In this study, GPT-4, the latest LLM released as of April 2023, when the study was conducted, was used among the large-scale language models developed by OpenAI. The system is based on the Generative Pretrained Transformer (GPT) architecture and leverages the transformer model. The core concept of GPT-3 was based on a transformer model with 175 billion parameters [14]. Although the exact number of parameters in GPT-3.5 and GPT-4 were not publicly disclosed, it has been adjusted through reinforcement learning from human feedback. GPT-4 was reported to have improved accuracy in its responses compared to GPT-3.5 [15,16]. It is important to note that the model is not specifically fine-tuned for each task but is designed to perform well across a wide range of natural language processing tasks, including question answering, summarization, and dialog generation. At the time used in this study, GPT-4 had accumulated information through September 2021.

JNEP is held once a year, and the question texts and answers are published by the Japanese Ministry of Health, Labour, and Welfare. JNEP has undergone several changes in format and content, reflecting the evolving role of pharmacists in health care. These changes have resulted in an increased emphasis on practical skills and knowledge, such as applying pharmaceutical knowledge in a clinical context and understanding relevant laws and ethical considerations. The JNEP criteria are revised approximately every 4 years, with the latest change having occurred at the 106th JNEP in 2020. Because GPT-4 used in this study has information until September 2021, the 107th and 108th JNEPs held in February 2022 and February 2023, respectively, were used.

Each examination consists of 345 questions, divided into 3 blocks: 90 essential questions, 105 pharmacy theory questions, and 150 practical pharmacy questions. The essential questions are questions that examine the basic knowledge required for pharmacists. The pharmacy theory questions are based on the theoretical knowledge necessary to evaluate and solve common problems encountered in pharmacists’ practice. The practical pharmacy questions are designed to assess basic, practical, and general knowledge for solving common problems in health care and public health. The practical pharmacy questions are often based on practical and clinical cases. Each question is from one of the following 9 areas: physics, chemistry, biology, hygiene, pharmacology, pharmaceutics, pathobiology, regulations, and practice. All questions are in the form of multiple-choice answers; however, each question has 1 or 2 correct answers, and if 2 correct answers are chosen, the answer must be complete to be considered correct. In addition to questions in which the examinee must answer correctly or incorrectly from each field, there are questions in which the examinee must perform calculations based on the conditions of the question as well as questions related to drug therapy in the presented case. The calculation questions require examinees to perform calculations from given conditions, and the results are chosen from a list of answer choices. All questions are written in Japanese. The passing criteria for the 107th JNEP included achieving a minimum of 70% correct answers in essential questions, a minimum of 30% correct answers in each area, a minimum of 217 (62.9%) correct answers overall, and a limit of 2 of the several contraindicated choices. Contraindicated choices, in this context, refer to certain choices within the questions that, if selected, could be ethically incorrect, potentially violate laws, or pose significant risks causing harm to patients and public health. Because information on contraindications was not disclosed by the question authors, it was decided to exclude them from the criteria used for evaluation. The passing criteria for the 108th JNEP included 70% or higher correct answers in essential questions, 30% or higher correct answers in each area, 235 (68.1%) or more correct answers overall, and no more than 2 contraindicated choices. In recent years, the pass rate for JNEP has averaged around 70%, with the pass rates for the 107th and 108th JNEPs standing at 68.0% (9607/14,124) and 69.0% (9602/13,915), respectively.

Question texts and answer choices of the target JNEP were entered into GPT-4, and the selected choices were tabulated as GPT-4’s answers. The question text was input into GPT-4 in Japanese. Figure 1 shows an image of the question input to GPT-4 and its response.

In the 107th JNEP, there was 1 question for which the answer was undefined, and it was excluded from the list of questions to analyze. During the study, GPT-4 did not allow the input of diagrams as information; therefore, all questions that required reading diagrams to come up with answers were excluded from the answer set. All questions that were not answered were given a score of 0, and the accuracy rate was calculated and compared with the passing criteria. In addition, as a subanalysis, the accuracy rate was calculated by excluding questions that GPT-4 could not answer and then was compared to the passing criteria. Moreover, the accuracy rates were compared according to the question content—whether they were multiple-choice, calculation, or case study questions. For each comparison, the Pearson chi-square test was performed using JMP Pro 16 (SAS Institute Inc).

