Xiaoni Liu1,2†Rui Lai3†Chaoling Wu4†Changjian Yan5Zhe Gan6Yaru Yang1Xiangtai Zeng7Jin Liu8Liangliang Liao8Yuansheng Lin9*
Hongmei Jing1*
Weilong Zhang1*- 1Department of Hematology, Lymphoma Research Center, Peking University Third Hospital, Beijing, China
- 2Department of Respiratory Medicine, First Affiliated Hospital Gannan Medical University, Ganzhou, China
- 3Department of Respiratory Medicine, The People's Hospital of Ruijin City, Ruijin, China
- 4Department of Respiratory Medicine, Affiliated Hospital of Jiujiang University, Jiujiang, China
- 5The Second Affiliated Hospital of Fujian Medical University, Quanzhou, China
- 6Gannan Medical University, Ganzhou, China
- 7Department of Thyroid and Hernia Surgery, First Affiliated Hospital of Gannan Medical University, Ganzhou, Jiangxi, China
- 8Department of Respiratory Medicine, Longnan First People's Hospital, Longnan, China
- 9Department of Emergency and Critical Care Medicine, Suzhou Hospital, Affiliated Hospital of Medical School, Nanjing University, Suzhou, China
Currently, there are still many patients who require outpatient triage assistance. ChatGPT, a natural language processing tool powered by artificial intelligence technology, is increasingly utilized in medicine. To facilitate and expedite patients’ navigation to the appropriate department, we conducted an outpatient triage evaluation of ChatGPT. For this evaluation, we posed 30 highly representative and common outpatient questions to ChatGPT and scored its responses using a panel of five experienced doctors. The consistency of manual triage and ChatGPT triage was assessed by five experienced doctors, and statistical analysis was performed using the Chi-square test. The expert ratings of ChatGPT’s answers to these 30 frequently asked questions revealed 17 responses earning very high scores (10 and 9.5 points), 7 earning high scores (9 points), and 6 receiving low scores (8 and 7 points). Additionally, we conducted a prospective cohort study in which 45 patients completed forms detailing gender, age, and symptoms. Triage was then performed by outpatient triage staff and ChatGPT. Among the 45 patients, we found a high level of agreement between manual triage and ChatGPT triage (consistency: 93.3–100%, p<0.0001). We were pleasantly surprised to observe that ChatGPT’s responses were highly professional, comprehensive, and humanized. This innovation can help patients win more treatment time, improve patient diagnosis and cure rates, and alleviate the pressure of medical staff shortage.
Introduction
Recently, the National Bureau of Statistics of China reported that there were over 8.42 billion outpatient visits in the country in 2022 (1). With such a large volume of patients seeking medical attention, effective triage becomes paramount for efficient and accurate diagnosis and treatment. Correct triage is crucial for the effective management of patients’ health conditions (2, 3). Traditional manual triage methods are often influenced by the experience and seniority of medical staff (4). However, intelligent triage systems, such as those based on AI, eliminate these potential biases (5). Studies have shown that smart phone triage applications can reduce the error rate in triage decisions, shorten consultation times, and help relieve the pressure on medical staff (5). Despite these advancements, the interaction mode of mobile App triage is still relatively fixed and may not provide personalized feedback to patients. In recent years, AI systems based on Chat Generation Pre-Training (ChatGPT) have gained significant attention and are increasingly being applied in healthcare settings (6). However, the application of ChatGPT in outpatient triage has not been fully explored. Therefore, this study aims to evaluate the utility of ChatGPT in outpatient triage. We hope to demonstrate the potential of ChatGPT to enhance triage accuracy, speed, and patient satisfaction, while also reducing the workload on medical staff.
Methods
This study employed a retrospective cohort study and a prospective cohort study.
Retrospective Cohort Study: In March 2023, we conducted a random sampling of 30 outpatient medical records out of the vast pool of 100,000, spanning across the departments of Internal Medicine, Surgery, Gynecology, Pediatrics, and the Emergency Department at the First Affiliated Hospital of Gannan Medical University. The symptoms (Supplementary Figure S1) of these 30 cases were representative of common clinical symptoms encountered in clinical practice (7, 8). ChatGPT was used to answer 30 corresponding questions, and the responses were then scored by 5 experts. All 30 responses were independently assessed by 5 experts and given a score, which was ultimately averaged to ensure accuracy and consistency.
Prospective Cohort Study: We provided a form with age, gender, and symptoms, and randomly assigned 45 outpatients to fill out. Based on the tabular information, triage was performed both manually and using ChatGPT. The consistency of manual and ChatGPT triage was evaluated by 5 experts, and statistical analysis was performed using the Chi-square test. The manual triage personnel included professionally trained nurses and healthcare-related personnel. The assessments of the 5 experts are independent.
