Vignette-based comparative analysis of ChatGPT and specialist treatment decisions for rheumatic patients: results of the Rheum2Guide study

The various manifestations and associated comorbidities of rheumatic diseases present considerable challenges for pharmacological treatment. When devising individualized treatment plans, clinicians must carefully evaluate an increasing array of options, considering patient-specific factors, safety criteria, and the latest research and guidelines [1]. The European Alliance of Associations for Rheumatology (EULAR) Committees consistently develop and publish treatment recommendations for various rheumatic conditions. The continuous refinement and extension of these EULAR recommendations ensures that practitioners have access to up-to-date and standardized guidance for managing rheumatic diseases and their specific manifestations.

However, aligning personalized treatment decisions with the latest EULAR recommendations and translating them into personalized treatment decisions is becoming more and more challenging. Increasing patient complexity often results in treatment decisions that resemble a trial-and-error approach, underscoring the confusion among rheumatologists and the critical need for decision support [2].

Rapid advancements in artificial intelligence have led to the development of powerful large language models (LLMs). Generative Pre-trained Transformer (GPT), an advanced LLM developed by OpenAI, is one of the most frequently consulted LLMs. Research into the use of GPT in medical decision-making is burgeoning [3, 4]. In rheumatology, GPT has already demonstrated the ability to pass rheumatology exams [5] and even exceeded the diagnostic accuracy of rheumatologists to diagnose inflammatory rheumatologic diseases based on medical history [6]. GPT-4 provided faster, higher quality and even more emphatic answers to frequently asked questions from patients with systemic lupus erythematosus [7]. EULAR acknowledged the potential of big data and called for benchmarking studies [8]. To our knowledge the ability of GPT in generating treatment plans for rheumatic diseases has not yet been investigated and critical scientific evaluations are urgently needed [9]. This study aimed to compare the acceptance, safety, guideline adherence, medical adequacy, overall quality, completeness and justification of first-line and second-line treatment plans generated by GPT-3.5, GPT-4 and rheumatologists for patients with inflammatory rheumatic diseases.

This study aimed to compare treatment plans generated by GPT-3.5, GPT-4 and rheumatologists for rheumatic patients. Overall, rheumatologist-derived treatment plans were preferred in the majority of cases by blinded rheumatologists. In addition, RB treatment plans were rated significantly better regarding all investigated parameters except for the safety of the first-line GPT-4 treatment plans and the logic of justification of the second-line GPT-4 treatment plans. Overall, inter-rater agreement was low, indicating notable disagreement among raters regarding the quality and preferences of treatment concepts. This underlines the potential of treatment decision support systems.

A closer examination of the individual treatment decisions revealed that rheumatologists’ plans were shorter and more nuanced, whereas the LLM treatment plans were lengthier and more formulaic. For instance, the rheumatology board (RB) made situational decisions favoring dose increases (case 10) of established therapies and local steroid injections (case 2). In contrast, the LLM consistently recommended changes in DMARD therapy.

Non-pharmacological therapies were not regularly mentioned by the LLM and the RB, despite physiotherapy being a fundamental component of treatment for conditions like axial spondylarthritis, as reflected in the 2022 guidelines[12].

Relevant comedication, such as cotrimoxazole for pneumocystis prophylaxis during treatment with cyclophosphamide or rituximab alongside prednisolone therapy[13], entecavir prophylaxis for hepatitis B status under rituximab therapy[14], antibiotic prophylaxis for latent tuberculosis and planned TNF inhibition[14] or vitamin D prophylaxis for prednisolone treatments[15], was sometimes, but not regularly recommended by the LLM (cases 5, 7, 13, 17–20).

In defense of the LLM, it can be argued that the absence of certain considerations was due to the lack of explicit prompts. However, comedication and non-pharmacological therapies were not excluded from the prompts, underscoring the importance of prompting using LLMs like OpenAI's ChatGPT in clinical practice: what is not explicitly mentioned may likely be overlooked. This necessitates situational optimization of prompting, implying that the required response is somewhat anticipated and that prior knowledge is needed to identify gaps in the responses. Venerito et al. recently suggested that prompt engineering, which involves the systematic design and refinement of prompts to enhance a model's performance on specific tasks, could become a crucial skill in rheumatology [16]. Future research could focus on developing comprehensive yet standardized and consented treatment plan prompts for inflammatory rheumatic diseases.