An ethics approval did not apply to this study. It should be noted that the examination questions and answers used in this study were originally produced and copyrighted by the Ministry of Health, Labour, and Welfare of Japan. These materials are publicly available and used for the purpose of academic research in this study. Any copyrights about the exam content belong to the Ministry of Health, Labour, and Welfare of Japan, and this study did not infringe upon these rights.

The results of inputting the 107th and 108th JNEP questions into GPT-4 showed that GPT-4 failed to meet some passing criteria. However, only for questions that GPT-4 could answer, its accuracy rate met all the passing criteria. In the past, LLMs have demonstrated the ability to answer several professional examinations at a passing level. For example, ChatGPT has been reported to be capable of answering questions of the law school and business management course examinations [17] and the final exam for the Master of Business Administration field [18] at a passing level. Furthermore, for the medical field, it has been reported that ChatGPT’s score on the USMLE is equivalent to the passing score of third-year medical students [7] or close to the passing standard [8]. In Japan’s National Medical Practitioners Qualifying Examination, the accuracy rate was also reported to be 55.0% [19]. The results of this study showed a higher accuracy rate than ChatGPT’s performance in medical examinations, as reported in previous studies. The main reason for this difference is thought to be the distinct LLM used and the varying knowledge requirements for physicians and pharmacists. The LLM used in this study was GPT-4, which was released on March 14, 2023. It is said to have had a more complex neural network and larger training data set than the older models [20], which may have led to the results of this study. Kasai et al [9] reported that GPT-4 achieves the best performance on Japan’s National Medical Practitioners Qualifying Examination questions compared to ChatGPT, ChatGPT-EN, and GPT-3, and it passed the exams of all 6 years [9]. In the study by Kasai et al [9] and this study, the questions were entered in Japanese, indicating that GPT-4 is highly effective in decoding content and providing accurate answers without translation into English. The GPT-4 Technical Report [21] reported different accuracy rates for questions in English and Japanese (85.5% vs 79.9%). Therefore, it is suggested that higher accuracy rates may be obtained by translating questions into English and then inputting them into GPT-4.

Although GPT-4 is not specifically trained or specialized in any particular field of study, it has demonstrated a certain level of ability to respond to questions in each of these areas. However, the accuracy rate varied depending on the field of questions. When limited to questions that GPT-4 could answer, accuracy rates tended to be higher for biology and pharmacology questions and lower for physics and chemistry. Nisar et al [22] reported the results of the ChatGPT test on pharmacology for undergraduate students, which showed that they adequately answered various questions on drug’s pharmacokinetics, mechanism of action, clinical uses, adverse effect, contraindications, and drug-drug interactions [22]. On the other hand, in physics questions, GPT-4 often answered incorrectly to questions about analysis techniques, such as liquid chromatography and electrophoresis, as well as questions about purity tests and determination methods, which are described in the Japanese Pharmacopoeia [23]. Although analysis methods and the Japanese Pharmacopoeia in English can be searched on the internet [23], it is a highly technical field, and GPT-4 may not have been adequately studied. In Antaki et al’s [24] report of the evaluation of ChatGPT answers to ophthalmology questions, the results were good for general medicine but not for highly specialized fields, such as neuro-ophthalmology and ocular pathology. Therefore, it is expected that GPT-4 would perform lower in highly specialized areas due to inadequate learning. Many of the chemistry questions included diagrams of chemical structures, and only 16% (8/50) of chemistry questions could be input into GPT-4. Therefore, it is impossible to clarify the performance of GPT-4 with the chemistry field from this result.

