Multidisciplinary blinded randomized expert evaluation of large language models for clinical diagnosis and management Authors: Peikai Chen, Jifu Cai, Jiaying Zhou, Shaoxi Chen, Chenguang Xu, Lihua Yuan, Xiaoying Dai, Xiaowei Chen, Yanzhe Wei, Xia Li, Shaofeng Gong, Xiaolong Liang, Jiancheng Yang, Jun Jin, Kanglin Dai, Yuzhen Cui, Guan-Ming Kuang, Jiansheng Xie, Libing Luo, Haibing Xiao, Shijie Yin, Jun Yang, Yulan Yan, Jianliang Chen, Yihua Chen, Qianshen Zhang, Qingshan Zhou, Lina Zhao, Min Wu, Xin Tang, Lei Rong, Zanxin Wang, Weifu Qiu, Yanli Wang, Liwen Cui, Xiangyang Li, Yong Hu, Huiren Tao, Nan Wu, David J. H. Shih Published: 2026 ABSTRACT BackgroundDirect clinical uses of large language models (LLMs) remain controversial, partly because of the lack of methodological rigor in assessing their risks and benefits in medicine.MethodsWe developed Medieval, a multidisciplinary, randomized, and blinded expert evaluation framework. A ten-point Dreyfus-based scoring scale linked to career stages of human physicians was designed to reflect response qualities. Seven advanced LLMs or their distilled versions that were released within a short time-frame ( ≤ 45 days) in early 2025 were tested. Incidence of fabricated medical facts were documented. Linear mixed-effects models and variance-stabilizing Bayesian generalized linear mixed models were employed to perform statistical analyses.ResultsWe first develop a high-quality question bank comprising 685 real and simulated clinical cases across 13 specialties. An expert panel of 27 clinicians (average years of services: 25.9) evaluated the 4795 model responses. We show that these LLM ratings (n = 9856) have excellent reliability (intraclass correlation coefficients \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ > $$\end{document}>0.9). Among the seven LLMs tested, Gemini 2.0 Flash achieved the highest raw scores. However, after adjusting for confounders, DeepSeek-R1 was the top-performing model with a mean score of 6.36 (95% confidence interval 6.03 − 6.69), a performance level equivalent to an early-career physician. Despite these strengths, 3–19% LLM responses were rated as incompetent and 40 instances of LLM hallucination were also identified.ConclusionsOur study shows that in spite of LLMs’ substantial potentials in medicine, their unguarded clinical application could present serious risks, which must be continuously monitored by human expert panels. The evaluation framework developed and validated in this study will facilitate such efforts. Introduction Top-ranked general-purpose large language models (or advanced LLMs1), such as ChatGPT, Claude, Grok, DeepSeek, and Qwen2, have demonstrated strong performance not only in general language tasks, but also in domain-specific reasoning, such as medicine3–5. These LLMs have promise in clinical applications including diagnosis, treatment planning, and patient interactions6,7. They have also been applied in clinical sub-specialties, including ophthalmology8, clinical oncology9,10, clinical genetics11, mental health12, neurology13,14, and orthopedics15. Additionally, LLMs can facilitate the diagnosis of rare diseases16,17. Despite these potential clinical applications, real-world adoption of LLMs in the clinical setting remain limited, due to a number of significant challenges, including factual inaccuracies18, hallucinations19, as well as cognitive and automation bias20. In the research setting, LLMs have been extensively benchmarked for clinical applications. These approaches include MMLU21, MedQA6,22, and MedMCQA23, where both questions and answers are available for training. Benchmarking against publicly available benchmark datasets often lead to model overfitting and optimistic bias of model performance, since data labels are readily available for training. Board-exam-style clinical questions have also been used to assess LLM performance in rheumatology24. However, these questions lack the complexity observed in real-world clinical cases. Moreover, the reliance on multiple-choice formats precludes an accurate assessment of clinical reasoning. Crowdsourced preference-based ranking platforms (.e.g ARENA25) have also been developed but the ratings are not provided by qualified medical experts, undermining the credibility of the ratings. Randomized controlled trials (RCTs) have also been conducted to compare the use of LLM vs. conventional tools (e.g. web-search or UpToDate)26. While RCTs remain essential for evidence-based medicine, their current design and implementation lag behind rapidly evolving technologies such as LLMs. To address these, we developed Medieval, a multidisciplinary expert evaluation framework, wherein: 1) real local clinical cases instead of public exam questions are used; 2) only qualified specialists are involved in scoring qualities of the open-ended responses; 3) a career stage equivalence metrics was used instead of Boolean