The 5 experts were all doctors who had worked in tertiary general hospitals for more than 5 years and held the qualification of attending physicians. They worked in departments such as respiratory medicine, hematology, oncology, pediatrics, and general surgery. They are assessed independently, first answering questions based on their own expertise and then evaluating ChatGPT’s responses. The independent evaluation by experts was based on the following principles: 1. Accuracy of ChatGPT triage; 2. Clarity of language expression; 3. Degree of first aid awareness; 4. Service attitude.
This study was conducted anonymously and without compensation, and was approved by the Ethics Committee of Ruijin People’s Hospital (approval No. 2023002). ChatGPT-3.5 was used in the study.
Results
The retrospective cohort study revealed that among the 30 answers reviewed by 5 experts, 17 received high scores (10 and 9.5 points), 7 received relatively high scores (9 points), and 6 received relatively low scores (8 and 7 points; Figure 1A). The 17 high-scored answers reflect comprehensive and professional analysis, hierarchical diagnosis and treatment systems, first aid concepts, and humanization. The 7 high-scored answers are generally professional and comprehensive but have room for improvement. The 6 relatively low-scored answers are relatively incomplete and unprofessional. These are shown in Figure 1A, Table 1, and Supplementary Table 1.
Figure 1. The reviewer’s evaluation of ChatGPT outpatient triage. (A) The reviewer’s rating of ChatGPT’s answers. (B) Reviewers assessed the consistency of manual and ChatGPT triage. Q, Questions. D, Doctor. Yes: ChatGPT triage in the same number as manual triage. No: Number of ChatGPT triage was inconsistent with manual triage. ****p<0.0001. The chi-square test was used in this study.
Table 1. ChatGPT’s evaluation of outpatient symptoms and the doctor’s comments on ChatGPT’s answer.
The prospective cohort study revealed that among these 45 outpatients, five specialists considered manual triage to be particularly consistent with ChatGPT triage. We found that 3 reviewers thought that the consistency of manual and ChatGPT triage was 100% (p<0.0001; Figure 1B), 1 reviewer thought that the consistency was 95.6% (p<0.0001; Figure 1B), and 1 reviewer thought that the consistency was 93.3% (p<0.0001; Figure 1B).
Overall, the results indicated that ChatGPT’s answers provided accurate and professional triage information to patients without providing misinformation or harmful information to patients.
Discussion
Outpatient triage is a necessary service in many parts of the world, especially where primary care systems are weak and primary care physicians work short weeks (9). Outpatient triage can improve treatment efficiency, reduce hospital queuing time, and improve medical efficiency, better meeting patients’ medical needs (3). With a shortage of medical staff, short consultation times for primary care physicians and non-24-h outpatient triage staff (9), we needed a tool that could help patients triage in real time. Traditional websites and apps can help triage patients, but the disadvantage is that the operation is complex, the information is broad and confusing, and cannot provide instant personalized feedback (10). However, this study shows that manual triage is highly consistent with ChatGPT triage and can provide professional, comprehensive, and humanized triage. ChatGPT can provide an interactive experience closer to human conversation, providing instant personalized feedback. Furthermore, ChatGPT possesses certain constraints, encompassing potential biases, reliability issues, privacy apprehensions, and ethical considerations surrounding its utilization (11, 12). Consequently, it is imperative to consistently update, train, and enhance ChatGPT to guarantee the security and credibility of the information it provides. Additionally, ethical frameworks ought to be formulated to tackle ethical quandaries stemming from its application in healthcare. This study lacks a large multicenter study. For future inquiries, it is envisaged that we shall amass specimens from numerous hospitals, regions, and centers, thereby augmenting the sample size and executing a multicenter study. Additionally, the triage of outpatient patients utilizing other AI models will be evaluated and contrasted with the triage by ChatGPT, providing a comprehensive comparison of their respective efficiencies. In the future, we hope that ChatGPT triage can be operated in healthcare facilities, so that patients can have a more convenient and faster medical experience. It is anticipated that ChatGPT will attain broader adoption in the medical sphere in the foreseeable future.
Data availability statement
The original contributions presented in the study are included in the article/Supplementary material, further inquiries can be directed to the corresponding authors.
Author contributions
XL: Conceptualization, Data curation, Formal analysis, Investigation, Methodology, Supervision, Validation, Writing – original draft, Writing – review & editing. RL: Conceptualization, Data curation, Formal analysis, Investigation, Methodology, Supervision, Validation, Writing – original draft, Writing – review & editing. CY: Data curation, Formal analysis, Methodology, Writing – original draft. ZG: Data curation, Methodology, Writing – original draft. YY: Data curation, Formal analysis, Supervision, Writing – original draft. XZ: Data curation, Formal analysis, Writing – original draft. JL: Data curation, Formal analysis, Writing – original draft. LL: Data curation, Formal analysis, Writing – original draft. HJ: Conceptualization, Data curation, Formal analysis, Funding acquisition, Investigation, Methodology, Project administration, Resources, Software, Supervision, Validation, Visualization, Writing – original draft, Writing – review & editing. WZ: Conceptualization, Data curation, Formal Analysis, Funding acquisition, Investigation, Methodology, Project administration, Resources, Software, Supervision, Validation, Visualization, Writing – original draft, Writing – review & editing. CW: Conceptualization, Data curation, Formal analysis, Investigation, Methodology, Supervision, Validation, Writing – original draft, Writing – review & editing. YL: Conceptualization, Data curation, Formal Analysis, Funding acquisition, Investigation, Methodology, Project administration, Resources, Software, Supervision, Validation, Visualization, Writing – original draft, Writing – review & editing.
