Introduction

Artificial intelligence (AI) has transformed the way we approach a huge number of industries and tasks ranging from consumer products to computer programming, as well as medicine [1,2]. Previously, the development of clinical AI models required significant time and resources with highly domain-specific training data [3]. However, with the release of ChatGPT by OpenAI in November 2022, the application of AI to many industries, including medicine, became far more accessible [4]. Within the medical field, ChatGPT has potential applications in medical education, clinical reasoning and research [5–7].

ChatGPT is a large language model (LLM) powered by OpenAI’s Generative Pre-trained Transformer(GPT)-3.5 or -4 and was seen as a major breakthrough in AI [8]. Previous iterations of AI language models were largely based on sequential style neural networks, such as recurrent neural networks (RNN) and long short-term memory (LSTM) neural networks [9–12]. With the introduction of the transformer architecture by Vaswani et al. in 2017, this has been widely adopted to develop LLMs such as GPT [13]. The key innovations with transformer architecture are two-fold. Firstly, they can process sequences non-sequentially. This means they do not ‘forget’ tokens far back in a sequence and the next word prediction considers the whole context not just the last few words. Secondly, self-attention, which weighs each token differently depending on the context, without which, the model could be basing its prediction on irrelevant information [13]. This ultimately means that LLMs can generate novel sequences never previously observed by the model [1,14]. As LLMs are trained on a vast data set, this allows ChatGPT to perform a wide array of tasks without the need for any additional specific training. It is able to generate computer code and analyse data, translate between languages, write discharge summaries and answer examination questions [8,15–18]. This had led to ChatGPT being tested on several medical examinations since its release to the public [1,19–26]. Most notably, ChatGPT-3.5 was shown to operate at or near pass level in the infamous United States Medical Licensing Examination (USMLE) [1]. GPT-4 was then released on 14^th March 2023, which can reportedly solve more difficult problems with greater accuracy due to a broader general knowledge and problem-solving ability [27]. Indeed, with the upgraded GPT-4, ChatGPT’s USMLE step 1 score increased from 64.5% to 88% [1,28].

Following its success on the USMLE, ChatGPT-4 has now passed numerous national undergraduate medical licensing examinations including those from the Australia, China, Iran, Japan, Korea, Peru, Saudi Arabia and the UK [19–26]. ChatGPT’s performance has also been tested on several postgraduate medical examinations.

We wanted to further evaluate the ability of ChatGPT-3.5 and -4 to undertake postgraduate medical exams without any additional training. We conducted a systematic literature review (SLR) of ChatGPT’s performance in UK postgraduate medical assessments to analyse the level of knowledge it is able to accurately report. To consolidate this review, we tested ChatGPT-3.5 and -4 on Part 1 of the Membership of the Royal College of Physicians (MRCP Part 1) examination and included these results in the analysis. MRCP Part 1 is a postgraduate UK medical qualification accessible to doctors with a minimum of 12 months experience in medical employment, that tests candidates using multiple choice questions [29]. MRCP Part 1 was selected firstly, as it has been demonstrated to be a reliable examination [30]. Secondly, because it is an internationally recognised postgraduate medical assessment attempted by around 30% of UK medical graduates which forms a critical part of career progression for aspiring physicians in the UK [31].

Methods

Results

Systematic literature review

Initially 1116 articles were retrieved with 12 included (Fig 1). This gave a total of 12 postgraduate UK examinations that ChatGPT performance was tested on. The examinations encompassed the following specialties: anaesthetics (n = 2), general practice (GP; n = 1), neurology (n = 1), obstetrics and gynaecology (n = 1), ophthalmology (n = 2), orthopaedics (n = 1), surgery (n = 1), multi-specialty (n = 1). The average grade at which these examinations are sat for a UK candidate, expressed as years after graduating, is 1.92 years.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Flow diagram of stages of systematic literature review.

Cochrane Library encompasses library of: Systematic reviews; systematic review protocols; controlled clinical trials.

https://doi.org/10.1371/journal.pone.0307372.g001

All examinations were in the format of a written exam, comprising single best answer or multiple choice questions, except one [40]. Li et al. tested ChatGPT-4 performance on Membership of the Royal College of Obstetricians and Gynaecologists (MRCOG) Part 3 virtual Objective Structured Clinical Exam (OSCE) circuit in which ChatGPT-4 outperformed two human participants with a score of 77.2%, compared with 73.7% for human participants [40].

Combined analysis

ChatGPT-3.5 failed eight out of nine postgraduate UK medical exams it performed, and only achieved a mark within or above passing range in MRCP Part 1. The average score for ChatGPT-3.5 was 56% (SD 0.084, range 43%-69.7%), with an average difference between GPT-3.5 score and pass mark of -5.0%.

