and mitigate hallucinations. The system is designed around four core modules:

2.1 Dataset Construction

A disease-specific dataset was developed comprising 300 curated medical prompts divided into three categories: symptoms (focusing on fatigue, weight gain, and menstrual health), medication/treatment, and laboratory testing/diagnosis. Ground truths were established using authoritative sources (NIDDK, CDC, ATA, MedlinePlus, DailyMed) and verified by three experienced endocrinologists. The dataset includes binary hallucination labels, severity scores, and trustworthiness scores.

2.2 System Architecture

The framework processes inputs through the following pipeline:

Prompt Encoder Module: Utilizes a domain-tuned language encoder (e.g., BioBERT or ClinicalBERT) to transform clinical queries into structured semantic embeddings. A softmax classifier routes prompts into specific domains (symptom, medication, or lab) to adjust hallucination thresholds accordingly.
Hallucination Detection Module: Implements a binary risk classifier $R(x)$ that identifies hallucinated responses. It employs semantic contradiction scoring ( $C_i = 1 - \text{Sim}(y_i, y^*_i)$ ) to measure factual similarity between the LLM output and verified reference answers. Features include linguistic overlap, negation detection, and clinical entity matching (ICD-10, drug mentions).
Trust Scoring Model: Introduces a time-weighted trust update mechanism that adapts to expert feedback. The trust score $T(y, t)$ is updated based on new human feedback $f_t(y)$ and a previous trust score $T_{prev}(y)$ , balanced by a memory parameter $\alpha$ . Crucially, the model incorporates exponential trust decay ( $e^{-\lambda \Delta t}$ ) to reduce trust scores over time in the absence of recent feedback, ensuring domain-specific reliability assessment.
Blockchain Auditing Layer: To ensure tamper-proof tracking, all response-feedback pairs are hashed and stored on a private, permissioned DAG-based blockchain (implemented on a permissioned Ethereum network using geth). Each block entry includes the trust score, feedback label, and timestamp, linked via cryptographic hash pointers to guarantee integrity and traceability.

2.3 Security and Formalization

The paper formalizes the robustness of hallucination detection against adversarial perturbations and defines trust chain integrity using DAG-based hash pointers. Access control is enforced via digital signatures from permitted expert users, and replay attacks are prevented using nonces.

3. Key Contributions

The paper outlines three primary contributions:

Human-in-the-Loop Trust Calibration: A framework for hallucination detection specifically tailored to hypothyroidism-related responses, utilizing iterative expert feedback to update trust scores.
Blockchain-Audited Feedback Loop: A secure logging mechanism that ensures the traceability and immutability of hallucination corrections and trust assessments.
Time-Weighted Trust Decay Model: A novel approach for assessing the reliability of LLM responses over time, accounting for the decay of trust in the absence of continuous expert validation.

4. Experimental Results

The framework was evaluated using GPT-4, Med-PaLM, and Claude across the 300 prompt dataset. Key performance metrics include:

Hallucination Rate Reduction: ChainTrust-LLM significantly reduced hallucination rates across all models:
- GPT-4: Decreased from 23.1% to 8.5% (63.2% reduction).
- Med-PaLM: Decreased from 28.7% to 10.1% (64.8% reduction).
- Claude: Decreased from 31.9% to 12.4% (61.1% reduction).
Trust Calibration Accuracy (TCA): The alignment between model trust predictions and expert ratings improved substantially:
- GPT-4: 75.4% $\rightarrow$ 91.2%
- Med-PaLM: 70.3% $\rightarrow$ 89.5%
- Claude: 67.8% $\rightarrow$ 87.1%
Inter-Rater Agreement: Fleiss' Kappa ( $\kappa$ ) improved from a baseline average of 0.65 to 0.87, indicating strong agreement among experts on the corrected outputs.
Symptom-Specific Detection: Accuracy in detecting hallucinations for specific symptoms improved significantly:
- Fatigue: 68% $\rightarrow$ 92%
- Weight Gain: 64% $\rightarrow$ 89%
- Menstrual Health: 61% $\rightarrow$ 87%
Classifier Performance: The hallucination risk classifier achieved an AUC of 0.934, demonstrating high sensitivity and specificity.
Blockchain Overhead: The system maintained low latency, with an average transaction latency of 211 ms and minimal gas costs, supporting real-time logging capabilities.

5. Significance and Claims

The authors claim that ChainTrust-LLM offers a robust, explainable, and safe solution for deploying medical language support models. Unlike previous post-hoc detectors, this framework operates as a real-time, feedback-based mitigation loop. By combining formal trust modeling with blockchain transparency, the system addresses the critical need for auditability in high-stakes medical AI.

The study emphasizes that while general hallucination detection is an active field, specific attention to underexplored clinical domains like hypothyroidism symptoms is vital. The integration of a time-decay trust model and immutable audit trails provides a mechanism to ensure that LLMs remain reliable over time, even as medical knowledge evolves or feedback patterns change. The authors conclude that the framework is a significant step toward trustworthy AI in healthcare, with future work planned to expand to other symptoms and utilize federated learning for broader applicability.