DeepSeek-assisted problem-based learning for glaucoma education in an undergraduate ophthalmology clerkship: a randomized educational pilot study Authors: Zheyao Gu, Ting Xi, Chunyuan Zhang Published: 2026 ABSTRACT AbstractThis study evaluated a teaching approach that combines the open-source large language model (LLM) DeepSeek with problem-based learning (PBL) in a glaucoma clerkship for medical students. We conducted a randomized educational pilot study with 74 fourth-year medical students. The PBL-DeepSeek group used the DeepSeek assistant PBL teaching method, whereas the traditional group was instructed via the traditional teaching method. Both groups completed the same glaucoma course that was taught by the same instructor. Student performance was assessed through theoretical examinations and mini-clinical assessment exercises (Mini-CEX). Questionnaires were distributed to gather student feedback on the DeepSeek-assisted PBL teaching approach. Compared with the traditional group, the PBL-DeepSeek group achieved significantly higher theoretical test scores (P = 0.003). Furthermore, the PBL-DeepSeek group demonstrated notable enhancements in interrogation skills, diagnostic reasoning, and general clinical proficiency (all P values < 0.05). However, the learners in the PBL-DeepSeek group expressed concerns that the advice regarding medical issues provided by artificial intelligence may not always be accurate. The DeepSeek-assisted PBL teaching model shows promise for enhancing learning outcomes, satisfaction and efficiency, suggesting that it could be used to support medical education. Future research should increase sample sizes and extend follow-up periods to validate these findings. Supplementary InformationThe online version contains supplementary material available at 10.1038/s41598-026-58447-5.Keywords: DeepSeek, Problem-based learning, Ophthalmology, Clinical teachingSubject terms: Diseases, Health care, Medical research Introduction Glaucoma is a significant public health issue and is well known as one of the leading causes of irreversible blindness worldwide1. This disease is characterized by progressive optic neuropathy with irreversible visual field defects2. From an educational standpoint, the complex pathophysiology of glaucoma, the uncertainty surrounding early diagnosis, and the need for long-term management strategies are significant challenges for medical students. Problem-based learning (PBL) is a widely applied student-centered learning model in medical education. Studies have shown that PBL can enhance student engagement and improve independent learning and clinical reasoning skills3,4. In ophthalmology education, problem-based instructional approaches have been increasingly incorporated across multiple clinical topics to improve student engagement and clinical reasoning skills5,6. Although the educational benefits of PBL are widely recognized, it also presents significant pedagogical challenges. Effective participation in PBL requires learners to independently identify their learning needs, retrieve information, and integrate knowledge from various sources7,8. In the early stages of learning, students might face problems such as ineffective information retrieval, challenges in finding reliable resources, and incomplete acquisition of knowledge. These challenges not only increase students’ cognitive load and decrease learning efficiency but can also have adverse effects on learning outcomes9,10. Moreover, PBL relies heavily on collaborative discussion, and poor knowledge acquisition in the early stages may result in uneven student participation and poor group interaction, which affects the depth and efficiency of learning11. These limitations are particularly apparent in glaucoma education, where students must integrate knowledge of anatomy, disease processes, interpretation of imaging, visual field analysis, and long-term treatment decisions. The cognitive complexity of glaucoma may therefore present a challenge to traditional PBL approaches in the absence of adequate educational support mechanisms. Recent advances in large language models (LLMs) may offer novel solutions to these challenges12. Certain educators believe that the integration of AI-driven teaching methods will solve many problems in the clinical training of medical students. This integration can, for example, provide personalized learning content on the basis of an individual’s progress, provide real-time feedback to support clinical decision-making, and simulate rare or complex cases, enabling everyone to further develop their medical skills13–15. DeepSeek is an open-source LLM developed by a Chinese AI research institute16. It has impressive reasoning skills, low deployment costs, and a streamlined and efficient architecture, characteristics that suggest exciting educational possibilities. DeepSeek’s main advantage in medical education is its practicality as compared to other common commercial LLMs (such as ChatGPT). In particular, DeepSeek facilitates on-premises installation and institutional customization, thus enabling stable integration in teaching activities without the need to rely on external platforms or subscription services that may not be available in certain educational contexts17. This increases the accessibility, continuity