We collected 1226 Chinese FFA reports and the corresponding clinical diagnoses of 728 patients from the Eye Center of the Second Affiliated Hospital of Zhejiang University (SAHZU) between August 2016 and September 2021. The clinical diagnosis of each eye was either classified as normal or one of the 6 primary retinal vascular diseases: diabetic retinopathy (DR), wet age-related macular degeneration, central serous chorioretinopathy (CSC), branch retinal vein occlusion (BRVO), central retinal vein occlusion (CRVO), and Vogt-Koyanagi-Harada disease (VKH). The clinical diagnosis was based on clinical information from the patients, primarily the FFA images and reports.

Ethical approval was obtained from the Ethics Committee in the SAHZU School of Medicine (2019-428). This research involves medical records data. We ensured that the medical records were deidentified and all private information was removed. The Institutional Review Board agreed to share access to the data with third parties, including sending it through application programing interfaces (APIs) provided by companies like OpenAI, or using it on web-based platforms like ChatGPT.

To diagnose the patient’s eye status based on the FFA report with ChatGPT, we designed a fixed instruction that concatenates the patient’s FFA report as the whole prompt for ChatGPT. The instruction consists of a specific task description and all alternative conditions. To fully exploit the potential of ChatGPT, we implemented different prompting strategies to investigate the potential effect and find the most appropriate way to apply it. First, we used the direct prompting strategy that requires ChatGPT to directly output the final option without explanations. Second, inspired by chain-of-thought prompting technology [23], we adopted a step prompting strategy to elicit the detailed reasoning process, which provides interpretability for the disease diagnosis. Finally, ChatGPT was primarily trained on English corpora and may have difficulty recognizing instructions and FFA reports in Chinese, as well as making use of internal knowledge. Therefore, we also rewrote the prompts in English while keeping the FFA reports in Chinese to conduct code-switching prompt examination. Therefore, we mainly investigated 4 prompt strategies: Direct-Chinese, Step-Chinese, Direct-English, and Step-English. The detailed prompts can be found in Multimedia Appendix 1.

To avoid the randomness of ChatGPT’s response, we set the inference temperature to 0 so as to choose the greedy decoding strategy via the API, making the response more focused and deterministic. Furthermore, we evaluated the robustness of ChatGPT to different languages by calculating the average performance of ChatGPT using different prompting methods. All tests were conducted on the same version of GPT3.5-Turbo-0301 using the official API of OpenAI. Figure 1 shows the overall workflow.

We designed a systematic evaluation scheme to assess the performance of ChatGPT. In addition to diagnostic performance, we also incorporated a combination of inference ability, omission of information, hallucinations, misinformation, and inconsistency assessments to evaluate the ChatGPT’s reasoning process, as follows:

For diagnostic evaluation, precision, recall, and F₁-score were calculated based on ChatGPT’s responses and gold clinical diagnosis. Additionally, to evaluate the diagnostic performance of ChatGPT, 2 ophthalmologists and 2 ophthalmology interns with 2 years of clinical experience from SAHZU were invited to diagnose 100 FFA reports, which were randomly selected according to the proportion of diseases. In terms of the evaluation on ChatGPT’s inference ability, the last 5 measurements were evaluated on the responses to the Step-Chinese and Step-English prompts by 2 ophthalmologists from SAHZU. We randomly selected 509 FFA reports (no more than 100 for each disease) and the corresponding ChatGPT outputs for evaluation. Before the formal evaluation, the ophthalmologists were requested to conduct an annotation with training and achieved a final interannotator agreement up to 94%. The statistical analysis between the Chinese and English prompts was performed with the use of Chi-square tests for categorical measurements and Student t tests or Wilcoxon rank-sum tests for continuous measurements, as appropriate. A 2-sided P<.05 was considered statistically significant.

The characteristics of the FFA reports and responses by ChatGPT are listed in Table 1. Direct-Chinese and Direct-English prompts received responses that directly provided the final options, and their mean (SD) lengths were 19.2 (4.4) tokens and 5.7 (1.7) tokens, respectively, while Step-Chinese and Step-English prompts received responses that provided the detailed reasoning process, and their mean (SD) lengths were 118.4 (71.8) tokens and 100.5 (36.9) tokens, respectively. Examples of different prompts and their responses are presented in Multimedia Appendix 1.

The Direct-English prompts achieved an overall precision of 79.61%, recall of 83.12%, and F₁-score of 80.05%, which was 9.58% higher than that achieved by the Direct-Chinese prompts (Table 2). The diagnostic performance varied significantly for each disease category. ChatGPT performed better in the normal and DR categories, with the F₁-scores exceeding 80%, but performed worse in the VKH and CSC categories, achieving F₁-scores of less than 4%. Additionally, the F₁-score in the BRVO category varied greatly, from 54.35% for Direct-Chinese prompts to 74.51% for Direct-English prompts.