or categorical ratings, and; 4) model identities of the responses are blinded. Applying Medieval to a cross-specialty bank of real clinical cases, we identify factors affecting LLM scores and clinical plausibility of LLMs adjusting for confounders. The strategies developed in this study has the potential to be deployed for multi-centered, larger-scale assessment schemes. Methods Medieval included three stages (Fig. 1). First, a multidisciplinary panel developed a question bank of real representative clinical cases. Second, LLMs were prompted with these cases and their responses were rated by clinical experts in a blinded manner. Third, LLM-generated responses were analyzed by advanced statistics.Fig. 1Overall study design of Medieval.D: the total number of disciplines.M: the number of models. N: number of questions. HIS: Hospital information system. Overall study design of Medieval. D: the total number of disciplines.M: the number of models. N: number of questions. HIS: Hospital information system. Development of a question bank of real cases Two types of questions were designed: (1) departmental OSCE (Objective Structured Clinical Examination) questions, (2) real clinical cases. The OSCE is an established instrument of clinical competence27 used for specialty board qualification examinations. The real cases were retrieved from the local hospital information system, from experienced clinicians to represent typical cases that are commonly encountered, as well as unusually rare and difficult cases. Key clinical information including age, gender, history of presenting complaint, symptoms, physical examinations, past medical history, family history, drug prescriptions, laboratory tests, radiological reports, genetic tests, etc. forms key components of the textual question stems. These multidisciplinary cases form a question bank, wherein each item was further reviewed by both the designer panel and an independent evaluator panel for adequacy and relevance, including level of difficulty (easy, medium, and hard), prevalence of the diseases (common or rare), and involvement of diagnostic, treatment, and operational reasoning. Model selection criteria We used the ARENA25 top-ranked general-purpose models (Feb 27, 2025, Supplementary Fig. S1B) as a starting-point to choose candidate models. We further enforce these restrictions: 1) models with API access; 2) models that were released within a short time frame of each other to avoid generational advantages of newer models; 3) avoiding excessive numbers of models (10 or more) to lower evaluation fatigue. As such, we short-listed to 4 advanced models and 3 smaller models (distilled versions of their advanced counterparts). They are: Gemini 2.0 Flash Thinking exp-01-21 (Gemini2F), ChatGPT 4o latest 2025-01-29 (ChatGPT-4oL), DeepSeek-R1 671B (DeepSeek R1, or DS-R1 671B), DeepSeek-R1-Distill-Llama-70B(DS70B), DeepSeek-R1-Distill-Qwen-32B (DS32B), Qwen 2.5 Max (Qwen Max), and Qwq Latest (Qwen 32B) (Supplementary Information S1.17). Grok was not included due to the unavailability of APIs at the time of the study. The models were released no more than 45 days apart. A ten-point scale of LLM qualities We develop a ten-point scale of LLM qualities based on the Dreyfus developmental model28, which spans novice, beginner, to expert and master. According to this physician-equivalence scale (Supplementary Information S1.2), we assigned: 0-1.5 for the medically untrained; 2–3.5 for novice (medical students); 4–5.5 for beginners (medical residents, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ < 3$$\end{document}<3 years after graduation); 6–6.5 for competent physicians (fellow, or physician, 3–6 years after graduation); 7-8 for proficient physicians (associate consultant or associate specialist, 6–10 years after graduation); 8.5–9 for experts (consultants, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ > 10$$\end{document}>10 years after graduation); and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\ge 9.5$$\end{document}≥9.5 for masters (leading experts). A mapping of the physician career stages in different medical training systems is provided in Supplementary Information S1.2. Evaluation of LLM responses After de-identification, a standardized prompt of such was designed: “You are a specialist in [discipline], please provide a professional diagnosis, treatment or surgical plan for the above case.” No additional restriction (e.g. verbosity of the LLM response) was imposed. Two examples of question prompts are included in Supplementary Information S1.16 for reference. Cases across all specialties were then aggregated, shuffled, and submitted to each LLM via application programming interfaces (APIs). The responses were then evaluated by the panel using the scale above. A number of measures were enforced to ensure rigor. First, single-blinding was enforced. Identifiable features such as emojis, excessive