and scalability of clinical training for medical students. Moreover, the contextualized and interactive design of DeepSeek is also aligned with the pedagogical needs of PBL18. However, like other LLMs, DeepSeek has several limitations, such as questions concerning data privacy, model reliability problems and the risk that it may provide misleading or incorrect information19,20. Nevertheless, DeepSeek is a suitable platform to support PBL in glaucoma education because of its balance of deployment ability, educational adaptability, and operational feasibility. Therefore, this study was designed to evaluate the educational effectiveness of DeepSeek-assisted PBL in teaching glaucoma during ophthalmology clerkship. Given the pilot nature of this study and the high degree of complexity of glaucoma education, the results must be interpreted within the study context and should not be generalized to other ophthalmic subspecialties without further validation. Methods Study design and setting This randomized educational pilot study was conducted among fourth-year undergraduate medical students at a public medical college in Jiangsu, China, from October to November 2025. Each student participated for two weeks, a period that aligns with the standard duration specified in our institution’s ophthalmology clerkship curriculum. All the students completed the pretest, the two-week teaching intervention, the postintervention theory exam, and the Mini-Clinical Evaluation Exercise (Mini-CEX) within the same timeframe. We did not conduct any long-term follow-up after the postintervention assessments. All the participants voluntarily signed written informed consent forms. Participants retained the right to withdraw from the study at any time, and withdrawal did not negatively affect their academic performance or learning opportunities. Participant data remained anonymous and confidential throughout the research process. Study participants and sampling The sample size was determined a priori using G*Power software on the basis of previous literature21,22 (d = 0.8, α = 0.05, power = 0.8, two-tailed independent sample comparison), yielding a required sample size of 26 participants per group. To account for potential attrition and ensure adequate statistical power, this study recruited 74 fourth-year clinical medical students (39 males and 35 females) majoring in clinical medicine at Nanjing Medical University. The inclusion criteria for the two groups were as follows: (1) fourth-year undergraduate clinical students with no prior exposure to AI-supported PBL training; (2) completion of foundational ophthalmology theory courses and achievement of a passing grade; (3) participation in all teaching activities of this course and completion of in-class, out-of-class, and final assessments; (4) no violations of rules or regulations during the course period; and (5) voluntarily completing the questionnaire survey. The exclusion criteria were as follows: (1) students who had not completed the foundational ophthalmology theory course; (2) students who did not achieve a passing grade; (3) students who encountered technical difficulties during AI learning courses (such as being unable to use electronic devices or access the internet); (4) students who had previously participated in courses incorporating AI-assisted PBL teaching; and (5) students who failed to complete the postcourse evaluation and questionnaire. Computerized randomization was used for group assignments, and all the participants were randomly assigned to either the PBL-DeepSeek group (37 participants) or the traditional group (37 participants). The PBL-DeepSeek group received PBL instruction assisted by DeepSeek, whereas the traditional group received traditional instruction. Notably, throughout the study, the PBL-DeepSeek group consistently used the DeepSeek-R1 model (DeepSeek-AI, 2025) without any version updates. Both groups received glaucoma curriculum using the third edition of the ophthalmology textbook edited by Ge Jian et al. and published by People’s Medical Publishing House and were taught by the same instructor. Owing to the nature of the educational intervention, blinding participants and instructors was not feasible. However, personnel responsible for assessing theoretical performance and evaluating Mini-CEX results, as well as those conducting statistical data analysis, remained unaware of group assignments to minimize potential bias. Teaching and learning programs Glaucoma is a complex disease. Effective management of glaucoma requires individualized treatment, and students must first have a solid understanding of ocular anatomy and clinical skills. This study focused on the core elements of glaucoma teaching: clinical presentation, medical history-taking, physical examination, diagnostic strategies, differential diagnosis, and treatment planning. Our goal was to deepen students’ understanding of glaucoma and improve their clinical reasoning and engagement during the ophthalmology rotation. A schematic overview of the study design and intervention workflow for both groups is shown in Fig. 1. Both groups underwent several phases of structured activities and the same assessments at the end of the rotation. Study design and intervention workflow. This chart compares the performance of the