ChatGPT-4 passed nine out of eleven examinations it performed with an average score of 76.2% (SD 1.23, range 50%-100%), with an average difference between GPT-4 score and pass mark of +13.56%. The specialties that ChatGPT-4 failed were orthopaedics and radiology. The specific examinations that ChatGPT-4 failed were the Fellowship of the Royal College of Surgeons in Orthopaedics (FRCS (Orth)) Part 1 and the Fellowship of the Royal College of Radiologists (FRCR) Part 1 examinations, which require a minimum of 6 years and 2 years postgraduate clinical experience respectively [37,41]. Therefore, all postgraduate examinations that require 0–1 year of clinical experience were passed by ChatGPT-4.

ChatGPT-4 outperformed ChatGPT-3.5 in all examinations that were taken by both LLMs. This was reported to be significantly greater in four examinations, including MRCP Part 1 (P<0.001) [37,42,43]. The above results are outlined in Table 1.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Summary of included studies with pass marks and average score of ChatGPT-3.5 and -4 included. Results from this study are included in the final row.

https://doi.org/10.1371/journal.pone.0307372.t001

Discussion

ChatGPT-4 performed above passing level on the majority of postgraduate UK medical exams that it was tested on, with no additional training. However, these results must be interpreted with caution.

Our results demonstrate that ChatGPT-4 can answer MRCP Part 1 questions at a level that is above the historical pass mark. The MELD analysis suggest that this performance is a reflection of ChatGPT’s ability to interpret and process information rather than simply recall answers. The candidate scores for MRCP are historically variable. It has previously been shown that performance in MRCP Part 1 varies with medical school for UK applicants, and score on Professional and Linguistic Assessments Board (PLAB) examinations for international applicants [31,49]. Previous studies have demonstrated that top performers in MRCP Part 1 are applicants who attended Oxford or Cambridge University, followed by those who scored ≥35 marks above the PLAB1 pass mark (PLAB1 A1 35+) [31,49]. Oxford and Cambridge students scored an average of around +5%-10% above the MRCP Part 1 pass mark, with PLAB1 A1 35+ cohort just below this [49]. Our data indicates that ChatGPT-4 would outperform all UK medical schools by a considerable margin with an average of +22.8% above the pass mark, at least two times above the average pass mark as those who attended Oxford University. In addition, ChatGPT-3.5 would be outperformed by candidates from Oxford and Cambridge Universities but its score in the sample paper would outperform all other UK universities and international applicant scores.

Despite the evidence that ChatGPT-4 can perform well in postgraduate UK medical exams, there are still concerns with using LLMs both in medical education and its application to clinical practice. Due to their design, LLMs can produce text that sounds correct, but it cannot guarantee that it is accurate [50]. This has been demonstrated with ChatGPT, and while GPT-4 is an improvement on GPT-3.5, it is still prone to hallucination and fabrication [51]. ChatGPT has been shown to inaccurately report the content of genuine publications and when asked to generate a short literature review, 18% of GPT-4 citations were fabricated [52,53]. With many medical students and doctors using ChatGPT as an adjunct for education and clinical practice, this propensity for fabrication becomes problematic. Indeed, many papers express concerns regarding factual inaccuracies and spread of misinformation that could occur with widespread use of LLMs such as ChatGPT [54]. This propensity for misinformation is further exemplified by LLM’s poor explanation accuracy; GPT-4 demonstrated an explanation accuracy of 65.9% dropping from the correct rate of 80.5% in the FRCR Part 2A examination [42]. This suggests that while ChatGPT-4 can pass many of these postgraduate examinations, it may not be able to explain the answers correctly at a passing level. Both these factors, fabrication and reduced explanation accuracy, highlight the potential limitations for LLMs including ChatGPT as learning tools. However, the effects of fabrication and hallucination could potentially be minimised with the use of prompts which require LLMs to provide the rationale behind their clinical decision making [55,56]. Savage et al. propose a new LLM workflow in which LLMs provide a clinical reasoning rationale before their output. This offers clinicians an interpretable means to assess if the answer given by the model is true or false. Incorrect model responses are often accompanied by rationales with factual inaccuracies, rendering inaccurate answers identifiable [55]. Savage et al. suggest LLMs providing clinical reasoning rationale is achievable through diagnostic reasoning Chain-of-Though (CoT) prompting. This is a method whereby input prompts are altered to instruct a LLM to divide tasks into smaller reasoning steps. This more accurately reflects the step-by-step cognitive processes utilised by clinicians in medical practice. Despite this, only GPT-4 and not GPT-3.5 was able to imitate advanced clinical reasoning processes to arrive at an accurate diagnosis when tested with diagnostic CoT prompts, and while it was observed that GPT-4 can imitate clinical reasoning thought processes it cannot apply clinical reasoning in a comparable way to a human [55].