line breaks, and whitespace were removed to further mask model identity. To ensure blinding integrity, case-response and response-LLM mappings were encrypted and shared with evaluators. Decryption keys were provided upon project completion to allow independent verification of blinding and confirmation of the absence of tampering. Second, bias-reduction measures were introduced. During the blinded evaluation, LLM responses to each case were displayed side-by-side electronically and in a randomized manner (Supplementary Fig. S1B). Additionally, the cases were shuffled between evaluation rounds to mitigate fatigue-related bias. Third, upon completion of the evaluation, evaluator consistencies were calculated within each specialty. Evaluators were alerted of potential inconsistencies (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$r < 0.3$$\end{document}r<0.3) in their ratings, and they were given the option to adjust or keep their original rating. The evaluator panel also rated each LLM response for the degree of hallucination based on a scale from 0 to 5, where 1 indicates any evidence of fabricated diseases, drugs, guidelines, consensus, and reference etc. A hallucination rating of 1 or above must be substantiated with solid evidence of fabrication, or it was adjusted to 0.5. Statistical and reproducibility The reliability of evaluator scores was assessed by estimating intraclass correlation coefficients (ICC) for absolute agreement under two-way random effect models29. ICC in the intervals <0.5, 0.5 to 0.75, 0.75 to 0.9, and >0.9 were considered poor, moderate, good, and excellent, respectively. Each question was rated by between 1 and 3 (average 2.07) independent evaluators. Linear mixed-effects model and logistic regression were used to estimate the effects of LLM model, verbosity (response length), and question characteristics on the score outcome variable (Supplementary Information 1.25). Implementation details We implemented Medieval at HKU-SZH, a tertiary general hospital affiliated with the University of Hong Kong. Thirteen clinical specialties were involved: accidents and emergency (A.E.), cardiology, cardiac surgery, pediatrics, neurology, nephrology, orthopedics, prenatal medicine, neonatal medicine (NICU), respiratory medicine, pediatric surgery, and rare diseases (Supplementary Fig. S1A). Each specialty included one case designer (average years of medicine practices: 14.8) and an independent evaluation panel of clinical experts (associate consultant or above). To balance statistical power and operational management, on average 24 to 25 questions per question-type and per specialty were designed (Supplementary Table S1A). All questions were manually inspected to avoid unintended information leak. The final cutoff time for question submission was 23:59 on February 28, 2025. Detailed rationales for case selection and question design are provided in Supplementary Information S1.3 to S1.15. The LLM queries were run between March 8 and 9 in 2025. The expert evaluations took place from March 10 to March 24. Ethical approval This study was performed in accordance with the Declaration of Helsinki, and ethical approval was obtained from the Ethics Committee of the University of Hong Kong – Shenzhen Hospital (reference no. Ethics-2025-078). Informed consent was waived. All personal identifiers in questions concerning real patients removed before submitting to models. No restricted data of human genetics was involved in this study. Results A multidisciplinary question bank of real clinical cases The 695 candidate questions proposed by the designer panel were reviewed for adequacy and relevance (Methods) by panel members (Supplementary Table S1). Of note, more real-world cases (31% rated hard) are harder than OSCE ones (9%) (Supplementary Fig. S2A). Some specialties (cardiology, ICU, respiratory medicine, prenatal diagnosis, pediatrics, and cardiac surgery, where over 50% were rated hard) are also more difficult than others. Overall, 18.2% (125/685) of questions pertained to rare diseases. Such cases were concentrated within specific specialties, constituting 100%, 97.7%, and 36.8% in prenatal diagnosis, rare disease category, and neurology, respectively. Relevancy with respect to diagnosis, treatment planning, or surgical operations were also assessed (Supplementary Fig. S2D–F). Ten of the 695 cases were excluded due to inadequacy (incompleteness or information leaks, Methods), leading to a high-quality cross-specialty question bank of 685 cases (48.9% recent clinical cases, 51.1% OSCE). Rigorous blinded evaluation of LLM responses by qualified experts For each items in the question bank, responses from seven LLMs were solicited using a standardized prompt (Methods), generating a total of 4795 LLM responses. A panel of 30 evaluators from 13 specialties then took part in the blinded evaluation process (Methods). One evaluator withdrew