PBL-DeepSeek group (left) and the traditional group (right) across multiple stages. Before class, the PBL-DeepSeek group received AI training and used DeepSeek to prepare case studies; meanwhile, the traditional group reviewed the textbook independently. In class, the PBL-DeepSeek group engaged in group discussions about AI-generated diagnostic results and conducted simulated patient consultations, whereas the traditional group attended lectures and completed individual exercises. Before the end of the course, the PBL-DeepSeek group summarized key points through group collaboration, whereas the traditional group received the course’s learning objectives and key points from the instructor. After class, the students in both groups completed assignments assigned by the instructor and underwent the same theoretical examination and Mini-CEX assessment. DeepSeek-assistant group Creation of study groups Students were divided into seven groups with five to six members each. Each group selected a leader and established an internal communication channel. Preclass training Before the course began, all the students in the PBL-DeepSeek group participated in a 30-minute training session on using DeepSeek (R1, 2025). The training covered how to formulate effective queries and, more importantly, how to critically evaluate AI-generated information by cross-checking it against standard ophthalmology textbooks, clinical guidelines, and instructor feedback. Preclass activities The instructor distributed the glaucoma learning objectives to each group. Students used DeepSeek to explore these objectives, which covered clinical manifestations, history-taking, physical examination, diagnostic methods, differential diagnosis, and treatment options. The AI system also posed open-ended questions to stimulate discussion, such as “Do all individuals with elevated intraocular pressure develop glaucoma?” and “Why do patients often fail to recognize early-stage glaucoma?” Students practiced history-taking through simulated consultations with AI-generated virtual patients. Faculty members reviewed all AI-generated content before class, participated in group discussions, and corrected any inaccurate or misleading outputs in real time. In-class activities Classroom discussion: Students in each group first discussed the questions prepared in advance using DeepSeek (R1, 2025). Each group then selected a representative to summarize the discussion and share the group’s solutions with the class. Finally, the teacher summarized and commented on all the issues discussed. Classroom discussion: Students in each group first discussed the questions prepared in advance using DeepSeek (R1, 2025). Each group then selected a representative to summarize the discussion and share the group’s solutions with the class. Finally, the teacher summarized and commented on all the issues discussed. Clinical skills training: After the discussion, the students practiced patient-centered consultations and related physical examinations. Clinical skills training: After the discussion, the students practiced patient-centered consultations and related physical examinations. Before the end of the course, each group collaboratively wrote a summary that combined teacher-led discussion with DeepSeek-generated analysis, emphasizing clinical applications such as treatment of acute angle-closure glaucoma. Before the end of the course, each group collaboratively wrote a summary that combined teacher-led discussion with DeepSeek-generated analysis, emphasizing clinical applications such as treatment of acute angle-closure glaucoma. Traditional group Students in the traditional group did not participate in small‑group learning activities before, during, or after class. Instruction was instructor-centered, with students learning passively. The teaching process had four steps. First, the teacher used multimedia courseware to introduce glaucoma and explained the relevant diagnostic concepts in detail, with a special emphasis on medical history-gathering and eye examination methods. Second, the teacher comprehensively explained the clinical manifestations, diagnostic criteria, differential diagnoses and treatment options with respect to the different types of glaucoma. Third, the students practiced consultations independently at the patient’s bedside and received relevant ophthalmology training in the examination room under instructor guidance. Before the end of the course, the instructor briefly summarized key learning points via a PowerPoint presentation to reinforce learning objectives. The two groups received the same amount of instruction time. Data collection tool and technique Objective evaluation was conducted through theoretical examinations and clinical skills assessments, whereas subjective evaluation was performed via questionnaire surveys. Theoretical examinations Before the glaucoma course, all the students completed a 50-item pretest to establish their baseline knowledge of glaucoma. At the end of the clerkship, they took a theory examination and a separate clinical skills assessment. Both tests used multiple-choice and true/false questions. All questions were drawn from the ophthalmology question bank of Nanjing Medical University, which is updated annually. The content was aligned with the 9th