Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Building a Human-Verified Clinical Reasoning Dataset via a Human LLM Hybrid Pipeline for Trustworthy Medical AI

Chao Ding, Mouxiao Bian, Pengcheng Chen, Hongliang Zhang, Tianbin Li, Lihao Liu, Jiayuan Chen, Zhuoran Li, Yabei Zhong, Yongqi Liu, Haiqing Huang, Dongming Shan, Junjun He, Jie Xu

TL;DR

This work tackles the lack of transparent reasoning in medical LLMs by introducing a human–LLM hybrid pipeline to produce a clinically validated dataset with expert‑verified chain‑of‑thought explanations. Seed medical exam questions are expanded via an LLM to generate thousands of QA pairs, then subjected to a multilayer review where clinicians and AI iteratively refine and verify content, culminating in consensus approval. The dataset finalizes at around 30,000 items (with an initial expansion to ~36,210) and is evaluated using a five‑dimensional rubric across medical correctness, reasoning structure, information sufficiency, terminology clarity, and clinical utility, demonstrating improvements in both answer accuracy and reasoning quality when models are trained on this data. The resource is positioned to advance safe, explainable medical AI by providing high‑quality, interpretable reasoning examples for model development and benchmarking, and is publicly available to catalyze further research and deployment in clinical settings.

Abstract

Despite strong performance in medical question-answering, the clinical adoption of Large Language Models (LLMs) is critically hampered by their opaque 'black-box' reasoning, limiting clinician trust. This challenge is compounded by the predominant reliance of current medical LLMs on corpora from scientific literature or synthetic data, which often lack the granular expert validation and high clinical relevance essential for advancing their specialized medical capabilities. To address these critical gaps, we introduce a highly clinically relevant dataset with 31,247 medical question-answer pairs, each accompanied by expert-validated chain-of-thought (CoT) explanations. This resource, spanning multiple clinical domains, was curated via a scalable human-LLM hybrid pipeline: LLM-generated rationales were iteratively reviewed, scored, and refined by medical experts against a structured rubric, with substandard outputs revised through human effort or guided LLM regeneration until expert consensus. This publicly available dataset provides a vital source for the development of medical LLMs that capable of transparent and verifiable reasoning, thereby advancing safer and more interpretable AI in medicine.

Building a Human-Verified Clinical Reasoning Dataset via a Human LLM Hybrid Pipeline for Trustworthy Medical AI

TL;DR

This work tackles the lack of transparent reasoning in medical LLMs by introducing a human–LLM hybrid pipeline to produce a clinically validated dataset with expert‑verified chain‑of‑thought explanations. Seed medical exam questions are expanded via an LLM to generate thousands of QA pairs, then subjected to a multilayer review where clinicians and AI iteratively refine and verify content, culminating in consensus approval. The dataset finalizes at around 30,000 items (with an initial expansion to ~36,210) and is evaluated using a five‑dimensional rubric across medical correctness, reasoning structure, information sufficiency, terminology clarity, and clinical utility, demonstrating improvements in both answer accuracy and reasoning quality when models are trained on this data. The resource is positioned to advance safe, explainable medical AI by providing high‑quality, interpretable reasoning examples for model development and benchmarking, and is publicly available to catalyze further research and deployment in clinical settings.

Abstract

Despite strong performance in medical question-answering, the clinical adoption of Large Language Models (LLMs) is critically hampered by their opaque 'black-box' reasoning, limiting clinician trust. This challenge is compounded by the predominant reliance of current medical LLMs on corpora from scientific literature or synthetic data, which often lack the granular expert validation and high clinical relevance essential for advancing their specialized medical capabilities. To address these critical gaps, we introduce a highly clinically relevant dataset with 31,247 medical question-answer pairs, each accompanied by expert-validated chain-of-thought (CoT) explanations. This resource, spanning multiple clinical domains, was curated via a scalable human-LLM hybrid pipeline: LLM-generated rationales were iteratively reviewed, scored, and refined by medical experts against a structured rubric, with substandard outputs revised through human effort or guided LLM regeneration until expert consensus. This publicly available dataset provides a vital source for the development of medical LLMs that capable of transparent and verifiable reasoning, thereby advancing safer and more interpretable AI in medicine.
Paper Structure (17 sections, 4 figures, 1 table)

This paper contains 17 sections, 4 figures, 1 table.

Table of Contents

  1. Introduction
  2. Methods
  3. Source Data and Initial Question Set
  4. LLM-Generated Reasoning and Dataset Expansion
  5. Multi-Stage Expert–AI Validation Pipeline
  6. Initial Human Review
  7. AI Re-Answering and Verification
  8. Five-Strike Error Trigger for Expert Panel
  9. Expert Panel Refinement
  10. Clinical Consensus Approval
  11. Results
  12. Dataset Content and Statistics
  13. Quality of Explanations (Rubric Evaluation)
  14. Baseline Model Performance and Utility of the Dataset
  15. Discussion
  16. ...and 2 more sections

Figures (4)

  • Figure 1: Deployment Landscape of DeepSeek in the Medical Domain Across China. This figure summarizes the adoption and application of the DeepSeek large language model in healthcare institutions nationwide. (Left): A pie chart showing the geographical distribution of over 700 medical institutions that have integrated DeepSeek. The top three provinces with the highest deployment density are Sichuan, Guangxi, and Guangdong. (Right): A radial sunburst chart illustrating the key application domains, including auxiliary diagnosis, report interpretation, intelligent triage, and medical record quality control. Each segment shows representative use cases and affiliated medical institutions across different clinical scenarios.
  • Figure 2: Human–LLM hybrid pipeline for constructing a clinically validated medical QA dataset with reasoning chains. The figure outlines the multi-stage workflow used to generate and validate a large-scale Chinese medical QA dataset. Starting from 3,621 seed questions drawn from national medical examinations, DeepSeek-R1 was prompted to generate step-by-step chain-of-thought (CoT) explanations. These rationales were used to synthesize 30,000 QA pairs. Medical experts then reviewed the QA items for correctness and clarity. Each item was re-answered by the LLM up to five times to detect formulation flaws. Items triggering the five-strike error mechanism underwent expert panel review, where CoTs were re-generated and scored across five dimensions: medical correctness, reasoning structure, information sufficiency, terminology clarity, and clinical utility. The validated dataset serves as a benchmark for training and evaluating trustworthy medical AI systems.
  • Figure 3: Distribution of clinical specialties represented in the medical QA dataset. The pie chart shows the proportional representation of 12 medical disciplines in the dataset. General Surgery accounts for 22.2% of all items, while Rheumatology, Endocrinology, and Oncology each contribute between 8–8.2%. A range of other core departments—including Cardiovascular Medicine, Nephrology, Gastroenterology, Hematology, and Emergency Services—are also represented, ensuring clinical breadth across both internal medicine and procedural specialties.
  • Figure 4: Five-dimensional expert evaluation of CoT quality.Visualizes the five-dimensional quality assessment of CoT explanations in the constructed medical QA dataset. Each CoT instance was scored across five expert-defined dimensions: medical correctness, reasoning structure, information sufficiency, terminology clarity, and clinical utility.