and two were removed according to our criteria for quality control (Methods, Supplementary Figs. S2a, S3, and S4). The 27 qualified evaluators averaged 25.9 years of clinical experience (Supplementary Fig. S1A). On average, it took each panel member 61.7 h (or 11.6 min/case) to complete the evaluations (Supplementary Fig. S4B). Panel members reconsidered 288 scores after initial ratings, leading to final scores with 1.2% improvement in evaluator-specific noises (Supplementary Information S2.1). After rigorous data quality control, we obtained 9856 scores on the LLM responses (Supplementary Figs. S3 to S4). Based on these ratings by the panel, the internal consistencies of the clinical case questions were high (Cronbach’s alpha = 0.89, CI: 0.87–0.90). We next assessed the reliability of the ratings using two-way random effects models (Fig. 2A). On average, each clinical expert rated 365 LLM responses. At the single rating level, the intraclass correlation coefficient (ICC) of the absolute agreement among raters was 0.476. Importantly, the average rating of LLM performance within each specialty had an absolute agreement ICC of 0.759, indicating good reliability. Further, the average rating for each LLM had an absolute agreement ICC of 0.995. Therefore, the average performance rating of each LLM is highly reliable.Fig. 2Analysis of rating scores of LLM responses rated by the expert evaluator panel.A Intraclass correlation coefficients for absolute agreement among evaluators in the expert panel at the level of single ratings or mean ratings averaged within each LLM response, specialty, or LLM model. B Comparison of nested linear mixed effects models fitted to the rating scores. Baseline model includes clinical case and evaluator as random effects. A fixed effect variable was included incrementally to build the final regression model. Statistical significance was assessed by the likelihood ratio test. C Adjusted rating scores for each LLM as estimated by a linear mixed-effects model, accounting for inter-case and inter-evaluator variability, as well as verbosity. Error bars indicate 95% confidence intervals (n = 9856) with two-sided t-test. Numerical values of the data in this figure can be found in Supplementary Data 232. Analysis of rating scores of LLM responses rated by the expert evaluator panel. A Intraclass correlation coefficients for absolute agreement among evaluators in the expert panel at the level of single ratings or mean ratings averaged within each LLM response, specialty, or LLM model. B Comparison of nested linear mixed effects models fitted to the rating scores. Baseline model includes clinical case and evaluator as random effects. A fixed effect variable was included incrementally to build the final regression model. Statistical significance was assessed by the likelihood ratio test. C Adjusted rating scores for each LLM as estimated by a linear mixed-effects model, accounting for inter-case and inter-evaluator variability, as well as verbosity. Error bars indicate 95% confidence intervals (n = 9856) with two-sided t-test. Numerical values of the data in this figure can be found in Supplementary Data 232. Gemini 2.0 achieved the highest raw rating scores Upon breaking down into different models, Gemini2F topped the leaderboard with an average raw score of 7.1 (standard error [s.e.] 0.04), followed by Qwen Max (mean 6.9), DS-R1 671B (mean 6.8), ChatGPT-4oL (6.7), Qwen 32B (6.6). DS32B and DS70B were considerably lower, at 5.9 and 5.7, respectively (Supplementary Fig. S5A). Post hoc Tukey analysis revealed multiple significant pairwise differences in rating scores across LLMs, with Gemini, Qwen Max, and DS-R1 671B outperforming ChatGPT-4oL, while DS32B and DS70B showed significantly lower performance than most other models (adjusted \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p < 0.001$$\end{document}p<0.001). Discipline-wise, Gemini was the winner in 10 out of all 13 specialties, followed by Qwen max (winner in 01b.AE and 09.Prenatal) and DS-R1 (winner in 01a.AE and 04.Neural) (Supplementary Fig. S5B). Response scores primarily explained by model strength and verbosity We used linear mixed-effects (LME) models to fit the rating scores. First, we fit a baseline model with case- and evaluator-specific variations as random effects only, and incrementally included fixed-effect variables in more complex models (Fig. 2B). Inclusion of the LLM variable significantly improved the log likelihood (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p < 0.001$$\end{document}p<0.001, likelihood ratio test). Further adding the specialty variable only marginally improved the fit, suggesting minor cross-specialty heterogeneities. We performed a full-model LME regression by further adding verbosity, case difficulty, question type, disease rarity, etc. as covariates (Supplementary Table S2). Response verbosity significantly affected