Accessibility settings

This paper is in the following e-collection/theme issue:

e-Learning and Digital Medical Education (1509) Digital Health, Telehealth and e-Innovation in Clinical Settings (679) Natural Language Processing (1206)

Published on in Vol 28 (2026)

Preprints (earlier versions) of this paper are available at https://preprints.jmir.org/preprint/92852, first published .
Medical team reviews X-rays, discussing patient diagnosis and treatment options.

Authoritative Textbook-Augmented Large Language Models for High-Altitude Public Health Medical Education in the Xizang Autonomous Region: Cross-Sectional Comparative Evaluation Study

Authoritative Textbook-Augmented Large Language Models for High-Altitude Public Health Medical Education in the Xizang Autonomous Region: Cross-Sectional Comparative Evaluation Study

Authors of this article:

Kun He1, 2 Author Orcid Image ;   Qiming Xiao1, 3 Author Orcid Image ;   Wangyang Chen1 Author Orcid Image ;   Lisha Jing3 Author Orcid Image ;   Yabing Wang4 Author Orcid Image ;   Shuai Li1, 3 Author Orcid Image ;   Daiyu Yang1, 3 Author Orcid Image ;   Hemiao Xu1, 3 Author Orcid Image ;   Ke Pang1, 3 Author Orcid Image ;   Ruoyu Xiao5 Author Orcid Image ;   Zhashilamu Liu5 Author Orcid Image ;   Deji Zhuoga5 Author Orcid Image ;   Ruxuan Chen6 Author Orcid Image ;   Jingyi Li7 Author Orcid Image ;   Long Chang8 Author Orcid Image ;   Yangzhong Zhou9 Author Orcid Image ;   Zhe Zhang10 Author Orcid Image ;   Ran Li11 Author Orcid Image ;   Lujing Ying1, 3 Author Orcid Image ;   Rutong Li1 Author Orcid Image ;   Hongwei Wang1 Author Orcid Image ;   Xin Yin12 Author Orcid Image ;   Ge Zhen2 Author Orcid Image ;   Siyi Cai13 Author Orcid Image ;   Qijun Shan14 Author Orcid Image ;   Qiang Wang1 Author Orcid Image ;   Danzeng Zhuoga15 Author Orcid Image ;   Ciren Yangjin2 Author Orcid Image ;   Gesang Luobu15 Author Orcid Image ;   Tu Ji16 Author Orcid Image ;   Dong Wu1, 2 Author Orcid Image
Article Authors Cited by Tweetations Metrics

    1Department of Gastroenterology, State Key Laboratory of Complex Severe and Rare Diseases, Peking Union Medical College Hospital, Shuaifuyuan, No.1, Dongcheng District, Beijing, Beijing, China

    2Department of Neurology, Peking Union Medical College Hospital, Beijing, Beijing, China

    3Department of Endocrinology, Key laboratory of Endocrinology of National Health Commission, Peking Union Medical College Hospital, Beijing, Beijing, China

    4Department of Gastroenterology, Air Force Medical Center, Chinese People's Liberation Army, Beijing, Beijing, China

    5Department of Orthopedics, Peking Union Medical College Hospital, Beijing, Beijing, China

    6Information Center, Peking Union Medical College Hospital, Beijing, Beijing, China

    7High Altitude Medical Research Institute, Tibet Autonomous Region People's Hospital, Lhasa, Tibet, China

    8Institute of Basic Medical Sciences, Chinese Academy of Medical Sciences & Peking Union Medical College, Beijing, Beijing, China

    9Department of Gastroenterology, Tibet Autonomous Region People's Hospital, Lhasa, Tibet, China

    10Chinese Academy of Medical Sciences & Peking Union Medical College, Beijing, Beijing, China

    11Department of Endocrinology, Beijing Friendship Hospital, Beijing, Beijing, China

    12Medical College, Xizang University, Lhasa, Tibet, China

    13Department of Pulmonary and Critical Care Medicine, Peking Union Medical College Hospital, Beijing, Beijing, China

    14Department of Cardiology, State Key Laboratory of Complex Severe and Rare Diseases, Peking Union Medical College Hospital, Beijing, Beijing, China

    15Department of Hematology, Peking Union Medical College Hospital, Beijing, Beijing, China

    16Department of Rheumatology and Clinical Immunology, National Clinical Research Center for Dermatologic and Immunologic Diseases, Peking Union Medical College Hospital, Beijing, Beijing, China

    Corresponding Author:

    Dong Wu, Prof Dr


    Background: Public health medical education is increasingly important in the low-resource, high-altitude Xizang Autonomous Region (Tibet). Traditional authoritative textbooks do not meet modern needs for accessibility and interactivity, whereas general large language models (LLMs) may hallucinate in specialized medical domains. Developing specialized LLMs for low-resource regions is also expensive and difficult.

    Objective: This study aimed to explore a novel approach to high-altitude public health medical education in the low-resource Xizang Autonomous Region that integrates modern LLMs and authoritative textbooks, using a comprehensive benchmark evaluation across multiple dimensions and retrieval-augmented generation (RAG) technology.

    Methods: We conducted a 2-stage cross-sectional comparative evaluation study to benchmark publicly available LLMs and evaluate the added value of textbook-augmented retrieval under standardized generation settings and blinded expert assessment. First, 4 publicly available LLMs (GPT-5.2 [OpenAI], Gemini 3.0 Pro [Google], DeepSeek R1 [DeepSeek], and Tencent HY 2.0 [Tencent]) were benchmarked using an 80-question benchmark on high-altitude public health medicine developed by authoritative medical specialists. Each question was asked 3 times, yielding 960 outputs; first responses (n=320) were scored under blinded conditions by 2 independent 8-member physician panels. A clinically weighted evaluation of multidimensional first-response scores (including comprehensiveness, accuracy, clarity, and relevance) and a composite consistency metric (including semantic similarity and algorithmic similarity) was administered. Second, 4 specific and prevalent authoritative textbooks on high-altitude public health medicine—Ward, Milledge and West’s High Altitude Medicine and Physiology, High Altitude Medicine: A Case-Based Approach, High Altitude Medicine, and High Altitude Medical Protection—were deployed as the external knowledge base for the evaluation-optimized model. Statistical analyses included Spearman ρ, Cronbach α, intraclass correlation coefficients, Friedman tests with Dunn multiple comparisons, and paired Wilcoxon signed-rank tests. The significance threshold was set at α=.05.

    Results: DeepSeek R1 was selected as the optimal base model for achieving the highest weighted score (5.61/10.00), followed by GPT-5.2 (5.51/10.00), Gemini 3.0 Pro (5.39/10.00), and Tencent HY 2.0 (4.71/10.00). The deployed retrieval-augmented model integrating the authoritative textbooks and the optimal LLM DeepSeek R1, HPHME-Xplus-RAG, achieved remarkable improvement in multidimensional scores compared to baseline DeepSeek R1 (median 8.00 [IQR 7.88‐8.00] vs median 7.63 [IQR 7.38‐7.88]; P<.001, r_rb=0.68, indicating a large effect).

    Conclusions: Integrating authoritative textbooks with an evaluation-optimized general LLM through an RAG framework showed strong performance for medical education in the low-resource Xizang Autonomous Region. Unlike prior studies that mainly evaluated general LLMs or used clinical guidelines to build RAG systems for diagnosis and treatment, this study used authoritative textbooks for the broader, guideline-scarce field of public health medical education. This work provides a replicable workflow—domain-authoritative knowledge+RAG+model optimization and evaluation—for low-resource settings, with practical implications for medical instructors and students, hospitals, and public health services seeking cost-effective, convenient, and trustworthy educational support.

    J Med Internet Res 2026;28:e92852

    doi:10.2196/92852

    Keywords

    public health medical education; large language model; authoritative textbooks; retrieval-augmented generation; high altitude 


    High-altitude public health medical education is essential for protecting individual well-being and regional public health in plateau settings. High-altitude regions, including the Xizang Autonomous Region (Tibet), are culturally and ecologically significant and have attracted increasing attention because plateau exposure poses substantial physical and psychological challenges to both residents and visitors [1-3]. Existing evidence shows that high-altitude environments may affect multiple physiological and behavioral domains, including cardiovascular function, cognitive performance, sleep quality, blood pressure regulation, neurological symptoms, musculoskeletal adaptation, inflammatory and immune responses, and mental health [4-10]. These multisystem effects can impair daily functioning, increase health risks during travel, work, and long-term residence, and create practical demands for early recognition, prevention, and self-management. Therefore, timely public health medical education is needed to raise awareness and promote preventive behaviors among medical instructors and students, frontline health care workers, residents, and travelers.