edition of Ophthalmology23, an ophthalmology textbook published by the People’s Medical Publishing House, and the predefined learning objectives of the glaucoma clerkship. Two faculty members independently reviewed all the questions and classified each question into one of four domains-pathophysiology, diagnostics, treatment, or clinical reasoning. Any categorization disagreements were resolved through discussion until a consensus was reached. Both tests demonstrated acceptable internal consistency, with a Cronbach’s α of 0.90. The same tests and scoring criteria were used for both groups. All tests were computer-scored to ensure standardization. Each test was scored out of 100 points, with a score less than 60 classified as failing. Clinical skills assessments The Mini-CEX is a widely used tool for assessing clinical competence. Its application in ophthalmology education has been previously reported24. We adapted the assessment sheet from the clinical skills evaluation scale developed by Batra et al.25 and modified it to fit the glaucoma context. The adapted form retained the original seven domains: history-taking, clinical examination, humanism and professionalism, clinical judgment, counseling, organization and efficiency, and overall clinical competence. Each domain is scored on a 9-point scale. Scores are grouped into three performance levels: “Unsatisfactory” (1–3 points), “Satisfactory” (4–6 points), and “Excellent” (7–9 points). The Cronbach’s α for the overall scale was 0.90, indicating acceptable internal consistency. To minimize rater bias, the Mini-CEX evaluations were completed independently by three faculty raters at the level of associate professor or higher. The final grade for each student was the average of the grades from the three graders. All the graders received standardized training on the scoring system. Interrater agreement was measured using meanpairwise Cohen’s kappa coefficient, which showed substantial agreement (mean Cohen’s κ = 0.74). Questionnaire for the PBL-DeepSeek group To evaluate the educational effectiveness of the DeepSeek-assisted PBL teaching method, we conducted an anonymous questionnaire survey for the PBL-DeepSeek group only. The questionnaire consisted of 10 questions (Table 3). Each question was answered on a 1–5 point scale ranging from “Strongly Disagree” to “Strongly Agree.” Table 3Results of the DeepSeek satisfaction survey questionnaire.ItemStrongly disagreeDisagreeNeutralAgreeStrongly agreeI am satisfied with the teaching process.22(59.46%)15(40.54%)This teaching method sparked my interests in learning.6(16.22%)19(51.35%)12(32.43%)I have gained a deeper understanding of glaucoma.2(5.41%)24(64.86%)11(29.73%)The teaching method of the course meets my learning needs.23(62.16%)14(37.84%)The course helps me improve my teamwork skills.4(10.81%)24(64.87%)9(24.32%)The feedback I receive during the course is clear and easy to understand.7(18.92%)29(78.38%)1(2.7%)I would like to continue using this approach to instruction in future courses.3(8.11%)25(67.57%)9(24.32%)The course content reflects the latest medical research and guidelines.30(81.08%)7(18.92%)I am concerned that the course might provide inaccurate medical advice.2(5.4%)6(16.22%)19(51.35%)8(21.62%)2(5.41%)The information provided in the course enhance my clinical reasoning skills.7(18.92%)26(70.27%)4(10.81%) Results of the DeepSeek satisfaction survey questionnaire. Statistical analysis method The experimental data were processed and analyzed using IBM SPSS Statistics for Windows, Version 23 (IBM Corp., Armonk, NY, USA). Quantitative results are reported as the mean values with standard deviation (X ± s), and intergroup differences were assessed through independent sample t tests. For the categorical variables, frequencies (n) and percentages (%) are provided, with chi-square tests used to evaluate group comparisons. Statistical significance was set at p < 0.05 for all analyses. Results Basic characteristics between groups The study included 74 fourth-year medical students (39 males and 35 females). The PBL-DeepSeek group included 37 students (19 males and 18 females), with an average age of 21.24 ± 0.43 years. The traditional group consisted of 37 students (20 males and 17 females), whose average age was 21.30 ± 0.52 years. No significant differences were observed between the two groups in terms of age, sex or glaucoma basic practice examination scores. Therefore, the data of the two groups are comparable, as presented in Table 1. Table 1Basic characteristics of two groups.ItemPBL-Deepseek group(n = 37)Traditional group(n = 37)p-valueAge21.24 ± 0.4321.30 ± 0.520.629Sex0.816Male1920Female1817Glaucoma Basics Practice Exam scores60.43 ± 6.2562.27 ± 7.480.255PBL Problem-Based Learning, *p < 0.05, PBL-Deepseek group vs. Traditional group. Basic characteristics of two groups. PBL Problem-Based Learning, *p < 0.05, PBL-Deepseek group vs. Traditional group. Comparison of theoretical examination scores between groups A statistical analysis of the theoretical examination scores revealed that students in the PBL-DeepSeek group achieved an average score of 81.71 ± 4.41 points, which was significantly higher than the average score of students in the traditional group (76.91 ± 8.31 points) (p = 0.003; Cohen’s d = 0.72). Mini-CEX assessment