Funding
The author(s) declare that financial support was received for the research, authorship, and/or publication of this article. Funding for the study was provided by National Natural Science Foundation of China (81800195), Key Clinical Projects of Peking University Third Hospital (BYSYZD2019026 and BYSYZD2023014), Beijing Xisike Clinical Oncology Research Foundation (Y-NCJH202201-0049), the special fund of the National Clinical Key Specialty Construction Program, P. R. China (2023).
Acknowledgments
We used OpenAI’s generative AI tool ChatGPT to answer questions from the research survey, and the study group translated these questions and ChatGPT responses from Chinese into English. The original documentation of ChatGPT’s answers has been provided in the Appendix.
Conflict of interest
The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.
Publisher’s note
All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.
Supplementary material
The Supplementary material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/fpubh.2024.1391906/full#supplementary-material
Abbreviations
AI, artificial intelligence; ChatGPT, Chat Generation Pre-Training; Q, Questions; n, problem number
References
2. Christ, M, Bingisser, R, and Nickel, CH. Emergency triage. An overview. Dtsch Med Wochenschr. (2016) 141:329–35. doi: 10.1055/s-0041-109126
3. Dondi, A, Calamelli, E, and Scarpini, S. Triage grading and correct diagnosis are critical for the emergency treatment of anaphylaxis Children (Basel) (2022) 9:12. doi: 10.3390/children9121794
5. Xie, W, Cao, X, Dong, H, and Liu, Y. The use of smartphone-based triage to reduce the rate of outpatient error registration: cross-sectional study. JMIR Mhealth Uhealth. (2019) 7:e15313. doi: 10.2196/15313
6. Haug, CJ, and Drazen, JM. Artificial intelligence and machine learning in clinical medicine, 2023. N Engl J Med. (2023) 388:1201–8. doi: 10.1056/NEJMra2302038
7. Finley, CR, Chan, DS, Garrison, S, Korownyk, C, Kolber, MR, Campbell, S, et al. What are the most common conditions in primary care? Systematic review. Can Fam Physician. (2018) 64:832–40.
8. Wändell, P, Carlsson, AC, Wettermark, B, Lord, G, Cars, T, and Ljunggren, G. Most common diseases diagnosed in primary care in Stockholm, Sweden, in 2011. Fam Pract. (2013) 30:506–13. doi: 10.1093/fampra/cmt033
9. Mohammadibakhsh, R, Aryankhesal, A, Jafari, M, and Damari, B. Family physician model in the health system of selected countries: a comparative study summary. J Educ Health Promot. (2020) 9:160. doi: 10.4103/jehp.jehp_709_19
10. Hill, MG, Sim, M, and Mills, B. The quality of diagnosis and triage advice provided by free online symptom checkers and apps in Australia. Med J Aust. (2020) 212:514–9. doi: 10.5694/mja2.50600
11. Dave, T, Athaluri, SA, and Singh, S. ChatGPT in medicine: an overview of its applications, advantages, limitations, future prospects, and ethical considerations. Front Artif Intell. (2023) 6:1169595. doi: 10.3389/frai.2023.1169595
Keywords: artificial intelligence, ChatGPT, triage outpatients, AI, triage
Citation: Liu X, Lai R, Wu C, Yan C, Gan Z, Yang Y, Zeng X, Liu J, Liao L, Lin Y, Jing H and Zhang W (2024) Assessing the utility of artificial intelligence throughout the triage outpatients: a prospective randomized controlled clinical study. Front. Public Health. 12:1391906. doi: 10.3389/fpubh.2024.1391906
Edited by:
Yanwu Xu, Baidu, ChinaReviewed by:
Josef Šedlbauer, Technical University of Liberec, CzechiaDag Øivind Madsen, University of South-Eastern Norway (USN), Norway
Copyright © 2024 Liu, Lai, Wu, Yan, Gan, Yang, Zeng, Liu, Liao, Lin, Jing and Zhang. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.
*Correspondence: Yuansheng Lin, linys202012@163.com; Hongmei Jing, hongmeijing@bjmu.edu.cn; Weilong Zhang, zhangwl2012@126.com
†These authors have contributed equally to this work