    However, public medical education in this setting faces substantial structural constraints, particularly in low-resource regions such as the Xizang Autonomous Region. Studies have shown insufficient general practitioner training and limited continuing medical education opportunities [11-13]. Local reports have also described difficulty accessing standardized lectures and textbooks because of limited educational resources, as well as shortages of qualified teachers and the limited relevance of standardized teaching materials to plateau-specific conditions [14,15]. In addition, existing formal guidelines and consensus statements in high-altitude medicine are fragmented and largely focused on acute altitude illness, travel medicine, or selected subgroups, while comprehensive guidance specifically addressing high-altitude public health medicine remains scarce [16-20].

    Although authoritative textbooks have long served as the primary medium for medical education in this field and encapsulate years of accumulated expertise, conventional materials alone increasingly struggle to meet modern expectations for accessibility, efficiency, and interactivity. Large language models (LLMs) have rapidly advanced driven by advances in deep learning and the Transformer architecture, supporting a broad range of medical tasks, including medical education, clinical decision-making, specialty-specific question answering, and diagnostic reasoning, although their performance varies substantially across specialties, tasks, and evaluation settings [21-39]. Despite this growing body of literature, evidence for LLM use in high-altitude medical education remains scarce. Therefore, although LLMs show considerable promise in spreading public health information, especially in settings where traditional outreach is limited, their role in high-altitude public health medical education remains insufficiently studied, with domain-specific accuracy and hallucinations probably being a key limitation [40,41]. Also, developing specialized large models for low-resource domains and regions is prohibitively expensive and difficult. This contrast underscores the need for hybrid approaches that combine the strengths of both paradigms.

    Retrieval-augmented generation (RAG) offers such a bridge. By integrating external, authoritative sources into the generative process, RAG allows an LLM to retrieve verified information before producing responses, providing a cost-effective approach compared to developing a new proprietary LLM for low-resource regions [42-47]. In high-altitude public health medical education, RAG has the potential to merge authoritative content with the flexibility of LLMs to deliver reliable and personalized information. This study aims to evaluate a novel approach that integrates the accuracy of authoritative textbooks with the efficiency and reasoning capacity of an evaluation-optimized LLM using the RAG mechanism for high-altitude public health medical education in this low-resource setting.


    Study Design

    This was a 2-stage cross-sectional comparative evaluation study conducted to select an optimal LLM for the RAG architecture and to evaluate its performance in high-altitude public health medical education. First, the performance of 4 extensively adopted publicly available LLMs in answering questions on high-altitude health medical education was evaluated. Based on the weighted results of the multidimensional first-response scores and composite consistency metrics with the weighting ratio determined by the clinical needs assessment, the LLM with the best performance was selected. Second, a retrieval corpus constructed from the specific and prevalent authoritative textbooks was deployed as an external data source for the selected LLM to develop a RAG architecture for medical education on high-altitude health, and the performance improvement of the RAG architecture compared to the LLMs was verified (Figure 1).

    ‎
    Figure 1. Study design of a 2-stage cross-sectional comparative evaluation study on large language models for high-altitude public health medical education in the Xizang Autonomous Region (Tibet), China. LLM: large language model; RAG: retrieval-augmented generation; HPHME-Xplus-RAG: High-Altitude Public Health Medical Education–Xplus (Xiehe in Chinese for Peking Union Medical College Hospital and Xizang for People’s Hospital of Xizang Autonomous Region)–Retrieval-Augmented Generation.

    Data Collection

    A total of 80 questions and corresponding reference answers were developed through an expert-driven, iterative process. Multidisciplinary specialists from Peking Union Medical College Hospital (PUMCH) and People’s Hospital of Xizang Autonomous Region collaboratively proposed and refined candidate questions on the basis of real-world diagnostic, therapeutic, preventive, and health education scenarios encountered in high-altitude practice. Rather than using random question sampling, we adopted a purposive selection strategy to ensure that the benchmark represented the major and practice-relevant domains of high-altitude public health medicine.

    During question set construction, particular attention was paid to broad domain coverage across the multisystem health effects of hypobaric hypoxia and plateau exposure. The final benchmark therefore included respiratory, cardiovascular, gastrointestinal, neurological, hematological, musculoskeletal, and infectious disease topics, as well as lifestyle, nutrition, and psychological acclimatization issues. In this way, the benchmark was designed to reflect the interdisciplinary nature of high-altitude public health medicine, spanning clinical medicine, physiology, prevention, and behavioral health (Figure 2). The preliminary question set was subsequently reviewed and validated by Xizang-aid physicians from PUMCH with direct experience in high-altitude health adaptation in order to confirm its contextual relevance, content breadth, and practical representativeness for the target setting. Each question was accompanied by an expert-defined reference answer, which served as the “gold standard” with a full score of 10 out of 10. These standard answers were based on authoritative sources, including clinical guidelines, textbooks, and medical specialists’ consensus.

    ‎
    Figure 2. Composition of the 80-question benchmark used in this cross-sectional comparative evaluation study of large language models for high-altitude public health medical education.

    Four LLMs—GPT-5.2 (OpenAI), Gemini 3.0 Pro (Google), DeepSeek R1 (DeepSeek), and Tencent HY 2.0 (Tencent)—were selected a priori because they were publicly available, stably accessible during the study period, and representative of major contemporary LLM families across both international and Chinese developer ecosystems [48-51]. Restricting stage 1 to standalone general-purpose models allowed a controlled comparison of intrinsic model performance before subsequent RAG augmentation. All models were accessed through their respective application programming interfaces under standardized conditions, with temperature set to 0.1 and top_p to 0.9 during generation to ensure result comparability, leaving all other hyperparameters at their developer defaults.

    Each question was posed 3 times to each LLM, generating 3 independent responses per model. Only the first response was scored to simulate real-world, on-demand information retrieval without iterative prompting, while all 3 were used for consistency analysis.

    Quantitative Evaluation of First Responses

    Four LLMs were anonymized and labeled with a code letter (A-D) to ensure blinding during assessment. Sixteen experts participated in the evaluation as 2 independent 8-member panels, one from PUMCH and one from People’s Hospital of Xizang Autonomous Region. All experts were trained and calibrated in advance using the scoring rubric. For each response, the 8 experts within each panel discussed the output and generated one panel-level consensus score using integer ratings from 0 to 10 across 4 predefined dimensions: comprehensiveness, scientific accuracy, relevance, and clarity. Thus, each response yielded 2 consensus scores, one from each panel. The final score for each LLM’s response to a given question was calculated as the average of the 4 dimension scores within each panel; the 2 panel-level consensus scores were then used for interpanel reliability analyses, and their mean was used for subsequent comparative analyses (Figure 3).

    ‎
    Figure 3. Physician scoring rubric used in the comparative evaluation of model-generated answers for high-altitude public health medical education.

    Consistency Analysis

    To assess the intramodel stability of the LLMs’ outputs, 3 responses generated for each question were compared using both lexical (word frequency–based) and structural similarity measures. Specifically, for each question, the 3 independently generated responses from the same model were compared in a pairwise manner (response 1 vs response 2, response 1 vs response 3, and response 2 vs response 3), and the resulting similarity scores were averaged across the 3 pairwise comparisons.

    For lexical similarity, text content similarity was calculated using term frequency–inverse document frequency vectorization followed by cosine similarity, implemented in Python using third-party libraries including sklearn and numpy. For structural similarity, Python’s built-in difflib.SequenceMatcher was used to quantify structural overlap between texts. In this structural similarity analysis, punctuation marks, spaces, and special characters were ignored so that superficial formatting differences would not disproportionately affect the similarity results. Both lexical similarity and structural similarity ranged from 0 to 1.

    However, purely algorithmic similarity metrics have inherent limitations when applied to complex, domain-specific texts such as medical and public health communication. Lexical similarity measures primarily focus on word overlap and frequency, which may overlook paraphrasing, synonym usage, and deeper contextual meaning. Two semantically identical sentences with different wording can thus receive low lexical similarity scores. Similarly, structural similarity metrics, which assess syntactic patterns or sentence structures, may fail to capture differences in the underlying semantic content or clinical implications of responses. These methods cannot reliably detect whether subtle but critical nuances—such as risk factors or treatment recommendations—are consistently conveyed.