the rating scores, with longer responses receiving higher scores. All other covariates did not seem to have influences on the ratings. Similar results were obtained when a full-model multiple linear regression was fitted (Supplementary Table S3). Taken together, the scores were primarily explained by intrinsic model strengths and verbosity. Advanced LLMs response qualities on par with early-stage physicians Based on the analysis above, an LME model with LLMs and verbosity as fixed-effects and case- and evaluator-specific variations as random effects, excluding interactions and question-specific variations, appears fittest. From this fitted model, we obtained the adjusted coefficient estimates, a reflection of the intrinsic strengths, for individual LLMs (Fig. 2C). LLMs with the highest adjusted scores were DS-R1 671B, Qwen Max, Gemini2F, and ChatGPT-4oL. Conversely, the two smaller models DS32B and DS70B significantly underperformed. Of note, the four advanced LLMs including DeepSeek R1, Qwen Max, Gemini 2F, and ChatGPT-4oL have achieved average rating scores of significantly above 6 (Fig. 2C), which is equivalent to that of a physician with 3-6 years of practices, based on our ten-point Dreyfus-based scale (Supplementary Information S1.3). Predictors of excellence in LLM responses To assess LLM performance in identifying high-quality answers, we define excellence to be score \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\ge 8$$\end{document}≥8 (expert or above, Supplementary Information S1.2), for which 1,862 instances were observed (18.9% out of 9,856). Gemini2F had the highest raw percentage of excellent responses (Table 1). We applied a generalized linear mixed-effects model with evaluator as random effects. The fitted model showed strong explanatory power (marginal \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${R}^{2}=0.48$$\end{document}R2=0.48; conditional \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${R}^{2}=0.66$$\end{document}R2=0.66), with evaluator effects accounting for 35% of variance (Table 2). Compared with the referent LLM (DS32B), DS-R1 671B (odds ratio [OR] 2.88; 95% CI 1.90 to 4.38), Qwen Max (OR 2.48), ChatGPT-4oL (OR 2.44), and Gemini2F (OR 2.45) had significantly higher odds of achieving rating score \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\ge 8$$\end{document}≥8 (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p < 0.01$$\end{document}p<0.01 for all). Increasing verbosity was also associated with higher probability of achieving excellence.Table 1Frequencies of excellence and incompetence in LLM responsesExcellenceIncompetenceModelsOtherwiseExcellent%OtherwiseIncompetent%DS70B1319896.32%113527319.39%DS32B13051037.32%123517312.29%QWEN32b113827019.18%1335735.18%ChatGPT-4oL112428420.17%1352563.98%QWENMAX103537326.49%1355533.76%DS-R1 671B112828019.89%1360483.41%Gemini2F94546332.88%1370382.70%Table 2Coefficient estimates under logistic mixed-effects model for predicting excellent LLM response, adjusted for evaluator random effectsExcellence (rating score \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\ge$$\end{document}≥ 8)Fixed-effectsOdds ratio95% C.I.\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p$$\end{document}p(Intercept)0.000.00, 0.01\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ < $$\end{document}<0.001Models (referent: model=DS32B, verbosity rank=1)model [DS70B]0.800.58, 1.100.173model [Qwen32B]1.901.22, 2.940.004model [ChatGPT-4oL]2.441.61, 3.71\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ < $$\end{document}<0.001model [Gemini2F]2.451.40, 4.270.002model [QwenMAX]2.481.56, 3.93\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ < $$\end{document}<0.001model [DS-R1 671B]2.881.90, 4.38\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ < $$\end{document}<0.001verbosity [rank=2]1.401.02, 1.940.038verbosity [rank=3]1.490.95, 2.320.081verbosity [rank=4]2.111.34, 3.300.001verbosity [rank=5]2.761.74, 4.37\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ < $$\end{document}<0.001verbosity [rank=6]4.152.57, 6.72\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ < $$\end{document}<0.001verbosity [rank=7]6.363.58, 11.30\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ < $$\end{document}<0.001Specialties (referent: AEa)specialty [AEb]0.250.01, 9.100.452specialty [PediSurg]0.300.01, 14.430.544specialty [ICU]1.020.06, 17.060.988specialty [Pedi]9.900.67, 147.150.096specialty [Neural]13.070.82, 209.280.069specialty [Rare]19.201.57, 234.820.021specialty [Ortho]21.891.48, 323.380.025specialty [Nephro]22.641.91, 267.710.013specialty [Cardio]44.673.03, 657.940.006specialty [Resp]50.243.42, 737.530.004specialty [CardiacSurg]59.754.08, 875.900.003specialty [Neonatal]82.355.61, 1209.210.001specialty [Prenatal]352.5024.02, 5173.71\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ < $$\end{document}<0.001Random effects Residual1.81 Evaluator1.33 \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${N}_{\mathrm{Evaluator}}$$\end{document}NEvaluator29 \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${N}_{\mathrm{Observations}}$$\end{document}NObservations9856Model verbosity refers to the text lengths of model responses, in ascending ranks. Table 1Frequencies of excellence and incompetence in LLM responsesExcellenceIncompetenceModelsOtherwiseExcellent%OtherwiseIncompetent%DS70B1319896.32%113527319.39%DS32B13051037.32%123517312.29%QWEN32b113827019.18%1335735.18%ChatGPT-4oL112428420.17%1352563.98%QWENMAX103537326.49%1355533.76%DS-R1 671B112828019.89%1360483.41%Gemini2F94546332.88%1370382.70% Frequencies of excellence and incompetence in LLM responses Table 2Coefficient estimates under logistic mixed-effects model for predicting excellent LLM response, adjusted for evaluator random effectsExcellence (rating score \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\ge$$\end{document}≥ 8)Fixed-effectsOdds ratio95% C.I.\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p$$\end{document}p(Intercept)0.000.00, 0.01\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ < $$\end{document}<0.001Models (referent: model=DS32B, verbosity rank=1)model [DS70B]0.800.58, 1.100.173model [Qwen32B]1.901.22, 2.940.004model [ChatGPT-4oL]2.441.61, 3.71\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ < $$\end{document}<0.001model [Gemini2F]2.451.40, 4.270.002model [QwenMAX]2.481.56, 3.93\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ < $$\end{document}<0.001model [DS-R1 671B]2.881.90, 4.38\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ < $$\end{document}<0.001verbosity [rank=2]1.401.02, 1.940.038verbosity [rank=3]1.490.95, 2.320.081verbosity [rank=4]2.111.34, 3.300.001verbosity [rank=5]2.761.74, 4.37\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ < $$\end{document}<0.001verbosity [rank=6]4.152.57, 6.72\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ < $$\end{document}<0.001verbosity [rank=7]6.363.58, 11.30\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ < $$\end{document}<0.001Specialties (referent: AEa)specialty [AEb]0.250.01, 9.100.452specialty [PediSurg]0.300.01, 14.430.544specialty [ICU]1.020.06, 17.060.988specialty [Pedi]9.900.67, 147.150.096specialty [Neural]13.070.82, 209.280.069specialty [Rare]19.201.57, 234.820.021specialty [Ortho]21.891.48, 323.380.025specialty [Nephro]22.641.91, 267.710.013specialty [Cardio]44.673.03, 657.940.006specialty [Resp]50.243.42, 737.530.004specialty [CardiacSurg]59.754.08, 875.900.003specialty [Neonatal]82.355.61, 1209.210.001specialty [Prenatal]352.5024.02, 5173.71\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ < $$\end{document}<0.001Random effects Residual1.81 Evaluator1.33 \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${N}_{\mathrm{Evaluator}}$$\end{document}NEvaluator29 \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${N}_{\mathrm{Observations}}$$\end{document}NObservations9856Model verbosity refers to the text lengths of model responses, in ascending ranks. Coefficient estimates under logistic mixed-effects model for predicting excellent LLM response, adjusted for evaluator random effects Model verbosity refers to the text lengths of model responses, in ascending ranks. Risks of incompetence in LLM responses We next assessed the probability that LLMs generate sub-par responses, by defining incompetence as having score \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ < 4$$\end{document}<4, the minimum of a beginning practicing physician in our Dreyfus-based scale (Supplementary Information S1.2). In total, 714 instances of incompetence were observed in the LLM responses (7.2%), with Gemini 2.0 generating responses with the lowest frequency of incompetence at 2.7% (Table 1). Interestingly, despite ranking only fourth highest in excellence (19.9%), DS-R1 671B ranked second lowest (3.4%) in incompetence. We fitted a logistic mixed-effects model with specialty and LLM as fixed effects, while adjusting for inter-rater variability. Given the low number of incompetence evens, to achieve numerical stability we employed a Bayesian Generalized Linear Mixed Model30 with weakly informative priors to analyze binary outcomes. We treated LLMs, verbosity, and specialty as fixed effects, and evaluator as random ones. Compared with the referent DS32B, DS-R1 671B (OR 0.43; 95% CI [0.27, 0.68]), ChatGPT-4oL (OR 0.52; 95% CI [0.33, 0.82]) demonstrated significantly lower odds of incompetence while DS70B was more likely to fail (OR: 1.98, 95% CI 1.57-2.49) (Supplementary Tables S4). Across specialties, LLMs are more likely to fail in ICU, neonatology, and pediatric surgery. Hallucination in LLM responses Another risk in deploying AI in clinical applications is hallucination, which is defined as fabricated medical facts31. Forty instances of hallucination were identified by evaluator panels (Supplementary Data 