In contrast, the Step-Chinese prompts achieved an overall precision of 76.24%, recall of 77.16%, and F₁-score of 75.61%, which was 2.15% higher than that achieved by ChatGPT for Step-English prompts. Compared with Direct-Chinese prompts, the F₁-score for Step-Chinese prompts was increased by 5.14% and provided the reasoning process, which is crucial for disease diagnosis. However, the diagnostic performance of Step-Chinese prompts in the normal and CRVO categories was far worse than that of Direct-Chinese prompts. This is mainly because, with Step-Chinese prompts, ChatGPT generated hallucinations for FFA reports in the normal category, which were wrongly diagnosed as CRVO. Figure 2 further demonstrates the confusion matrices of ChatGPT for the 4 prompting strategies.

Figure 3 shows the average F₁-score of ophthalmologists, ophthalmology interns, ChatGPT with English prompts (Direct-English and Step-English), and ChatGPT with Chinese prompts (Direct-Chinese and Step-Chinese). Although ChatGPT performed better than experts for some disease types (eg, Direct-English and Step-English prompts for the normal and CRVO categories and all prompts for the BRVO category), the overall performance of ophthalmologists was the best (89.35%), followed by ophthalmology interns (82.69%), ChatGPT with Direct-English and Step-English prompts (76.76%), and ChatGPT with Direct-Chinese and Step-Chinese prompts (73.04%).

Table 3 presents the outcomes of ChatGPT’s inference ability, as evaluated by 2 ophthalmologists. The results of continuous measurements were presented descriptively as mean (SD) values. Based on the 509 FFA reports, Step-English prompts exhibited a tendency to require more reasoning steps for each report (P<.001, Wilcoxon rank-sum test). Although the average number of inference errors was similar (0.4 per report) between Step-Chinese and Step-English prompts (P=.88, Wilcoxon rank-sum test), Step-English prompts brought less incompleteness (44.31%), hallucinations (0.59%), and misinformation (1.96%) (all P<.001, Chi-square test). Instead, Step-Chinese prompts, which involved fewer reasoning steps, were more prone to having key information from the original report overlooked during the reasoning process, although this difference lacked statistical significance (P=.68, Chi-square test). In addition, a few generated diagnoses were marked as inconsistent with the reasoning process, with no statistical difference observed between Step-Chinese and Step-English prompts. Overall, compared with Step-Chinese prompts, ChatGPT demonstrated stronger inference abilities for Step-English prompts, particularly with regard to the average number of reasoning steps, incompleteness, hallucinations, and misinformation.

Using different prompt strategies introduces some variability in ChatGPT's responses to a given FFA report. Hence, we evaluated the robustness of ChatGPT with different language prompts through calculating the average diagnostic performance for 4 prompting methods: Direct, Step, Step (more detail), and Step (2-step) (Multimedia Appendix 1). The precision, recall, and F₁-score, presented descriptively as mean (SD) values, were compared between Chinese and English prompts using Student t tests. As shown in Table 4, the results indicated significant differences in the recall (P=.03) and F₁-score (P=.04) between Chinese and English prompts, while no significant difference was observed in the precision (P=.18). The mean (SD) F₁-scores of ChatGPT with Chinese and English prompts were 70.02% (3.54%) and 76.47% (2.61%), respectively. In short, the diagnostic performance of ChatGPT with English prompts was better and more robust.

To the best of our knowledge, this is the first study to evaluate ChatGPT’s performance on non-English clinical text for retinal vascular disease diagnosis. We have developed a systematic evaluation scheme that encompasses objective diagnostic performance, professional inference abilities, and comparisons with the diagnostic ability of experts. Through large-scale experiments and analysis, we found the potential of LLMs as medical assistants to provide diagnosis, and identified challenges faced by ChatGPT in the field of health care, especially regarding language disparity.

Our results demonstrated that ChatGPT can preliminarily diagnose retinal vascular diseases based on Chinese FFA reports and achieved a high F₁-score of 80.05% at best. However, the diagnostic performance of ChatGPT varied significantly among different diseases and prompting languages. The performance for common DR was significantly better than that for the more uncommon VKH, which is relatively low in incidence and more difficult to diagnose. Another noteworthy phenomenon is the language disparity of ChatGPT. Given that the development and validation of ChatGPT predominantly relied on English-centric data sets [31] and that non-English medical corpora are even more scarce, compared to with English prompts, ChatGPT exhibited a significant decline in diagnostic performance with Chinese prompts, with a 6.45% decrease in F₁-score. This language disparity poses challenges for the effective application of ChatGPT in non-English clinical settings.

Meanwhile, the diagnosis accompanied by reasoning steps did not necessarily lead to performance improvement; F₁-scores decreased by 6.59% for English prompts but increased by 5.14% for Chinese prompts. This disparity may be attributed to ChatGPT’s training being mainly on English corpora, with Direct-English prompts enabling a straightforward mapping from input to diagnosis. In contrast, Step-English prompts tended to bring more mistakes than benefits through multistep internal reasoning. However, for Chinese prompts, the scarcity of Chinese training data results in limited knowledge for disease diagnosis. Step-Chinese prompts, with the requirement of a reasoning process, can effectively compensate for incomplete and incorrect reasoning caused by limited knowledge, although they may introduce some noise. The performance gap between different diseases and prompting strategies demonstrates the potential unfairness brought by the overrepresentation of the major diseases, languages, and countries. This limitation hinders the global applicability of ChatGPT, particularly in non–English-speaking countries.