    Due to these shortcomings, relying solely on algorithmic similarity can misrepresent true consistency and overlook clinically relevant differences. To overcome this, the final consistency evaluation incorporated both quantitative and physician-assessed semantic similarity, which involves expert reviewers evaluating how closely the meaning and clinical relevance of different model-generated answers align, which is fine-tuned with reference to the rubric in SemEval-2017 Task 1 [52]. This expert judgment better captures the practical, real-world equivalence of responses, reflecting nuances that automated metrics miss. Semantic similarity was scored on a 0‐5 scale and subsequently normalized to a 0‐1 scale for integration with the algorithmic metrics. The final composite score was calculated with the following weighting: physician-assessed semantic similarity (60% weight, detailed in Figure 3), structural similarity (20% weight), and lexical similarity (20% weight). This composite score was used to better reflect the meaningful consistency of model responses in clinical and public health communication contexts.

    Development and Evaluation of the New RAG Architecture

    Before model output generation, physician scoring, and comparative analysis, the expert panels prospectively agreed to use a clinically weighted composite approach to determine the optimal LLM. Because the first-response score was reported on a 0‐10 scale, whereas the composite consistency metric was presented on a normalized 0‐1 scale, the composite consistency metric was linearly rescaled to a 0‐10 scale before model selection. The final model-selection score was then derived by weighting the first-response score at 60% and the rescaled composite consistency score at 40%. This weighting was selected during study design to reflect the intended real-world use case, in which first-response quality was considered slightly more important than response consistency.

    Four widely used authoritative books on high-altitude public health medicine, Ward, Milledge and West’s High Altitude Medicine and Physiology, High Altitude Medicine: A Case-Based Approach, High Altitude Medicine, and High Altitude Medical Protection, were selected as the external knowledge source for the LLM [53-56]. These authoritative books are essential reading material for individuals entering high-altitude regions and serve as a fundamental reference for medical professionals working in high-altitude medicine in China and around the world. The textbooks were obtained in PDF format from official online sources accessible to the research team. The files were used only in a closed internal research environment to construct and test the retrieval corpus, and no copyrighted content was redistributed or released publicly.

    The RAG pipeline was constructed using the LangChain framework, with the text-embedding-ada-002 model used to generate text embeddings. Outputting 1536-dimensional vectors with an 8192-token context window, this model effectively captures the semantic information of medical texts. ChromaDB was used as the vector storage backend. Following chunking, all document segments were transformed into vectors via the embedding model and stored in a local ChromaDB instance to facilitate efficient similarity retrieval. Authoritative textbooks were segmented by chapters and paragraphs using a Recursive Character Text Splitter, with a chunk size of 512 tokens and an overlap of 128 tokens. An initial retrieval of 5 fragments was subsequently reranked, from which the top 3 most relevant segments were selected and fed into the LLM to further mitigate noise. To ensure focused and comparable outputs, we used the identical question set, prompt templates, and hyperparameter settings from the first evaluation phase, configuring only a temperature of 0.1 and a top_p of 0.9, with all other hyperparameters kept at developer defaults.

    To reduce expectancy bias in the second-stage evaluation, the assessment of HPHME-Xplus-RAG, which stands for High-Altitude Public Health Medical Education–Xplus (Xiehe in Chinese for Peking Union Medical College Hospital and Xizang for People’s Hospital of Xizang Autonomous Region)–Retrieval-Augmented Generation, was conducted after a washout period of 2 weeks following the initial 4-model evaluation and was coordinated by a study member who was not involved in model development. During this reassessment, the physician panels were informed only that an additional model output set would be evaluated using the same prespecified scoring rubric and consensus procedure as in stage 1. The identity of the system was not disclosed at the time of scoring. Explicit model identifiers and other source-revealing labels were removed from the evaluation materials before assessment. This procedure provided temporal separation and source masking, although it did not constitute an identical fully blinded parallel comparison to the A-D anonymized first-stage assessment.

    Following a similar comparative approach, average scores and scores in 4 dimensions of the RAG architecture and selected LLM were compared. To further validate the performance improvement of the RAG architecture over the selected LLM, an additional analysis was conducted to investigate the performance disparity between them across different text-generation scenarios. The 80 questions were partitioned into 2 cohorts based on text generation task types: factual and reasoning questions. Factual questions were defined as those primarily requiring recall or retrieval of explicit information from authoritative sources, whereas reasoning questions were defined as those requiring integration, inference, or application of knowledge to generate the answer (the group label for each question is provided in Multimedia Appendix 1). A subsequent comparative evaluation was performed within each cohort.

    Statistical Analysis

    Descriptive statistics, including measures of central tendency (mean and median) and variability (minimum, maximum, SD, SEM, and coefficient of variation), were calculated for each model’s performance. Interpanel reliability was evaluated on the 2 panel-level consensus scores using Spearman ρ, Cronbach α, and the intraclass correlation coefficient (ICC). ICC single and ICC average were calculated using a 2-way mixed-effects model to assess the agreement of a single panel-level consensus score and the reliability of the mean of the 2 panel-level consensus scores, respectively. To test for statistically significant differences between models, the Friedman test followed by the Dunn multiple comparisons test and the paired Wilcoxon signed-rank test were used. In addition to P values, we reported effect sizes to quantify the magnitude of differences. For Friedman tests, Kendall W was calculated as a measure of concordance and effect size (Friedman statistic/n(k–1)). For paired Wilcoxon signed-rank tests between HPHME-Xplus-RAG and DeepSeek R1, matched-pairs rank-biserial correlation (r_rb, (R+ – R-) / (R++R-)) was reported. Unless otherwise indicated, nonnormally distributed continuous variables are presented as median (IQR). All analyses were conducted using IBM SPSS (version 29.0), with exact significance levels determined via Monte Carlo simulation. A 2-sided α level of .05 was used throughout.

    Ethical Considerations

    This study was reviewed and approved by the Ethics Committee of PUMCH (approval number I-25PJ1246) and the Ethics Committee of the People’s Hospital of the Xizang Autonomous Region (approval number ME-TBHP-25-KJ-124). Although the primary study materials were LLM-generated responses rather than patient data, the study involved 16 medical specialists as structured evaluators, and therefore formal human subjects research ethics review and approval were obtained. No patients, patient records, protected health information, or identifiable public user data were involved. Before participation, all medical specialists were informed of the study purpose, evaluation procedures, voluntary nature of participation, and intended use of the evaluation results, and all agreed to participate. The approved ethics protocol permitted the use of their evaluation scores for this study without additional consent. To protect privacy and confidentiality, individual evaluator identities were not linked to scoring results, and only aggregated panel-level scores and summary statistics are reported. No monetary or nonmonetary compensation was provided. The paper and supplementary materials contain no identifiable images, screenshots, names, account identifiers, or other information that could identify individual participants or users.


    Overview and Reliability Verification of the Evaluation Between the Two Evaluation Panels

    We queried each of the 4 models independently 3 times with the set of 80 questions on medical education on high-altitude health, and their outputs were collated in Multimedia Appendix 2.

    In a simulation of real-world conditions, the models’ first responses underwent a blind assessment by 2 independent evaluation panels. Both panels scored the answers of DeepSeek R1 as the best, followed by the answers of Gemini 3.0 Pro, GPT-5.2, and Tencent HY 2.0 (Table 1). Notably, the minimum score of 1.00 assigned to Tencent HY 2.0 by both evaluation panels corresponded to Question 34. On manual review of the raw output, this response was interpreted as a genuine model failure due to informational and clinical insignificance.