of clinical skills Both groups of students completed the Mini-CEX assessment in an average time of 29.21 ± 1.71 min, with each student receiving feedback within 2.74 ± 1.92 min. Compared with the traditional group, the PBL-DeepSeek group showed statistically significant improvements in counseling skills, professional clinical competence, clinical judgment and overall clinical competence (all P < 0.05). Effect sizes were moderate to large, with Cohen’s d values ranging from 0.49 to 0.77, indicating that the observed differences are educationally meaningful and not merely statistically significant. Table 2 summarizes the detailed comparison of the Mini-CEX scores between the two groups. Table 2Mini-CEX assessment of clinical skills examination results of two groups.ItemPBL-Deepseek group(N = 37)Traditional group(N = 37)P-valueCohen’s dMedical history interview skills5.86 ± 1.324.92 ± 1.280.002*0.72Physical examination skills6.62 ± 1.065.86 ± 1.440.012*0.60Humanistic care/professionalism5.92 ± 2.114.97 ± 1.760.04*0.49Clinical judgment6.32 ± 1.235.43 ± 1.390.005*0.68Organizational effectiveness6.57 ± 1.725.30 ± 1.560.001*0.77Overall clinical competence5.22 ± 1.554.38 ± 1.660.028*0.52Counseling skills5.43 ± 1.634.41 ± 1.850.013*0.59Mini-CEX Mini-clinical Evaluation Exercise, PBL Problem-Based Learning. *p < 0.05, PBL-Deepseek group vs. Traditional group. Mini-CEX assessment of clinical skills examination results of two groups. Mini-CEX Mini-clinical Evaluation Exercise, PBL Problem-Based Learning. *p < 0.05, PBL-Deepseek group vs. Traditional group. Analysis of the questionnaire survey results Of the questionnaires collected, 37 were complete and valid; thus, our completion rate was 100%. The students in the PBL-DeepSeek group offered high praise regarding this teaching method. All of these students believed that this teaching method was well suited to their learning needs. With respect to the course content, 100% of the students felt that the DeepSeek-provided materials were closer to the latest medical research and guidelines, and 81.1% of the students considered the provided course explanations and answers to be clear and comprehensible. In terms of future preferences, 91.9% of the students found PBL courses with DeepSeek more attractive and hoped that this method would continue to be used in future PBL courses. However, 27% of the students expressed concerns about the accuracy of the medical advice that an AI course might provide. Table 3 details the survey results with respect to the PBL-DeepSeek teaching method. Discussion This study evaluated a DeepSeek-assisted PBL curriculum for glaucoma clerkship training. Our findings suggest that AI-assisted PBL may improve student engagement, information accessibility, and clinical reasoning performance. However, we must be careful about what these findings mean. The PBL-DeepSeek group had several advantages over the control group: students worked in small groups, they used the PBL format, and they received AI literacy training. Previous research26,27 has confirmed that group learning in itself can enhance engagement, facilitate peer discussion, and strengthen clinical reasoning skills. Therefore, the observed benefits should not be interpreted as evidence that AI alone produces these effects but rather as the synergistic outcome of a comprehensive intervention. Nevertheless, the present findings may suggest several potential educational applications of LLMs in ophthalmology education. As an open-source LLM, DeepSeek may offer practical implementation advantages in certain institutional settings, including local deployment capability, scalability, and adaptability to institution-specific educational requirements. Compared with certain proprietary AI systems, locally deployable models may be more feasible in resource-limited educational environments where financial, technical, or data-governance constraints restrict access to commercial platforms28. In addition, AI-generated clinical simulations may provide supplementary exposure to uncommon ophthalmic conditions that are infrequently encountered during routine clinical rotations, potentially enriching case-based discussion and supporting broader clinical exposure during undergraduate training29. However, this study revealed that 27% of students in the PBL-DeepSeek group expressed concern about potential inaccuracies in AI-generated medical information. These findings may reflect medical students’ emerging awareness of the limitations of LLMs, including algorithm-generated “hallucinations” or clinically misleading outputs. Given their limited clinical experience, medical students do not yet possess the ability to reliably identify such inaccuracies. A recent study30 reported similar findings, indicating that learners’ skepticism alone is insufficient to effectively guard against AI-generated medical misinformation. This clearly demonstrates that AI-generated educational content still requires teacher supervision, expert validation, and ongoing quality control, particularly in clinically complex fields, such as ophthalmology. To mitigate these risks, the AI-assisted learning activities investigated in this study were implemented within a teacher-led educational framework. Instructors reviewed AI-generated materials prior to classroom use, actively