From the perspective of clinical practice, ChatGPT’s diagnostic performance still did not reach the level of ophthalmologists or even ophthalmology interns. It is worth noting that ChatGPT may be conservative in disease diagnosis. Despite the instruction restriction (must identify one), certain responses involved multiple conditions or indicated an inability to conclude based on existing information. Notably, although ChatGPT can derive a reasoning process to improve clinical interpretability, ophthalmologists identified some harmful mistakes, such as generating medical findings not mentioned in the original reports and quoting incorrect prior knowledge. More in-depth investigation and careful regulation are required before applying ChatGPT in the health care domain. Also, it is imperative to incorporate more extensive and higher-quality clinical data sets and knowledge into ChatGPT [32-34].

Prior work in using artificial intelligence (AI) for the automated diagnosis of retinal vascular disease has yielded promising outcomes [35]. However, since some hospitals struggle to produce qualified FFA reports [36] and require ophthalmologists with extensive clinical experience or retinal specialists, the majority of these studies have predominantly focused on analyzing FFA images. Ryu et al [37] introduced an end-to-end deep convolutional neural network–based method specifically designed for the automatic detection of DR and the assessment of lesion status. Similarly, Ding et al [38] proposed a pipeline for detecting retinal vessels in FFA images using deep neural networks. Moreover, Li et al [39] presented a weakly supervised learning-based method for detecting fluorescein leakage, eliminating the need for manual annotation of leakage areas. In contrast to research predominantly centered on lesion detection or specific disease diagnoses, Zhao et al [40] developed an AI system capable of automating image phase identification, diagnosing 4 different types of retinal diseases, and segmenting ischemic areas using FFA images. In our study, we used ChatGPT with 4 different prompting strategies based on FFA reports to diagnose a series of retinal diseases. Notably, when using an English prompt for direct diagnosis, ChatGPT exhibited impressive performance in the classification of retinal vascular diseases, requiring no additional training.

Beyond diagnostic accuracy, researchers have dedicated efforts to enhance the interpretability of disease diagnoses [41,42]. The widely used method for this purpose is heatmap visualization [38-40], used to accentuate characteristic regions crucial for disease diagnosis. This method may not capture the nuanced interplay of features critical for accurate diagnosis, leading to a potential loss of information and subtlety in the interpretative process. In this study, ChatGPT showed promise in enhancing the interpretability of disease diagnoses by explaining the process of diagnostic reasoning step by step. Its capacity to generate human-readable responses also allows for a more intuitive understanding of the AI diagnostic process.

ChatGPT has been used for various applications in clinical care and research. While numerous studies have demonstrated promising outcomes in complex medical tasks, including the United States Medical Licensing Exam (USMLE) [17,43], simplifying imaging reports for patients [27] and aiding decision-making [44,45], it is crucial to note that ChatGPT exhibits certain limitations. In the execution of the aforementioned tasks, ChatGPT occasionally produces errors, such as hallucinations or incomplete information [46]. However, the preceding studies were limited to the application and evaluation of ChatGPT solely within English medical contexts, neglecting an exploration of its effectiveness in non-English clinical scenarios. This study fills this gap by leveraging Chinese FFA reports to assess ChatGPT’s diagnostic performance and inference abilities for retinal vascular diseases in a non-English clinical environment and exploring the appropriate prompt languages and strategies.

Our study has several limitations. First, we did not fully utilize all the information available in clinical scenarios to conduct a diagnosis, such as more detailed FFA images, which may have reduced the diagnostic accuracy due to incomplete information. Since ChatGPT cannot analyze images, we will further evaluate the capabilities of multimodal models in subsequent research. Second, this study was not conducted in clinical practice. A prospective clinical trial can better examine an LLM’s clinical benefit; we leave this to our future work.

This study conducted extensive experiments to evaluate the diagnostic capabilities of ChatGPT in retinal vascular diseases, including objective diagnostic performance and professional reasoning analysis evaluated by ophthalmologists. ChatGPT with English prompts for direct diagnosis performed best, achieving results close to the diagnostic performance of ophthalmology interns with 2 years of clinical experience. On the contrary, due to limited Chinese training data and knowledge, ChatGPT with Chinese prompts led to incomplete reasoning and poor diagnostic performance, which demonstrates that there is a significant language disparity in the application of ChatGPT in clinical environments. Additionally, although ChatGPT can derive a reasoning process with a low error rate, mistakes such as misinformation and hallucinations still exist, which will mislead the diagnose of retinal vascular diseases. This study generally reveals the potential of LLMs to serve as a helpful medical assistant to provide diagnosis in non-English clinical environments, but also demonstrates the potential limitations and the need to continually explore more robust LLMs in ophthalmology practice.