    Table 1. Descriptive statistics for the panel-level consensus scores assigned to the first responses generated by 4 publicly available large language models in a 2-stage cross-sectional comparative evaluation study on high-altitude public health medical education.
    First responses scoringGPT-5.2Gemini 3.0 ProDeepSeek R1Tencent HY 2.0
    Evaluation panel 1Evaluation panel 2Evaluation panel 1Evaluation panel 2Evaluation panel 1Evaluation panel 2Evaluation panel 1Evaluation panel 2
    Minimum4.755.256.006.506.756.751.001.00
    Median (IQR)7.00 (0.3125)7.00 (0.5)7.25 (0.75)7.50 (0.75)7.50 (0.5)7.75 (0.5)6.25 (1.0625)6.50 (0.5)
    Maximum7.757.758.008.008.008.007.257.25
    Mean (SD)6.88 (0.42)6.94 (0.38)7.33 (0.43)7.37 (0.41)7.53 (0.36)7.63 (0.35)6.06 (0.90)6.36 (0.82)
    Coefficient of variance (%)6.145.445.815.534.804.5414.8312.92

    To ensure that the 2 evaluation panels adhered to the same criteria, we assessed interpanel reliability based on the 2 panel-level consensus scores for the answers provided by the 4 LLMs (Table 2). Spearman ρ and corresponding P values revealed strong and statistically significant correlations between their scores, suggesting that the first responses of the 4 LLMs were evaluated in the same way. Similarly, Cronbach α, ICC single, and ICC average suggested high reliability. All Cronbach α values were >0.7, suggesting strong internal consistency reliability of the 80 questions on medical education on high-altitude health. ICC single and ICC average values were also calculated using a 2-way mixed-effects model to assess the agreement between the two evaluation panels and the reliability of using the average scores provided by the 2 evaluation panels. All ICCs were statistically significant.

    Table 2. Interpanel reliability for the panel-level consensus scores assigned to the first responses generated by 4 publicly available large language models in a cross-sectional comparative evaluation study on high-altitude public health medical education.
    First responses scoringSpearman ρP valueCronbach αICCa singleP valueICC averageP value
    GPT-5.20.789<.0010.9060.822<.0010.902<.001
    Gemini 3.0 Pro0.678<.0010.7940.658<.0010.794<.001
    DeepSeek R10.764<.0010.8490.712<.0010.832<.001
    Tencent HY 2.00.778<.0010.9210.805<.0010.892<.001

    aICC: intraclass correlation coefficient.

    Furthermore, we computed similar descriptive statistics and rigorously assessed interrater reliability for each of the 4 scoring dimensions, including comprehensiveness, scientific accuracy, clarity, and relevance (Tables S3-S10 in Multimedia Appendix 3). Furthermore, the reliability between the two evaluation panels in each dimension was also sufficient.

    Comparison Among the First Responses of 4 LLMs

    An average score was computed across the 2 evaluation panels for each question and LLM. The answers of DeepSeek R1 were scored as the best, followed by the answers of Gemini 3.0 Pro, GPT-5.2, and Tencent HY 2.0 (Table 3).

    Table 3. Descriptive statistics for the mean first-response scores averaged across 2 independent physician-panel consensus ratings for 4 publicly available large language models evaluated on an 80-question benchmark in high-altitude public health medical education.
    First responses scoringGPT-5.2Gemini 3.0 ProDeepSeek R1Tencent HY 2.0
    Minimum5.006.387.001.00
    Median (IQR)7.00 (0.375)7.38 (0.53125)7.63 (0.5)6.34 (0.875)
    Maximum7.758.008.007.25
    Mean (SD)6.91 (0.38)7.35 (0.38)7.58 (0.33)6.21 (0.83)
    Coefficient of variance (%)5.545.174.3513.34

    Owing to the nonnormal distribution of the rating data, we conducted comparison among the average scores and individual dimension scores of first responses of 4 LLMs by using the Friedman test followed by the Dunn multiple comparisons test (Figure 4). The average scores of the 4 LLMs exhibited strong variation (P<.001), with a Kendall W of 0.75, indicating a large overall effect. DeepSeek R1 outperformed the other 3 LLMs (P=.01 vs Gemini 3.0 Pro, P<.001 vs GPT-5.2 and Tencent HY 2.0), Gemini 3.0 Pro outperformed GPT-5.2st and Tencent HY 2.0 (P<.001 vs both), and GPT-5.2 outperformed Tencent HY 2.0 (P<.001; Figure 4).

    Similarly, the Friedman test revealed significant differences among the average scores of the 4 LLMs in the 4 dimensions (P<.001 for all), with Kendall W values of 0.66 for comprehensiveness, 0.99 for scientific accuracy, 0.54 for relevance, and 0.62 for clarity; furthermore, DeepSeek R1 outperformed GPT-5.2 and Tencent HY 2.0 in all the 4 dimensions (P<.001 for all), but only outperformed Gemini 3.0 Pro in relevance (P=.02) and no significant difference versus Gemini 3.0 Pro in the other 3 dimensions. Gemini 3.0 Pro outperformed GPT-5.2 and Tencent HY 2.0 in all 4 dimensions (P<.001 for all except P=.006 vs GPT-5.2 in relevance and P=.004 vs GPT-5.2 in clarity). GPT-5.2 outperformed Tencent HY 2.0 in all 4 dimensions (P<.001 for all except P=.002 in relevance; Figure 4). Performance rankings of the average scores held consistent for both the overall score and each of the 4 individual dimensions, with DeepSeek R1 being the best, followed by Gemini 3.0 Pro, GPT-5.2, and Tencent HY 2.0 (Figure 4).

    ‎
    Figure 4. Comparison of first-response performance across 4 publicly available large language models. (A) Overall average scores across the 4 evaluation dimensions. Box plots show the median (center line), IQR (box), and range (whiskers); n=80 questions per model. (B) Mean scores for comprehensiveness, scientific accuracy, relevance, and clarity (bars indicate mean, SD; n=80 per model). (C) Radar plot summarizing overall and dimension-specific performance. *P<.05; **P<.01; ***P<.001; ns, not significant.

    Comparison of the Consistency of 3 Responses of 4 LLMs

    To assess output stability, we measured textual consistency among 3 responses per LLM using a composite consistency index incorporating both subjective semantic and objective algorithmic consistency metrics. Similarly, we adopted the Friedman test followed by the Dunn multiple comparisons test to evaluate disparities in the composite consistency scores. In order to dissect its contributions, the test was also conducted on its 2 subcomponents: the subjective and the objective consistency scores (Figure 5). For the composite consistency, the Friedman test was significant (P<.001; Kendall W=0.43, indicating a moderate overall effect). GPT-5.2 achieved the highest consistency, with 0.350 (0.303‐0.376) versus 0.242 (0.227‐0.264) for Gemini 3.0 Pro, 0.275 (0.238‐0.290) for DeepSeek R1, and 0.243 (0.224‐0.281) for Tencent HY 2.0. GPT-5.2 outperformed the other 3 LLMs (P<.001 for all). DeepSeek R1 outperformed Tencent HY 2.0 and Gemini 3.0 Pro (P=.03 vs Tencent HY 2.0 and P=.004 vs Gemini 3.0 Pro). No significant difference was detected between the consistency scores of Tencent HY 2.0 and Gemini 3.0 Pro (Figure 5).

    Similarly, Friedman tests revealed strong differences among the subjective and objective consistency scores of the 4 LLMs (P<.001 for both, with Kendall W values of 0.24 for subjective consistency scores, 0.57 for objective consistency scores); in comparison of the subjective consistency scores, DeepSeek R1 outperformed the other 3 LLMs (P<.001 for all). Tencent HY 2.0 outperformed GPT-5.2 (P=.04) and no statistical differences were detected in Gemini 3.0 Pro versus Tencent HY 2.0 and Gemini 3.0 Pro versus GPT-5.2 (Figure 5). As for the comparison of the objective consistency scores, GPT-5.2 outperformed the other 3 LLMs (P<.001 for all); however, no statistical differences were detected between the objective consistency scores of the other LLMs (Figure 5).

    ‎
    Figure 5. Comparison of response consistency across 4 publicly available large language models. (A) Composite consistency scores across 4 LLMs. Box plots show the median (center line), IQR (box), and range (whiskers); n=80 questions per model. (B) Subjective semantic consistency scores (mean, SD). (C) Objective algorithmic consistency scores (mean, SD). *P<.05; **P<.01; ***P<.001; ns, not significant.

    Comparison Between the Answers of DeepSeek R1 and HPHME-Xplus-RAG

    The top-performing model was selected based on a composite model-selection metric, which combined the first-response score and the composite consistency metric in a 6:4 weighting ratio after the consistency metric had been rescaled from the normalized 0‐1 scale to the 0‐10 scale used for first-response scoring. DeepSeek R1 was finally selected as the best-performing LLM among the 4 LLMs (5.61 vs 5.51 for GPT-5.2, 5.39 for Gemini 3.0 Pro, and 4.71 for Tencent HY 2.0).

    With the 4 authoritative books being ingested into a local retrieval corpus for DeepSeek R1, HPHME-Xplus-RAG, a RAG model to support high-altitude public health medical education, was jointly deployed by researchers from PUMCH and People’s Hospital of Xizang Autonomous Region (hereafter referred to as Xplus) [57].

