The diagnostic accuracy of differential diagnoses generated by artificial intelligence (AI) chatbots, including the generative pretrained transformer 3 (GPT-3) chatbot (ChatGPT-3) is unknown. This study evaluated the accuracy of differential-diagnosis lists generated by ChatGPT-3 for clinical vignettes with common chief complaints. General internal medicine physicians created clinical cases, correct diagnoses, and five differential diagnoses for ten common chief complaints. The rate of correct diagnosis by ChatGPT-3 within the ten differential-diagnosis lists was 28/30 (93.3%). The rate of correct diagnosis by physicians was still superior to that by ChatGPT-3 within the five differential-diagnosis lists (98.3% vs. 83.3%, p = 0.03). The rate of correct diagnosis by physicians was also superior to that by ChatGPT-3 in the top diagnosis (53.3% vs. 93.3%, p < 0.001). The rate of consistent differential diagnoses among physicians within the ten differential-diagnosis lists generated by ChatGPT-3 was 62/88 (70.5%). In summary, this study demonstrates the high diagnostic accuracy of differential-diagnosis lists generated by ChatGPT-3 for clinical cases with common chief complaints. This suggests that AI chatbots such as ChatGPT-3 can generate a well-differentiated diagnosis list for common chief complaints. However, the order of these lists can be improved in the future.

Background: The diagnostic accuracy of differential diagnoses generated by artificial intelligence chatbots, including ChatGPT models, for complex clinical vignettes derived from general internal medicine (GIM) department case reports is unknown. Objective: This study aims to evaluate the accuracy of the differential diagnosis lists generated by both third-generation ChatGPT (ChatGPT-3.5) and fourth-generation ChatGPT (ChatGPT-4) by using case vignettes from case reports published by the Department of GIM of Dokkyo Medical University Hospital, Japan. Methods: We searched PubMed for case reports. Upon identification, physicians selected diagnostic cases, determined the final diagnosis, and displayed them into clinical vignettes. Physicians typed the determined text with the clinical vignettes in the ChatGPT-3.5 and ChatGPT-4 prompts to generate the top 10 differential diagnoses. The ChatGPT models were not specially trained or further reinforced for this task. Three GIM physicians from other medical institutions created differential diagnosis lists by reading the same clinical vignettes. We measured the rate of correct diagnosis within the top 10 differential diagnosis lists, top 5 differential diagnosis lists, and the top diagnosis. Results: In total, 52 case reports were analyzed. The rates of correct diagnosis by ChatGPT-4 within the top 10 differential diagnosis lists, top 5 differential diagnosis lists, and top diagnosis were 83% (43/52), 81% (42/52), and 60% (31/52), respectively. The rates of correct diagnosis by ChatGPT-3.5 within the top 10 differential diagnosis lists, top 5 differential diagnosis lists, and top diagnosis were 73% (38/52), 65% (34/52), and 42% (22/52), respectively. The rates of correct diagnosis by ChatGPT-4 were comparable to those by physicians within the top 10 (43/52, 83% vs 39/52, 75%, respectively; P=.47) and within the top 5 (42/52, 81% vs 35/52, 67%, respectively; P=.18) differential diagnosis lists and top diagnosis (31/52, 60% vs 26/52, 50%, respectively; P=.43) although the difference was not significant. The ChatGPT models’ diagnostic accuracy did not significantly vary based on open access status or the publication date (before 2011 vs 2022). Conclusions: This study demonstrates the potential diagnostic accuracy of differential diagnosis lists generated using ChatGPT-3.5 and ChatGPT-4 for complex clinical vignettes from case reports published by the GIM department. The rate of correct diagnoses within the top 10 and top 5 differential diagnosis lists generated by ChatGPT-4 exceeds 80%. Although derived from a limited data set of case reports from a single department, our findings highlight the potential utility of ChatGPT-4 as a supplementary tool for physicians, particularly for those affiliated with the GIM department. Further investigations should explore the diagnostic accuracy of ChatGPT by using distinct case materials beyond its training data. Such efforts will provide a comprehensive insight into the role of artificial intelligence in enhancing clinical decision-making.