ChatGPT performance appeared to be enhanced with factual recall and diminished when required to apply more complex reasoning. ChatGPT performed well on shorter questions as well as those testing anatomy and pharmacology, while it performed poorly in physiology and legal or ethical questions [37,38,41]. Indeed, in MRCP Part 1, ChatGPT-4 performed worst in questions asking for further investigations, with 29.4% incorrect. This was thought to be a consequence of limited exposure, where ChatGPT is able to recall facts similar to those in the data set it was trained on, but struggles with more complex questions due to lack of higher-order thinking and limited clinical experience [41]. In contrast to this, other studies found that there was no significant difference in performance between higher and lower order questions (P = 0.816 for GPT-3.5, P = 0.427 for GPT-4) and equal proficiency when answering basic science and clinical questions [42,43]. Of the questions ChatGPT-4 got consistently incorrect on MRCP Part 1, there was no statistically significant difference between the number of incorrectly answered factual recall and complex reasoning questions (Z = 1.0263, P = 0.30302). Interestingly, ChatGPT’s performance was found to mimic performance of humans candidates, with worse results on questions humans found more challenging [39]. These differences in performance could be explained firstly due to an undertrained model, with lack of representation of data in questions where ChatGPT performed poorly [1,43]. Indeed, one study found that ChatGPT performed significantly better on general medicine than neuro-ophthalmology, suggesting lack of training on more esoteric topics [39]. However, with the improved performance of GPT-4 compared with -3.5 and a more up-to-date data set, this issue will likely become less prevalent. It could also be suggested that an insufficiency in human judgement at the initial reinforcement stages of model development, which is a significant liability of LLMs, may be additionally responsible for this change. In short this means that AI ability becomes concomitant with human ability, and would explain why ChatGPT performed worse on questions humans found more challenging [1,39,57]. This reduction in LLM’s performance in more complex reasoning could be mitigated through similar methods used to minimise hallucination. The Tree-of-Thoughts (ToT) system proposes the construction of a tree-like structure in which each node represents a partial solution in the problem-solving process. The tree comprises coherent sequences of language called “thoughts”. Each thought is then evaluated by the LLM using a deliberative reasoning process similar to that of type 2 human reasoning [56,58]. The ToT method has been applied to LLMs undertaking various complex reasoning tasks, including sudoku and crosswords, and the results suggest there cannot be non-hallucinatory, reliable human-like reasoning in LLMs without a strategy that employs a ToT, or similar system, as well as elements that mimic human-like working memory [56].

The majority of examinations discussed and analysed ChatGPT’s factual knowledge and reasoning. However, a number of medical examinations require adequate clinical skills, patient manner and professionalism–qualities required to be a safe and effective clinician [45]. It has been suggested that part of the reason ChatGPT was unable to perform well on some examination questions was due to lack of clinical experience [41]. Despite its lack of experience, ChatGPT was able to outperform human candidates at a postgraduate virtual OSCE and achieved strong marks in communication, for which empathy is essential [40]. This demonstrates ChatGPT’s ability to rapidly assemble complex clinical information into a coherent response [40]. Not only was this strong performance a surprise, but ChatGPT’s success within a field in which human interaction and communication is fundamental, has led people to suggest that we may need to change the way medical examinations are performed [43].

The strong performance of ChatGPT-4 on postgraduate UK medical exams highlights the huge potential for collaboration with AI for decision making, as well as medical education [38,44]. With the noticeable progress made from GPT-3.5 to -4 it is probable that future iterations of GPT will be even more competent [37]. Although LLMs have limitations to their utility, their ability to process complex information designed to assess high-level clinical reasoning, demonstrates the large potential AI models have as an adjunct role in medicine. It is essential for doctors to be aware of the limitations of LLMs, such as ChatGPT, so that the models can be used and applied effectively and safely into medical practice.

Conclusions

ChatGPT-4 performed at above passing level for the majority of UK postgraduate medical examinations it was tested on, with no additional training. ChatGPT-4’s performance on MRCP(UK) Part 1 greatly outperformed the historical average performance of applicants from all medical schools.

ChatGPT-4 outperformed ChatGPT-3.5 in all examinations that they both undertook, in keeping with their respective performances on previous examinations. While improved with GPT-4, ChatGPT is prone to hallucinations, fabrications and reduced explanation accurary which could limit its potential as a learning tool. The potential for these errors are an inherent part of LLMs, thus, despite the impressive trajectory of ChatGPT from -3.5 to -4, this may always be a limitation for the medical applications of LLMs, however there is potential to minimise these effects through strategies that mimic human deliberative reasoning such as Chain-of-Thought and Tree-of-Thought search systems.

It appears that ChatGPT performed better on factual recall than higher order complex clinical scenarios, and its performance mimicked the performance of human participants, performing worse on more difficult questions. This convincing performance is likely to improve with future iterations of ChatGPT, whereby ChatGPT may become an increasingly valuable medical education resource if applied appropriately. However, it is essential that clinicians are aware of the limitations of LLMs, such as ChatGPT, if these models are to be applied safely to medical practice.

Supporting information

Acknowledgments

The authors would like to acknowledge OpenAI for providing access to ChatGPT-3.5 and -4 and the memorisation effets Levenshtein detector (MELD) algorithm, which were used in this study.