    After a 2-week washout period, we additionally queried HPHME-Xplus-RAG with the same 80-question set, and these outputs underwent a source-masked reassessment coordinated by an independent study member; the answers were collated in Multimedia Appendix 2. We computed descriptive statistics and rigorously assessed interrater reliability of the average scores of the answers provided by HPHME-Xplus-RAG and the scores in the 4 dimensions, including comprehensiveness, scientific accuracy, clarity, and relevance (Tables S1-S10 in Multimedia Appendix 3). Also, the reliability between the two evaluation panels was sufficient.

    We used the paired Wilcoxon signed-rank test to compare the overall ratings between HPHME-Xplus-RAG and DeepSeek R1, which were nonnormally distributed. In addition, the same test was applied to compare ratings within each of the 4 scoring dimensions and the 2 question categories (Figure 6). HPHME-Xplus-RAG achieved higher overall first-response scores (median 8.00 [IQR 7.88‐8.00]) over DeepSeek R1 (median 7.63 [IQR 7.38‐7.88]). The paired Wilcoxon signed-rank test showed that this difference was statistically significant (P<.001), with a matched-pairs rank-biserial correlation of r_rb=0.68, indicating a large effect (Figure 6). HPHME-Xplus-RAG outperformed DeepSeek R1 in comprehensiveness (P<.001; r_rb=0.60), scientific accuracy (P<.001; r_rb=0.62), relevance (P<.001; r_rb=0.94), and clarity (P<.001; r_rb=0.77), indicating a large effect across all 4 dimensions (Figure 6). The average scores and scores across the 4 dimensions of the answers of HPHME-Xplus-RAG were all higher than those of DeepSeek R1, and the whole improvement was mainly due to the enhancement in clarity and relevance (Figure 6).

    To further assess the new model’s performance enhancement in different task scenarios, we divided the question set into 2 groups, factual questions and reasoning questions, and administered paired Wilcoxon signed-rank tests within each group for comparison. Paired Wilcoxon signed-rank tests revealed that HPHME-Xplus-RAG (median 7.88 [IQR 7.75‐8.00] in factual questions and median 8.00 [IQR 7.88‐8.31] in reasoning questions) outperformed DeepSeek R1 (median 7.63 [IQR 7.50‐7.88] in factual questions and median 7.50 [IQR 7.25‐8.00] in reasoning questions in both 2 kinds of questions; P<.001 for both question groups; r_rb=0.76 for factual questions and r_rb=0.82 for reasoning questions, both indicating a large effect). And HPHME-Xplus-RAG exhibited a larger performance gain over DeepSeek R1 in reasoning questions compared with factual questions (Figure 6).

    ‎
    Figure 6. Comparison between the baseline model DeepSeek R1 and HPHME-Xplus-RAG. (A) Overall average first-response scores (box plots show the median, IQR, and range; n=80 questions per model). (B) Mean scores in comprehensiveness, scientific accuracy, relevance, and clarity (bars indicate mean, SD; n=80 per model). (C) Radar plot summarizing overall and dimension-specific performance. (D) Subgroup comparison according to question type, including factual questions (n=51) and reasoning questions (n=29). *P<.05; **P<.01; ***P<.001; HPHME-Xplus-RAG: High-Altitude Public Health Medical Education–Xplus (Xiehe in Chinese for Peking Union Medical College Hospital and Xizang for People’s Hospital of Xizang Autonomous Region)–Retrieval-Augmented Generation; ns, not significant.

    Principal Findings

    High-altitude public health medical education is of critical importance in low-resource regions, yet it faces substantial challenges [1-12,14-20]. This study systematically evaluated 4 LLMs (DeepSeek R1, GPT-5.2, Gemini 3.0 Pro, and Tencent HY 2.0) on a set of 80 expert-designed questions covering high-altitude public health medical education. Among the 4 models, DeepSeek R1 achieved the highest weighted composite score, integrating multidimensional first-response performance with composite consistency. Building upon this, we developed HPHME-Xplus-RAG, a RAG-based system integrating 4 authoritative textbooks with DeepSeek R1 as the base model. HPHME-Xplus-RAG demonstrated significant and consistent improvement in overall multidimensional performance compared to baseline DeepSeek R1, confirming that integrating domain-authoritative knowledge via RAG can meaningfully enhance LLM performance in a resource-limited, highly specialized medical domain where specialized model development is neither practical nor cost-effective [30,43,58].

    Comparison to Relevant Literature, Interpretations, and Implications

    This study systematically unveils and examines the generative capabilities of LLMs within the domain of high-altitude medicine. Similar to findings reported in other domains, the results of this study strongly support that general LLMs possess the fundamental capability to perform medical education and other tasks; however, their performance remains inconsistent and the overall quality requires further improvement [31-33]. Studies evaluating LLMs in other medical disciplines have also demonstrated the strong performance of the DeepSeek R1 model, while some reports have noted its relatively poorer results in specific specialty tasks [31,34-36]. Differences in model selection, task design, and evaluation results across studies suggest that the capabilities of general LLMs vary across medical specialties and task scenarios. For instance, Tordjman et al [31] demonstrated that DeepSeek R1 achieved an accuracy of 0.92 on the United States Medical Licensing Examination and showed particularly strong diagnostic reasoning performance, though it underperformed relative to ChatGPT-o1 in radiology report summarization. Similarly, a clinical decision-making benchmark involving 125 patient cases found that DeepSeek R1 performed comparably to or better than GPT-4o across diagnosis and treatment recommendation tasks [35].

    Consistent with the majority of previous studies, the application of RAG technology led to a significant improvement in model performance within the target domain [59-64]. It is worth noting that this study innovatively used authoritative textbooks as the external knowledge source, providing a new perspective on the selection of authoritative domain knowledge for RAG-based systems and ensuring that authoritative books, which embody knowledge accumulated over generations, can be preserved in a way that keeps pace with the digital era. This approach is particularly applicable to domains lacking formal clinical guidelines, where authoritative textbooks serve as the most reliable and comprehensive knowledge source available.

    Notably, RAG significantly improved performance across all 4 evaluated dimensions (Figure 6), indicating that the integration of authoritative textbooks enhanced the overall quality of model responses. However, this improvement was not uniformly distributed across dimensions. The overall gain appeared to be driven primarily by larger improvements in clarity and relevance, although significant benefits were also observed in comprehensiveness and scientific accuracy (Figure 6). This pattern suggests that retrieval from authoritative textbooks may be particularly effective in helping the model generate responses that are better focused, more interpretable, and more closely aligned with the educational intent of the questions, while still strengthening factual grounding. In addition, the larger gain observed for reasoning questions than for factual questions suggests that the retrieved domain knowledge may support not only information recall but also the integration and application of knowledge in more complex response-generation scenarios. Mechanistically, this is consistent with the nature of the external knowledge base used in our RAG pipeline: authoritative textbooks provide systematic, well-structured, and context-rich information that can guide the model toward more coherent and educationally relevant outputs. While the current retrieval pipeline performed well, future work could further explore strategies such as query expansion or multidocument summarization to optimize information extraction for highly complex queries.

    Regarding the magnitude of improvement, the observed gain of approximately 0.37 points on the 10-point scale (~4% relative improvement) is modest compared to the larger absolute gains reported in studies using binary accuracy metrics, although the effect is large enough across all dimensions. For example, Ferber et al [60] reported an improvement from 57% to 84% correct responses with RAG applied to oncology guidelines, and Alexandrou et al [64] reported accuracy improvements of approximately 16‐21 percentage points for DeepSeek R1 and ChatGPT-4o on acute coronary syndrome guidelines. However, those studies used binary accuracy outcomes on a different measurement scale, making direct comparison of absolute effect size inappropriate. In addition, the relatively high baseline performance of DeepSeek R1 in our study (median 7.63/10, IQR 0.5) inherently limited the room for further observable improvement. As performance approaches the upper end of the scale, achieving a breakthrough-level absolute increase becomes progressively more difficult, reflecting a typical ceiling effect. Therefore, although the absolute score change was not dramatic, the consistently significant improvements across all evaluated dimensions, together with the large effect sizes, indicate that the benefit of RAG in our setting was both robust and clinically meaningful.