Background: In the evolving field of health care, multimodal generative artificial intelligence (AI) systems, such as ChatGPT-4 with vision (ChatGPT-4V), represent a significant advancement, as they integrate visual data with text data. This integration has the potential to revolutionize clinical diagnostics by offering more comprehensive analysis capabilities. However, the impact on diagnostic accuracy of using image data to augment ChatGPT-4 remains unclear. Objective: This study aims to assess the impact of adding image data on ChatGPT-4’s diagnostic accuracy and provide insights into how image data integration can enhance the accuracy of multimodal AI in medical diagnostics. Specifically, this study endeavored to compare the diagnostic accuracy between ChatGPT-4V, which processed both text and image data, and its counterpart, ChatGPT-4, which only uses text data. Methods: We identified a total of 557 case reports published in the American Journal of Case Reports from January 2022 to March 2023. After excluding cases that were nondiagnostic, pediatric, and lacking image data, we included 363 case descriptions with their final diagnoses and associated images. We compared the diagnostic accuracy of ChatGPT-4V and ChatGPT-4 without vision based on their ability to include the final diagnoses within differential diagnosis lists. Two independent physicians evaluated their accuracy, with a third resolving any discrepancies, ensuring a rigorous and objective analysis. Results: The integration of image data into ChatGPT-4V did not significantly enhance diagnostic accuracy, showing that final diagnoses were included in the top 10 differential diagnosis lists at a rate of 85.1% (n=309), comparable to the rate of 87.9% (n=319) for the text-only version (P=.33). Notably, ChatGPT-4V’s performance in correctly identifying the top diagnosis was inferior, at 44.4% (n=161), compared with 55.9% (n=203) for the text-only version (P=.002, χ2 test). Additionally, ChatGPT-4’s self-reports showed that image data accounted for 30% of the weight in developing the differential diagnosis lists in more than half of cases. Conclusions: Our findings reveal that currently, ChatGPT-4V predominantly relies on textual data, limiting its ability to fully use the diagnostic potential of visual information. This study underscores the need for further development of multimodal generative AI systems to effectively integrate and use clinical image data. Enhancing the diagnostic performance of such AI systems through improved multimodal data integration could significantly benefit patient care by providing more accurate and comprehensive diagnostic insights. Future research should focus on overcoming these limitations, paving the way for the practical application of advanced AI in medicine.

Background: The efficacy of artificial intelligence (AI)-driven automated medical-history-taking systems with AI-driven differential-diagnosis lists on physicians’ diagnostic accuracy was shown. However, considering the negative effects of AI-driven differential-diagnosis lists such as omission (physicians reject a correct diagnosis suggested by AI) and commission (physicians accept an incorrect diagnosis suggested by AI) errors, the efficacy of AI-driven automated medical-history-taking systems without AI-driven differential-diagnosis lists on physicians’ diagnostic accuracy should be evaluated. Objective: The present study was conducted to evaluate the efficacy of AI-driven automated medical-history-taking systems with or without AI-driven differential-diagnosis lists on physicians’ diagnostic accuracy. Methods: This randomized controlled study was conducted in January 2021 and included 22 physicians working at a university hospital. Participants were required to read 16 clinical vignettes in which the AI-driven medical history of real patients generated up to three differential diagnoses per case. Participants were divided into two groups: with and without an AI-driven differential-diagnosis list. Results: There was no significant difference in diagnostic accuracy between the two groups (57.4% vs. 56.3%, respectively; p = 0.91). Vignettes that included a correct diagnosis in the AI-generated list showed the greatest positive effect on physicians’ diagnostic accuracy (adjusted odds ratio 7.68; 95% CI 4.68–12.58; p < 0.001). In the group with AI-driven differential-diagnosis lists, 15.9% of diagnoses were omission errors and 14.8% were commission errors. Conclusions: Physicians’ diagnostic accuracy using AI-driven automated medical history did not differ between the groups with and without AI-driven differential-diagnosis lists.