232), including fabricated drugs and dosages (n = 5), conditions (n = 2), subtyping (n = 3), physical examination details in OSCE items (n = 7), misinterpretation of variant pathogenicity (n = 15). These instances span across multiple specialties, including pediatrics, nephrology, neurology, prenatal diagnosis, respiratory medicine, and orthopedics. In prenatal diagnosis, 17 instances of hallucinations were identified. The most common errors involved misinterpretation or incorrect application of variant classification criteria such as PVS1, PS3/4, PM1/2/4, and PP1/3. For example, responses to Q3, Q7 and Q23 by Qwen 32B, Q8 by ChatGPT-4oL, and Q20 by Gemini, misused the ACMG-ACGS criteria33,34 and the ClinGen Sequence Variant Interpretation framework. In Q11, five of the LLMs (except Qwen Max and DeepSeek R1) misinterpreted the function of DHTKD1, a relatively understudied gene with limited and conflicting literature. Additional examples of hallucination include the misattribution of a kidney-specific phenotype (Q9) to mitochondrial disease by Qwen 32B, and a misdiagnosis by DS70B of a case (Q19) as spinocerebellar ataxia and falsely equated with Joubert syndrome. In pediatrics, 12 instances of hallucination (9 in OSCE items and 2 in real cases) were identified, often involving ignorance of the appropriate drug use or dosing for pediatric patients. In Q5, Qwen32B recommended doxycycline at nearly twice the evidence-based dose specified in national guidelines. ChatGPT-4oL incorrectly endorsed levofloxacin use in children under 18, contradicting national guidelines for anti-microbial prescriptions. DeepSeek models (70B and 32B) fabricated family and exposure histories (Q4), while Qwen 32B hallucinated physical exam findings that were not reported in a clinical case (Q9). Instances of hallucination were also reported in diagnostic reasoning. For example, in Q25 of pediatrics, Qwen32B diagnosed NK cell deficiency based solely on a single low count, ignoring contextually normal immunoglobulins and a known history of coconut oil aspiration. In contrast, DeepSeek R1 671B identified the same laboratory finding but interpreted it conservatively. In Q38, ChatGPT-4oL cited incorrect reference ranges for G6PD, leading to a false diagnosis of G6PD deficiency. In Q11 of respiratory medicine, LLMs suggested that osimertinib-related ILD requires drug discontinuation, which oversimplified clinical decision-making and contradicted current clinical guidelines35,36. Furthermore, instances of fabricated literature were reported in orthopedics, and fabricated drugs in nephrology and neurology, where non-existent drug names were suggested by LLMs. Above all, these identified hallucinations in LLM responses are sufficiently egregious to constitute medical malpractice, highlighting the need for human expert oversight. Discussion LLMs hold great promises for medicine but their safe and ethical use is paramount. Computational validations alone are insufficient. Human evaluations by clinical experts are critical guardrails against misuses. Our study presents a comprehensive multi-disciplinary expert evaluation methodological framework (Medieval) for clinical applications of LLM. Medieval has multiple contributions. First, as compared with recent studies37,38 that uses open question sets, Medieval uses real clinical cases not known to the models, which help reduces model memorization39. Second, as compared with crowdsourcing approaches, our study set stringent entry bars for evaluator qualifications to ensure the quality and reliability of the rating scores. Third, instead of multiple-choice testing, our framework uses content-based evaluations, which allows the logical reasoning process and clinical practicality of the LLM responses to be scrutinized. Fourth, we developed a Dreyfus-based scale to link the quality of responses to the career stage of a human physician, assigning meanings to scores understandable to both medical professionals and the general public. Fifth, rigorous statistical analyses were performed to delineate the intrinsic model strengths from possible confounders. Lastly, we implemented our approach on a large question set that spans wide spectrum of specialties to enhance the representativeness of the results, introduced blinding and randomization to reduce confounding, and evaluated the most highly ranked LLMs at the time of our evaluation within a short time window ( ≤ 45 days) to reduces the risk of bias due to time lag. Applying our approach yielded interesting findings. Based on >\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathrm{9,800}$$\end{document}9, 800 ratings, we found that Gemini 2.0 Flash achieved the highest ratings but its verbosity may also have conferred advantages. With LME models adjusting for evaluator variability and verbosity, we found that DeepSeek R1 had the highest rating. Delineating confounders, advanced LLMs including