    The study goes beyond conventional accuracy assessments by establishing a 4-dimensional evaluation framework encompassing comprehensiveness, scientific accuracy, relevance, and clarity [65]. Although only the models’ first responses were scored to simulate real-world application scenarios, the study innovatively incorporated a consistency evaluation. Rather than relying solely on algorithmic similarity, we proposed a composite consistency metric that integrates quantitative algorithmic indicators with qualitative expert semantic assessments. Furthermore, based on clinical needs, the study adopted a weighted standard that combines first-response scores with composite consistency, providing a more comprehensive and realistic reflection of the models’ reliability and stability in actual clinical settings [65].

    In addition, this study applies cutting-edge generative artificial intelligence (AI) technology to a specific, critical, and resource-limited domain—high-altitude medicine—providing a replicable paradigm for AI applications in specialized medical fields. It demonstrates how advanced medical expertise, such as authoritative textbooks, can be rapidly transformed into accessible educational resources through technological innovation. By linking the generation process to authoritative high-altitude medicine literature, the system enhances interpretability and credibility, marking a key step toward the practical integration of AI into clinical and educational workflows.

    Limitations

    Several limitations of this study warrant acknowledgment. First, this study focused solely on text generation performance within the context of high-altitude public health medical education; the models’ capabilities across other modalities and task types were not assessed. Second, the entire evaluation was conducted by medical experts, while the intended end users, such as medical instructors and students, frontline health care workers, residents, and travelers, were not directly involved in evaluating system outputs, which may limit the ecological validity of the findings. In addition, the single-system reassessment design cannot fully exclude expectancy effects, even with source masking and temporal separation. Future studies should incorporate real-world user evaluations to complement expert assessments with more rigorous blinded methodology.

    Another limitation relates to the benchmark and generalizability of the findings. Although the 80-question benchmark was purposively designed to cover the major domains of high-altitude public health medicine, the selected question set may still not fully capture the breadth and heterogeneity of real-world problems in this field, with an average of approximately 11 questions per domain. Future studies should therefore expand the benchmark to include a broader and more diverse range of clinical, preventive, educational, and public health scenarios, so as to strengthen domain-specific evaluation and real-world generalizability. Moreover, because the benchmark was developed and validated within the specific geographic, cultural, and practice context of the Xizang Autonomous Region, caution is warranted when extrapolating these findings to other low-resource regions, health systems, or medical specialties. Future studies should therefore conduct cross-regional and cross-domain external validation.

    Several practical and methodological issues also remain to be addressed. The study does not assess practical deployment considerations such as response latency, cost per query, or system reliability under real-world conditions. Moreover, as LLM application programming interfaces are subject to continuous updates, longitudinal comparisons may be affected by model versioning, complicating reproducibility over time. Finally, the observed score distribution, with maximum values concentrated in the 8‐9 range rather than at the full score of 10, raises the possibility of a ceiling effect. Although the current evaluation framework was sufficient for the aims of this study—being developed through expert discussion, built upon established approaches in the literature, and implemented using a 10-point scale with 0.25-point intervals that provided reasonably fine discrimination—the compression of scores near the upper bound may still limit its sensitivity for distinguishing among very high-performing models. As stronger LLMs continue to emerge, such clustering at the top of the scale may make further absolute improvements harder to detect, even when meaningful qualitative differences exist. Nevertheless, despite using a clinically oriented quantitative framework that combined first-response scoring with composite consistency weighting, the inherent probabilistic nature of LLM outputs cannot be eliminated [43,58]. Future research may therefore benefit from adopting more sensitive rating instruments, such as scales with a wider range, to improve discriminative precision when evaluating increasingly capable systems.

    Conclusions

    In summary, this study presents a replicable workflow integrating domain-authoritative knowledge, RAG, and large-model optimization and evaluation for high-altitude public health medical education in the low-resource Xizang Autonomous Region. By integrating authoritative textbooks with an evaluation-optimized general LLM through a RAG framework, the study demonstrated strong performance in a highly specialized educational domain with limited resources. Unlike prior RAG studies that mainly relied on clinical guidelines or focused on diagnosis and treatment tasks, this work used authoritative textbooks for the broader, guideline-scarce field of public health medical education, thereby transforming textbook knowledge into a reliable, specialized, and scalable external knowledge source for LLMs. This approach is particularly suited to highly specialized domains such as high-altitude medicine, where formal clinical guidelines remain limited, and offers a practical way to preserve and deliver authoritative knowledge in the digital era. More broadly, the paradigm of domain-authoritative knowledge+RAG+model optimization and evaluation provides a scalable framework for building trustworthy AI tools in other resource-limited medical and professional settings. With further real-world validation, systems such as HPHME-Xplus-RAG may support clinical and educational practice by providing cost-effective, convenient, and reliable access to medical knowledge.

    Acknowledgments

    JT is co-corresponding author (email: ji.tu@ibms.pumc.edu.cn; phone: 86 18519395290). The authors would like to thank Tao Wang for his invaluable advice in the preparation of this manuscript.

    The authors confirm that no generative artificial intelligence was used in the preparation of this manuscript.

    Funding

    This research was supported by Beijing Research Ward Excellence Program (BRWEP2024W034010103); National Key Research and Development Program of China (2024YFA0918504); Science and Technology Projects of Xizang Autonomous Region, China (XZ202501JD0021, XZ202501ZY0072); Non-profit Central Research Institute Fund of Chinese Academy of Medical Sciences (2024-RW320-01); National High Level Hospital Clinical Research Funding (2025-PUMCH-C-048); Medical Artificial Intelligence Educational Reform Project of Institute of Basic Medical Sciences (2022jcjg0104); The Ministry of Education 2022 Baidu industry-academy cooperation Collaborative Education Project - Medical artificial intelligence curriculum construction (182215PC08768); The 2025 Medical Education Scholars Program at Peking Union Medical College (2025mesp010); 2024 CCF-Baidu Open Fund Project: Construction and Application Evaluation of Benchmark Dataset for Multimodal Collaborative Medical Large Language Model (CCF-BAIDUOF202418).

    Data Availability

    The shareable materials generated and analyzed during this study are provided in the multimedia appendices. Multimedia Appendix 1 contains the 80-question benchmark with labels and corresponding reference answers; Multimedia Appendix 2 contains the model-generated outputs; and Multimedia Appendix 3 contains the supplementary statistical tables, including dimension-specific descriptive statistics and reliability analyses. Additional noncopyrighted materials, such as deidentified scoring tables used for analysis and brief statistical analysis and figure-generation code, are available from the corresponding author upon reasonable request.

    The copyrighted textbook contents used to construct the retrieval-augmented generation knowledge base cannot be publicly shared or redistributed. These materials were used only in a closed internal research environment for retrieval-corpus construction.

    Authors' Contributions

    DW, TJ, and KH conceived the study. KH, QX, WC, and LJ designed the study. KH, YW, RC, JL, LC, YZ, ZZ, Ran Li, LY, Rutong Li, HW, XY, GZ, SC, QS, QW, DZ, CY, and GL designed the questions and performed the evaluation. SL, DY, HX, KP, RX, ZL, and DZ contributed to the statistical analysis. KH, QX, WC, and LJ wrote the original manuscript. DW and TJ revised the manuscript. KH, QX, WC, and LJ contributed equally to this work and are joint first authors. DW and TJ contributed equally and are joint corresponding authors and the guarantors. The corresponding author attests that all listed authors meet authorship criteria and that no others meeting the criteria have been omitted.

    Conflicts of Interest

    None declared.

    Multimedia Appendix 1

    The 80-question benchmark with task-type labels and corresponding reference answers used for model evaluation.

    XLSX File, 62 KB

    Multimedia Appendix 2

    Outputs of DeepSeek R1, Gemini 3.0 Pro, Tencent HY 2.0, GPT-5.2, and HPHME-Xplus-RAG queried with the 80-question compiled set.

    XLSX File, 1665 KB

    Multimedia Appendix 3

    Supplementary descriptive statistics and interpanel reliability analyses for overall and dimension-specific scores. The tables include results for the 4 large language models and HPHME-Xplus-RAG.

    DOCX File, 48 KB

    Checklist 1

    iCHECK-DH checklist.