Background: The potential of artificial intelligence (AI) chatbots, particularly ChatGPT with GPT-4 (OpenAI), in assisting with medical diagnosis is an emerging research area. However, it is not yet clear how well AI chatbots can evaluate whether the final diagnosis is included in differential diagnosis lists. Objective: This study aims to assess the capability of GPT-4 in identifying the final diagnosis from differential-diagnosis lists and to compare its performance with that of physicians for case report series. Methods: We used a database of differential-diagnosis lists from case reports in the American Journal of Case Reports, corresponding to final diagnoses. These lists were generated by 3 AI systems: GPT-4, Google Bard (currently Google Gemini), and Large Language Models by Meta AI 2 (LLaMA2). The primary outcome was focused on whether GPT-4’s evaluations identified the final diagnosis within these lists. None of these AIs received additional medical training or reinforcement. For comparison, 2 independent physicians also evaluated the lists, with any inconsistencies resolved by another physician. Results: The 3 AIs generated a total of 1176 differential diagnosis lists from 392 case descriptions. GPT-4’s evaluations concurred with those of the physicians in 966 out of 1176 lists (82.1%). The Cohen κ coefficient was 0.63 (95% CI 0.56-0.69), indicating a fair to good agreement between GPT-4 and the physicians’ evaluations. Conclusions: GPT-4 demonstrated a fair to good agreement in identifying the final diagnosis from differential-diagnosis lists, comparable to physicians for case report series. Its ability to compare differential diagnosis lists with final diagnoses suggests its potential to aid clinical decision-making support through diagnostic feedback. While GPT-4 showed a fair to good agreement for evaluation, its application in real-world scenarios and further validation in diverse clinical environments are essential to fully understand its utility in the diagnostic process.

Background: Diagnostic errors are significant problems in medical care. Despite the usefulness of artificial intelligence (AI)–based diagnostic decision support systems, the overreliance of physicians on AI-generated diagnoses may lead to diagnostic errors. Objective: We investigated the safe use of AI-based diagnostic decision support systems with trust calibration by adjusting trust levels to match the actual reliability of AI. Methods: A quasi-experimental study was conducted at Dokkyo Medical University, Japan, with physicians allocated (1:1) to the intervention and control groups. A total of 20 clinical cases were created based on the medical histories recorded by an AI-driven automated medical history–taking system from actual patients who visited a community-based hospital in Japan. The participants reviewed the medical histories of 20 clinical cases generated by an AI-driven automated medical history–taking system with an AI-generated list of 10 differential diagnoses and provided 1 to 3 possible diagnoses. Physicians were asked whether the final diagnosis was in the AI-generated list of 10 differential diagnoses in the intervention group, which served as the trust calibration. We analyzed the diagnostic accuracy of physicians and the correctness of the trust calibration in the intervention group. We also investigated the relationship between the accuracy of the trust calibration and the diagnostic accuracy of physicians, and the physicians’ confidence level regarding the use of AI. Results: Among the 20 physicians assigned to the intervention (n=10) and control (n=10) groups, the mean age was 30.9 (SD 3.9) years and 31.7 (SD 4.2) years, the proportion of men was 80% and 60%, and the mean postgraduate year was 5.8 (SD 2.9) and 7.2 (SD 4.6), respectively, with no significant differences. The physicians’ diagnostic accuracy was 41.5% in the intervention group and 46% in the control group, with no significant difference (95% CI −0.75 to 2.55; P=.27). The overall accuracy of the trust calibration was only 61.5%, and despite correct calibration, the diagnostic accuracy was 54.5%. In the multivariate logistic regression model, the accuracy of the trust calibration was a significant contributor to the diagnostic accuracy of physicians (adjusted odds ratio 5.90, 95% CI 2.93‐12.46; P<.001). The mean confidence level for AI was 72.5% in the intervention group and 45% in the control group, with no significant difference. Conclusions: Trust calibration did not significantly improve physicians’ diagnostic accuracy when considering the differential diagnoses generated by reading medical histories and the possible differential diagnosis lists of an AI-driven automated medical history–taking system. As this was a formative study, the small sample size and suboptimal trust calibration methods may have contributed to the lack of significant differences. This study highlights the need for a larger sample size and the implementation of supportive measures of trust calibration.