The LLM treatment plans contained several incorrect or at least debatable recommendations. ChatGPT-4 repeatedly listed hydroxychloroquine (quensyl) as a treatment option for RA (cases 1 and 4), despite prompts for guideline-compliant treatments. According to the EULAR 2022 guideline [1], hydroxychloroquine should only be chosen if the other csDMARDs are contraindicated or not tolerated. Rituximab was sometimes given in incorrect dosages or application intervals (cases 14 and 19). GPT-4 recommended the JAK inhibitor baricitinib to a patient with breast cancer and deep vein thrombosis (case 3), despite the EMA recommendations in this regard and the results of the ORAL Surveillance study [17]. In the patient with high-risk antiphospholipid syndrome (APS) and a history of venous thromboembolism while taking the direct oral anticoagulant (DOAC) apixaban, the RB advocated a switch to a vitamin K antagonist, while the LLM would continue the DOAC. While the EULAR guideline allows for continuation of therapy in the absence of arterial thromboses [18], it does not specifically recommend it. However all four raters unanimously preferred changing the therapy to a vitamin K antagonist. This preference appears to have no alternative given the uncertain study situation regarding DOACs and APS [19], as well as current German guidelines [20] and warnings [21].

In the case of SSc (case 15), it is important to note that at the time of the study the current guideline was from 2017 [22] and did not yet include therapies such as tocilizumab [23], rituximab [24], and antifibrotic therapy with nintedanib [25], whose effectiveness are proven. Both the LLM and the RB selected these newer guideline-incompliant therapy options, demonstrating the LLM's capacity for intelligent adaptation and application of up-to-date medical knowledge.

Not only the medical treatment decisions, but also their justifications were rated significantly better. Compared to the LLM, the RB's justifications were more concise, more concrete, and directly related to the patient case. In contrast, the LLM often provided general explanations about why the chosen medication is effective for the disease. Elsewhere, the RB's answers were more critical, for example by pointing to a differential diagnosis, namely drug abuse with an unusual dual-positive MPO- and PR3-ANCA positive status in the context of small vessel vasculitis (case 18).

The lack of significant differences between GPT-3.5 and GPT-4 was surprising, as other studies, such as those evaluating performance on the Spanish rheumatology exam [5], showed that the more advanced GPT-4 performed significantly better. Compared to OpenAI's free version ChatGPT-3.5, the paid version GPT-4 includes more recent content. The significantly higher word count in ChatGPT-4 responses did not result in significantly better evaluations. None of the treatment decisions made by the LLMs included therapies approved after 2022. In contrast, the RB sometimes opted for more modern treatments, such as anifrolumab, which was approved by the EMA for SLE in 2022 (case 14), and bimekizumab (approved by the EMA for SpA in 2023).

Our results are consistent with study results from other disciplines, in which ChatGPT produced partly presentable results with regard to therapy decisions in e.g. (breast) cancer [10, 26] and urological diseases [27], and partly made inadequate, inaccurate and dangerous therapy decisions. Large performance variations were previously observed between different LLMs and according to the individual cases [28], in line with our results. Our study demonstrates that ChatGPT is still inadequate for routine use in therapy decision-making in rheumatology and cannot replace physicians in the foreseeable future. Despite unresolved ethical concerns about accountability, transparency, and health data security, it is impressive how quickly large language models (LLMs) can generate high-quality and safe treatment plans even for complex and rare inflammatory rheumatology cases.

The further development of LLMs and, for example, the addition of retrieval-augmented generation (RAG), which provides the model with additional clinical information, and grounding, which links the model’s responses to real-world data and established knowledge bases [29], could also improve LLMs with regard to treatment decisions in rheumatology [30]. Additional input of multimodal data, such as imaging data could further enhance LLM performance. Truhn et al. previously reported that GPT-4 generated clinically useful orthopedic recommendations solely based on MRI reports [31].

The study has several strengths and limitations. To our knowledge, this is the first study comparing rheumatology treatment concepts generated by LLMs and rheumatologists. The inclusion of various different diseases with varying difficulty levels and clinical pitfalls represent methodological strengths of this study. Furthermore, the multicenter nature of the study represent strengths. A major confounder is that the consistently longer responses from the LLM might have inadvertently unblinded the raters concerning treatment plan source. While more specific prompting including a certain answer length could have been employed, simple and uniform prompts were intentionally chosen to simulate real-world usage, where the LLMs would be used without extensive time and expert knowledge in prompt engineering. Future studies are warranted with a similar answer length to enable better blinding of raters. The board represented a group-based decision, enabling a solid gold-standard for comparison, yet most clinical decisions are seldomly derived in groups. Low inter-rater agreement is a limitation of this study, however this supports the need for decision support. Future studies could investigate whether LLMs might increase overall agreement, treatment confidence, and guideline adherence.

LLMs demonstrated the ability to generate mostly safe and high-quality treatment concepts for various rheumatic diseases. While models like GPT-4 showed promise in creating treatment plans, they frequently fell short of the nuanced, situation-specific decisions made by experienced rheumatologists. Consequently, rheumatologists generally preferred the treatment concepts generated by the RB over those from LLMs, highlighting that LLMs cannot replace rheumatologists in making therapeutic decisions for rheumatic diseases. Future research should aim to enhance LLMs' performance through standardized prompting strategies and investigate the impact of LLM usage on clinical decisions.