    DOCX File, 28 KB
    1. Dorji T, Giri S, Tshering U, et al. Challenges in the management of high-altitude illnesses and emergencies in Bhutan and Nepal. Travel Med Infect Dis. 2023;56:102660. [CrossRef] [Medline]
    2. Bardwell WA, Ensign WY, Mills PJ. Negative mood endures after completion of high-altitude military training. Ann Behav Med. Feb 2005;29(1):64-69. [CrossRef] [Medline]
    3. Rajesh YS, Bhat PS, Srivastava K, Kanitkar M. Study of psychological effects of deployment at extreme high altitude area on soldiers. Med J Armed Forces India. Oct 2021;77(4):426-430. [CrossRef] [Medline]
    4. Levine BD, Zuckerman JH, deFilippi CR. Effect of high-altitude exposure in the elderly: the Tenth Mountain Division study. Circulation. Aug 19, 1997;96(4):1224-1232. [CrossRef] [Medline]
    5. Wang L, Sang L, Cui Y, et al. Effects of acute high-altitude exposure on working memory: a functional near-infrared spectroscopy study. Brain Behav. Dec 2022;12(12):e2776. [CrossRef] [Medline]
    6. Liu Y, Zhang JH, Gao XB, et al. Correlation between blood pressure changes and AMS, sleeping quality and exercise upon high-altitude exposure in young Chinese men. Military Med Res. Dec 2014;1(1):19. [CrossRef]
    7. Kumar K, Shijith KP, Ahmad FMH. Bell’s palsy at high altitude -- an unsuspected finding. Military Med Res. Dec 2016;3(1). [CrossRef]
    8. Qu H, Yu LJ, Wu JT, et al. Spine system changes in soldiers after load carriage training in a plateau environment: a prediction model research. Military Med Res. Dec 2020;7(1). [CrossRef]
    9. Hussain MS, Tripathi V. Smoking under hypoxic conditions: a potent environmental risk factor for inflammatory and autoimmune diseases. Military Med Res. Dec 2018;5(1). [CrossRef]
    10. Mallet RT, Burtscher J, Richalet JP, Millet GP, Burtscher M. Impact of high altitude on cardiovascular health: current perspectives. Vasc Health Risk Manag. 2021;Volume 17:317-335. [CrossRef]
    11. Yang S, Zhao H, Zhang H, et al. Current status and continuing medical education need for general practitioners in Tibet, China: a cross-sectional study. BMC Med Educ. Mar 8, 2024;24(1):265. [CrossRef] [Medline]
    12. Hou J, Ke Y. Addressing the shortage of health professionals in rural China: issues and progress comment on “Have health human resources become more equal between rural and urban areas after the new reform?”. Int J Health Policy Manag. Mar 7, 2015;4(5):327-328. [CrossRef] [Medline]
    13. Tong W, Zhang X, Zeng H, Pan J, Gong C, Zhang H. Reforming China’s secondary vocational medical education: adapting to the challenges and opportunities of the AI era. JMIR Med Educ. Aug 15, 2024;10:e48594. [CrossRef] [Medline]
    14. Chen H, Wang Z, Zheng L, Ciren J, Zhang W, Zhu T. Anesthesiology education: a discussion on the current reform for training in the hypoxic environment of the highlands of China. Front Educ. Jan 13, 2023;7:1041500. [CrossRef]
    15. Li X, Zhang YL, Shen L, Huang YG. Remote video lecture education improves anesthesiology residency training in Tibet of China. Chin Med J. Jan 20, 2020;133(2):245-246. [CrossRef]
    16. Pollard AJ, Niermeyer S, Barry P, et al. Children at high altitude: an international consensus statement by an ad hoc committee of the International Society for Mountain Medicine, March 12, 2001. High Alt Med Biol. 2001;2(3):389-403. [CrossRef] [Medline]
    17. Parati G, Agostoni P, Basnyat B, et al. Clinical recommendations for high altitude exposure of individuals with pre-existing cardiovascular conditions. Eur Heart J. May 1, 2018;39(17):1546-1554. [CrossRef]
    18. León-Velarde F, Maggiorini M, Reeves JT, et al. Consensus statement on chronic and subacute high altitude diseases. High Alt Med Biol. 2005;6(2):147-157. [CrossRef] [Medline]
    19. Luks AM, Beidleman BA, Freer L, et al. Wilderness Medical Society Clinical Practice Guidelines for the prevention, diagnosis, and treatment of acute altitude illness: 2024 update. Wilderness Environ Med. Mar 2024;35(1_suppl):2S-19S. [CrossRef] [Medline]
    20. Gatterer H, Villafuerte FC, Ulrich S, Bhandari SS, Keyes LE, Burtscher M. Altitude illnesses. Nat Rev Dis Primers. Jun 20, 2024;10(1):43. [CrossRef] [Medline]
    21. Luitse D, Denkena W. The great Transformer: examining the role of large language models in the political economy of AI. Big Data Soc. Jul 2021;8(2):20539517211047734. [CrossRef]
    22. Peng C, Yang X, Chen A, et al. A study of generative large language model for medical research and healthcare. NPJ Digit Med. Nov 16, 2023;6(1):210. [CrossRef] [Medline]
    23. Wang D, Zhang S. Large language models in medical and healthcare fields: applications, advances, and challenges. Artif Intell Rev. Sep 2024;57(11). [CrossRef]
    24. Meng X, Yan X, Zhang K, et al. The application of large language models in medicine: a scoping review. iScience. May 17, 2024;27(5):109713. [CrossRef] [Medline]
    25. Mu LJ, Wang TT, Miao YD. Advancements in AI-driven oncology: assessing ChatGPT’s impact from GPT-3.5 to GPT-4o. Int J Surg. Jan 1, 2025;111(1):1669-1670. [CrossRef] [Medline]
    26. Vaswani A, Shazeer N, Parmar N, et al. Attention is all you need. Presented at: NIPS’17: Proceedings of the 31st International Conference on Neural Information Processing Systems; Dec 4-9, 2017:6000-6010; Long Beach, CA, USA. [CrossRef]
    27. Blank IA. What are large language models supposed to model? Trends Cogn Sci (Regul Ed). Nov 2023;27(11):987-989. [CrossRef]
    28. Fu Q, Yuan X, Han X, Wang W, Zhang J, Yuan X. Enhancing the diagnostic accuracy of sacroiliitis: a machine learning approach applied to computed tomography imaging. Br J Hosp Med. Aug 30, 2024;85(8):1-13. [CrossRef]
    29. Fan S, Hong H, Lu S, et al. Artificial intelligence-based predictive model for relapse in acute myeloid leukemia patients following haploidentical hematopoietic cell transplantation. J Transl Int Med. Jun 2025;13(3):253-266. [CrossRef] [Medline]
    30. Vrdoljak J, Boban Z, Vilović M, Kumrić M, Božić J. A review of large language models in medical education, clinical decision support, and healthcare administration. Healthcare (Basel). Mar 10, 2025;13(6):603. [CrossRef] [Medline]
    31. Tordjman M, Liu Z, Yuce M, et al. Comparative benchmarking of the DeepSeek large language model on medical tasks and clinical reasoning. Nat Med. Aug 2025;31(8):2550-2555. [CrossRef] [Medline]
    32. Rao A, Kim J, Kamineni M, Pang M, Lie W, Succi MD. Evaluating chatgpt as an adjunct for radiologic decision-making. Radiology and Imaging. Preprint posted online on Feb 7, 2023. [CrossRef]
    33. Zhou S, Luo X, Chen C, et al. The performance of large language model-powered chatbots compared to oncology physicians on colorectal cancer queries. Int J Surg. Oct 1, 2024;110(10):6509-6517. [CrossRef] [Medline]
    34. Wu X, Huang Y, He Q. A large language model improves clinicians’ diagnostic performance in complex critical illness cases. Crit Care. Jun 6, 2025;29(1):230. [CrossRef] [Medline]
    35. Sandmann S, Hegselmann S, Fujarski M, et al. Benchmark evaluation of DeepSeek large language models in clinical decision-making. Nat Med. Aug 2025;31(8):2546-2549. [CrossRef] [Medline]
    36. Han T, Adams LC, Bressem KK, Busch F, Nebelung S, Truhn D. Comparative analysis of multimodal large language model performance on clinical vignette questions. JAMA. Apr 16, 2024;331(15):1320-1321. [CrossRef] [Medline]
    37. Aster A, Laupichler MC, Rockwell-Kollmann T, Masala G, Bala E, Raupach T. ChatGPT and other large language models in medical education - scoping literature review. Med Sci Educ. Feb 2025;35(1):555-567. [CrossRef] [Medline]
    38. Aydin S, Karabacak M, Vlachos V, Margetis K. Large language models in patient education: a scoping review of applications in medicine. Front Med. Oct 29, 2024;11:1477898. [CrossRef]
    39. Zhang Y, Xie X, Xu Q. ChatGPT in medical education: bibliometric and visual analysis. JMIR Med Educ. Oct 7, 2025;11:e72356. [CrossRef] [Medline]
    40. Zhui L, Fenghe L, Xuehu W, Qining F, Wei R. Ethical considerations and fundamental principles of large language models in medical education: viewpoint. J Med Internet Res. Aug 1, 2024;26:e60083. [CrossRef] [Medline]
    41. Abouammoh N, Alhasan K, Aljamaan F, et al. Perceptions and earliest experiences of medical students and faculty with ChatGPT in medical education: qualitative study. JMIR Med Educ. Feb 20, 2025;11:e63400. [CrossRef] [Medline]
    42. Fan W, Ding Y, Ning L, et al. A survey on RAG meeting llms: towards retrieval-augmented large language models. 2024. Presented at: KDD ’24; Aug 25-29, 2024:6491-6501; Barcelona, Spain. [CrossRef]
    43. Huang L, Yu W, Ma W, et al. A survey on hallucination in large language models: principles, taxonomy, challenges, and open questions. ACM Trans Inf Syst. Mar 31, 2025;43(2):1-55. [CrossRef]
    44. Wang D, Liang J, Ye J, et al. Enhancement of the performance of large language models in diabetes education through retrieval-augmented generation: comparative study. J Med Internet Res. Nov 8, 2024;26:e58041. [CrossRef] [Medline]
    45. Miao Y, Zhao Y, Luo Y, Wang H, Wu Y. Improving large language model applications in the medical and nursing domains with retrieval-augmented generation: scoping review. J Med Internet Res. Oct 21, 2025;27:e80557. [CrossRef] [Medline]
    46. Chen X, Wang L, You M, et al. Evaluating and enhancing large language models’ performance in domain-specific medicine: development and usability study with DocOA. J Med Internet Res. Jul 22, 2024;26:e58158. [CrossRef] [Medline]
    47. Li H, Huang J, Ji M, Yang Y, An R. Use of retrieval-augmented large language model for COVID-19 fact-checking: development and usability study. J Med Internet Res. Apr 30, 2025;27:e66098. [CrossRef]
    48. Update to GPT-5 system card: GPT 5.2. OpenAI; 2025. URL: https://cdn.openai.com/pdf/3a4153c8-c748-4b71-8e31-aecbde944f8d/oai_5_2_system-card.pdf [Accessed 2026-05-26]
    49. Gemini 3 pro model card. Google. 2025. URL: https://storage.googleapis.com/deepmind-media/Model-Cards/Gemini-3-Pro-Model-Card.pdf [Accessed 2026-01-03]
    50. Tencent HY 2.0 product oeverview. Tencent Cloud. 2025. URL: https://cloud.tencent.com/document/product/1729/104753 [Accessed 2026-01-23]
    51. DeepSeek-R1: incentivizing reasoning capability in llms via reinforcement learning. GitHub. 2025. URL: https://github.com/deepseek-ai/DeepSeek-R1/blob/main/DeepSeek_R1.pdf [Accessed 2026-01-03]
    52. Cer D, Diab M, Agirre E, Lopez-Gazpio I, Specia L. SemEval-2017 task 1: semantic textual similarity multilingual and crosslingual focused evaluation. In: Bethard S, Carpuat M, Apidianaki M, Mohammad SM, Cer D, Jurgens D, editors. Presented at: Proceedings of the 11th International Workshop on Semantic Evaluation (SemEval-2017); Aug 3-4, 2017:1-14; Vancouver, Canada. [CrossRef]
    53. Gerily G. High Altitude Medicine. 2nd ed. Peking University Medical Press; 2020. URL: https://www.pumpedu.com/shop/5722.html [Accessed 2026-06-04] ISBN: 9787565922251
    54. Yuqi G. High Altitude Medical Protection Knowledge. Science Press; 2017. URL: https://www.books.com.tw/products/CN11431887 [Accessed 2026-06-04] ISBN: 9787030522313
    55. Hidalgo J, Re S, D’Almeida AG, editors. High Altitude Medicine: A Case-Based Approach. Springer International Publishing; 2023. [CrossRef] ISBN: 9783031350917
    56. Luks AM, Ainslie PN, Lawley JS, Roach RC, Simonson TS. Ward, Milledge and West’s High Altitude Medicine and Physiology. 6th ed. CRC Press; 2021. [CrossRef] ISBN: 978-0-429-44433-3
    57. High altitude medicine: a collaborative model for books. Peking Union Medical College Hospital, People’s Hospital of Tibet Autonomous. URL: http://pumc.vip:8001/PUMCH/ [Accessed 2026-06-15]
    58. Shah SV. Accuracy, consistency, and hallucination of large language models when analyzing unstructured clinical notes in electronic medical records. JAMA Netw Open. Aug 1, 2024;7(8):e2425953. [CrossRef] [Medline]
    59. Zakka C, Shad R, Chaurasia A, et al. Almanac - retrieval-augmented language models for clinical medicine. NEJM AI. Feb 2024;1(2). [CrossRef] [Medline]
    60. Ferber D, Wiest IC, Wölflein G, et al. GPT-4 for information retrieval and comparison of medical oncology guidelines. NEJM AI. May 23, 2024;1(6):AIcs2300235. [CrossRef]
    61. Lewis M, Thio S, Dobson RJ, Denaxas S. Grounding large language models in clinical evidence: a retrieval-augmented generation system for querying UK NICE clinical guidelines. arXiv. Preprint posted online on Oct 3, 2025. [CrossRef]
    62. Kresevic S, Giuffrè M, Ajcevic M, Accardo A, Crocè LS, Shung DL. Optimization of hepatological clinical guidelines interpretation by large language models: a retrieval augmented generation-based framework. NPJ Digit Med. Apr 23, 2024;7(1):102. [CrossRef] [Medline]
    63. Ke YH, Jin L, Elangovan K, et al. oRetrieval augmented generation for 10 large language models and its generalizability in assessing medical fitness. arXiv. Preprint posted online on Oct 11, 2024. [CrossRef]
    64. Alexandrou M, Kumar S, Mahtani AU, et al. Performance of large language models on the acute coronary syndrome guidelines using retrieval-augmented generation. JACC Cardiovasc Interv. Oct 27, 2025;18(20):2458-2467. [CrossRef] [Medline]
    65. Shool S, Adimi S, Saboori Amleshi R, Bitaraf E, Golpira R, Tara M. A systematic review of large language model (LLM) evaluations in clinical medicine. BMC Med Inform Decis Mak. Mar 7, 2025;25(1):117. [CrossRef] [Medline]