participated in PBL discussions and provided real-time clarification when potential inaccuracies or clinically misleading outputs were identified. Additionally, students were required to critically cross-reference AI-generated information against standard ophthalmology textbooks, establish clinical guidelines, and provide instructor feedback rather than treat AI-generated answers as authoritative clinical guidance. These practices align with recent research findings emphasizing the importance of human supervision, expert validation, and governance frameworks in the integration of LLMs into medical education31. Therefore, AI systems should be viewed as supplementary educational tools rather than substitutes for expert-guided clinical instruction and professional judgment. Future implementation of AI-assisted educational models should prioritize the establishment of structured faculty supervision and oversight mechanisms, as well as structured AI literacy training, to help students correctly understand the limitations of AI-generated content. This study has several limitations. First, the sample size was estimated on the basis of a relatively large, anticipated effect size, whereas educational intervention effects are often smaller in practice. In addition, this pilot study was conducted at a single institution in China and focused exclusively on glaucoma education, which may limit the generalizability of the findings to other ophthalmology subspecialties, educational settings, or cultural contexts. Second, the follow-up period was relatively short and did not allow assessment of the long-term sustainability of the educational effects. Third, the adapted Mini-CEX has not been externally validated, and the theoretical examination did not involve a formal content validity evaluation. Fourth, the postintervention questionnaire was administered only to the DeepSeek group, without baseline assessment or control-group comparisons, limiting interpretation of the subjective perception data. Finally, the PBL-DeepSeek group employed group learning and the PBL framework and participated in AI literacy training, which prevented us from drawing conclusions regarding the independent role of AI. Therefore, those findings should be considered preliminary and hypothesis-generating rather than definitive evidence of educational effectiveness. Future studies should include multicenter studies and employ factorial designs, larger sample sizes, longer follow-up periods, and cross-cultural validation across diverse educational settings. Conclusion Integrating DeepSeek with PBL could improve medical students’ understanding of ophthalmic diseases. However, further research with larger sample sizes and additional evaluation methods is needed to validate the effectiveness of the DeepSeek-PBL model in ophthalmology education. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1 Acknowledgements We are grateful to the educators and students who participated in both the course and the control-group portions of this study. Within the instructional setting, the DeepSeek-R1 model (Beijing DeepSeek Technology Co., Ltd.) served as a supportive generative AI tool. However, it is important to note that this technology played no role in the research process itself-neither in its conception, design, or execution nor in the analysis of data or the drafting of this manuscript. Abbreviations Problem-based learning Mini-clinical assessment exercises Large language model Artificial intelligence Author contributions Data curation: Zheyao Gu, Ting Xi; Data analysis: Zheyao Gu, Chunyuan Zhang; Methodology: Ting Xi, Zheyao Gu, Chunyuan Zhang; Project administration: Zheyao Gu, Ting Xi; Writing—original draft: Zheyao Gu; Writing—review and editing: Ting Xi, Chunyuan Zhang. Funding The author(s) received no financial support for the research, authorship, or publication of this article. Data availability Owing to confidentiality agreements signed with all the participants, the dataset generated in this study is not publicly available. However, it may be requested from the corresponding author upon reasonable request. All the relevant data are within the manuscript. Declarations Competing interests The authors declare no competing interests. Ethical approval and consent to participate This study was conducted in accordance with the Declaration of Helsinki. The study protocol was reviewed by the Ethics Committee of Suzhou Municipal Hospital Affiliated to Nanjing Medical University, which determined that the study qualified for exemption from formal ethical review as a low-risk educational intervention involving no patients, clinical procedures, or sensitive health information. Written informed consent was obtained from all individual participants included in the study. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Contributor Information Zheyao Gu, Email: 1181992191@qq.com. Ting Xi, Email: szslyyxt@163.com. References Associated Data Supplementary Materials Supplementary Material 1 Data Availability Statement Owing to confidentiality agreements signed with all the participants, the dataset generated in this study is not publicly available. However, it may be requested from the corresponding author upon reasonable request. All the relevant data are within the manuscript.