DeepSeek R1, Gemini 2.0 Flash, ChatGPT-4oL, and Qwen Max were found to match to human physicians with 3-6 years of post-qualification experience, while smaller models such as DS70B and DS32B significantly underperformed. In spite of their remarkable average performances, we showed that both advanced and smaller LLMs exhibit pathological behaviors. Our expert panel identified a concerning 7.2% rate of diagnostic incompetence and 40 instances of fabricated medical facts, including erroneous diagnoses, drug recommendations, and genetic interpretations. For example, in a simulated A.E. case with a typical inferior ST-segment elevation myocardial infarction (STEMI) complicated by right ventricular infarction and hypotension (90/50 mmHg), most LLMs offered routine reminders to monitor blood pressure when using nitroglycerin. In contrast, DeepSeek R1 demonstrated a more profound understanding of the pathophysiology, by integrating the patient’s low blood pressure and signs of right heart failure such as jugular venous distension, to infer the right ventricular involvement and explicitly recommend avoiding nitroglycerin uses. This latter recommendation fully aligns with the American College of Cardiology guidelines40. Additionally, our expert panel also reported fabrications by LLM that pertain to variant site interpretation in prenatal diagnosis. This result is consistent prior studies on the LLM applications in clinical genetics16,41. Our results show that while the average performance of some LLMs matches human physicians, their high variability and lack of a safety floor preclude their immediate unsupervised uses. We offer some additional technical notes below. First, our LME analyses treated the scores as real values, though in fact they are 0.5-intervaled from 0 to 10. We performed a robustness check by running a Cumulative Link Mixed Model (CLMM) treating the scores as ordinal values, and confirmed that the relative ratios between two key random effects (questions and evaluators) are highly close in both models (Supplementary Information S2.2), supporting the robustness of the LME analyses. Second, one way to improve the LME fitting for more accurate estimations of intrinsic LLM strengths is to further increase the panel size (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$k\approx 2$$\end{document}k≈2 per specialty in our study). This is equivalent to increasing the intraclass correlation coefficient (ICC) of the absolute agreement among raters, or ICC(A, k), to the excellent level ( > 0.9) by increasing \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$k$$\end{document}k. Assuming the observed inter-evaluator and inter-specialty variabilities are representative of the larger evaluation schemes, we showed that (details in Supplementary Information S1.26), \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$k\ge 20$$\end{document}k≥20 is required. This translates to an ideal panel size 10 times larger than the current one, with each performing similar workload of evaluations as the current study. Although challenging to implement, this helps guide future planning of larger studies. Third, in our progress monitoring, the estimated time used per response (11.6 min) may be somewhat overestimated, because the evaluators have competing obligations during the evaluation period. Overall, we projected that routine expert-based monitoring is realistic. We are aware of some limitations. First, as compared with multiple-choice question-sets with hard ground-truths, scoring open-ended responses by human experts is more subjective. Second, the verbosity of LLM responses is positively associated with rating scores by human evaluators, which widens the confidence intervals for true competence estimation despite advanced LME analyses. This can be optimized by structured prompt templates in future studies. Third, while tested models were released within 45 days, training data currency differences may affect performance independent of reasoning capabilities. Fourth, this work is single-centered and monolingual. Although monolingual studies were common in the literature1, multi-centered and cross-country studies will surely enhance the reproducibility. Lastly, individual models are known to be biased42. Future studies may also involve multiple models or multiple queries as recent studies showed evident improvements with such strategies43. Conclusion This study establishes a rigorous, blinded “human-in-the-loop” framework to evaluate advanced LLMs in medicine using real cases and by real human experts. Our findings reveal that while top-tier models demonstrate sophisticated clinical reasoning capacities, they also exhibit critical variability and safety risks. Further methodological developments and continuous assessments are needed before wider clinical adoptions of LLMs. Supplementary information Supplementary information Supplementary information