    ‎
    AI: artificial intelligence
    HPHME-Xplus-RAG: High-Altitude Public Health Medical Education–Xplus (Xiehe in Chinese for Peking Union Medical College Hospital and Xizang for People’s Hospital of Xizang Autonomous Region)–Retrieval-Augmented Generation
    ICC: intraclass correlation coefficient
    LLM: large language model
    PUMCH: Peking Union Medical College Hospital
    RAG: retrieval-augmented generation

    Edited by Stefano Brini; submitted 04.Feb.2026; peer-reviewed by Wendong Xiao, Wenhao Qi, Yunguo Yu; final revised version received 13.May.2026; accepted 14.May.2026; published 16.Jun.2026.

    Copyright

    © Kun He, Qiming Xiao, Wangyang Chen, Lisha Jing, Yabing Wang, Shuai Li, Daiyu Yang, Hemiao Xu, Ke Pang, Ruoyu Xiao, Zhashilamu Liu, Deji Zhuoga, Ruxuan Chen, Jingyi Li, Long Chang, Yangzhong Zhou, Zhe Zhang, Ran Li, Lujing Ying, Rutong Li, Hongwei Wang, Xin Yin, Ge Zhen, Siyi Cai, Qijun Shan, Qiang Wang, Danzeng Zhuoga, Ciren Yangjin, Gesang Luobu, Tu Ji, Dong Wu. Originally published in the Journal of Medical Internet Research (https://www.jmir.org), 16.Jun.2026.

    This is an open-access article distributed under the terms of the Creative Commons Attribution License (https://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work, first published in the Journal of Medical Internet Research (ISSN 1438-8871), is properly cited. The complete bibliographic information, a link to the original publication on https://www.jmir.org/, as well as this copyright and license information must be included.