In this survey, there were not only simple correct or incorrect questions about events but also many questions in which a case was presented and the question was about pharmacotherapy for the case. The accuracy rate of the case questions that GPT-4 could answer was 74.8%, which was as high as the percentage for all questions except for the case questions. This indicates that GPT-4 could be used by pharmacists to support their pharmacotherapy practice in clinical settings. However, in discussing the limitations of GPT-4 in real-world pharmacy practice, several factors should be considered. Although GPT-4 demonstrated strong performance on standardized exam questions, its effectiveness in handling diverse clinical scenarios and patient-specific factors may be limited. This is primarily due to the challenge of processing a wide range of patient information that extends beyond the scope of exam questions. In real-world clinical settings, patient data include detailed medical history, medication history, laboratory data, and allergy information, which change with time. It is unclear whether GPT-4 can accurately process such diverse information. Another important consideration is the lack of communication skills with other health care professionals. In real clinical practice, pharmacists collaborate and exchange information with various members of the health care team. However, GPT-4 cannot mimic this collaborative communication with other professionals. It was reported that ChatGPT lacks thoughtful reasoning like humans [25] and cannot evaluate information critically [26]. Consequently, the utility of GPT-4 in team-based health care provision may be limited. By recognizing and addressing these limitations, a more comprehensive evaluation of GPT-4’s practical applicability in clinical settings can be achieved. It is essential to acknowledge that GPT-4’s effectiveness in handling the complexities of real-world clinical practice, including diverse patient scenarios and collaboration with other health care professionals, needs further consideration and exploration. In addition, it is important to note that GPT-4, as it currently stands, is not compliant with patient privacy information [26].

It is important to note that the accuracy rate of GPT-4 was not 100%, and caution is needed regarding ethical issues related to the input of patients’ personal information [26]. It is reported that ChatGPT is prone to a phenomenon known as “hallucination,” which involves the generation of scientifically false content that appears sound to nonexperts [26]. Therefore, it is risky to rely completely on the generated content.

In addition, the accuracy rate for questions for which the participants were required to perform calculations under the indicated conditions was low (42.9%). In some cases, the results were incorrect due to the omission of values with different units, and in other cases, the results were correct; however, GPT-4 made a mistake in selecting the option with the closest value. In a previous report [17], it has been noted that there were surprisingly erroneous answers to the calculation questions, and the answers to the calculation questions were considered unreliable. Furthermore, questions that included diagrams could not be entered in this survey and were excluded from the answers, but diagram recognition is essential to a pharmacist’s ability to infer drug characteristics from the structural formula of a substance or to predict drug changes from chemical reaction formulas. These limitations should be considered when using GPT-4 in clinical practice. It is expected that in the future, LLMs will be developed to be capable of recognizing diagrams and photos as information. Furthermore, although the level of knowledge required of pharmacists was assessed by having them answer the JNEP questions using GPT-4, this may not entirely reflect the knowledge and suggestions that pharmacists are required to provide in clinical settings. GPT-4 does not have an inherent knowledge of which answers are right or wrong but rather generates responses based on patterns and information present in its training data. Therefore, it cannot provide responses beyond the information present in its training data available on the web. However, it is important to note that pharmacists may often face questions and scenarios that are not readily available on the internet. In addition to the limitations discussed above, it is important to acknowledge that the GPT-4 model used in this study was pretrained until September 2021 and does not have access to the internet or other resources beyond that date. Given the rapidly changing nature of fields like pharmacy, which sees the introduction of new medications annually and the release of updated treatment guidelines every few years, it is essential to recognize that GPT-4 may not be up to date with the latest information. This study provides insights into GPT-4’s capabilities within its training data timeframe, and therefore, caution should be exercised when applying its results to real-world clinical practice, and reliance on the most current sources and specialized knowledge is necessary. To evaluate whether GPT-4 can be used as an auxiliary tool for pharmacist work in the future, verification using more detailed work data sets (eg, patient counseling, records of inquiries from physicians, drug interaction analysis, and examples of questionable prescriptions) is required.

In conclusion, GPT-4 showed that some passing criteria were not met in terms of the accuracy rate for all JNEP questions, but the accuracy rates for the questions that GPT-4 could answer met all of the passing criteria. Nevertheless, recognizing the limitations of the current GPT-4 model is crucial, particularly in terms of its performance in answering highly specialized questions, calculation questions, and questions requiring diagram recognition. Furthermore, exploring the practical applicability of GPT-4 in real-world clinical settings is essential by evaluating its performance on more detailed work data sets (eg, patient counseling, records of inquiries from physicians, drug interaction analysis, and examples of questionable prescriptions). By addressing these limitations and validating its performance in a broader range of tasks, GPT-4 could become a more reliable and effective tool for pharmacists in their day-to-day practice.