Our study suggests that GPT-4 can produce valuable treatment recommendations for common knee and shoulder conditions. The recommendations were largely up-to-date, consistent, clinically useful/relevant, and aligned with the most recent clinical and scientific evidence.
We observed signs of reasoning and inference across multiple key findings. For example, GPT-4 correctly deduced that meniscus tears may be associated with bone marrow edema (as a sign of excessive load transmission). Hence, its recommendation to “address focal bone marrow edema: As this issue could be related to the medial meniscus tear […]” was entirely plausible.
Similarly, GPT-4 demonstrated considerable foresight as it recommended organizing post-surgical care and rehabilitation for the patient with multi-ligament knee injuries and imminent surgery. Whether this recommendation can be regarded as “planning” is questionable, though, as true planning abilities in the non-medical domain are still limited4,16. Instead, these recommendations are likely based on the schematic treatment regime that GPT-4 encountered in its training data.
Interestingly, GPT-4 recommended lifestyle modifications, i.e., weight loss and low-impact exercise, and assistive devices (such as braces, canes, or walkers) for shoulder degeneration. While these are sensible and appropriate recommendations for knee osteoarthritis, such recommendations are of doubtful value in shoulder osteoarthritis. In patients with shoulder osteoarthritis or degeneration, exercises to improve the range of motion were not recommended, even though they are indicated22. Again, this observation is likely attributable to the statistical modeling behavior of GPT-4, given the epidemiologic dominance of knee OA over shoulder OA.
Additional limitations of GPT-4 became apparent when the model was tasked to make treatment recommendations for patients with complex conditions or multiple relevant findings.
Critically, the patient with septic arthritis of the knee was not recommended to seek immediate treatment. This particular treatment recommendation, or rather the failure to stress its urgency, is negligent and dangerous. Septic arthritis constitutes a medical emergency, which may lead to irreversible joint destruction, morbidity, and mortality. Literature studies report mortality rates of 4% to 42%23,24,25. Furthermore, because of the stated cartilage damage in this patient, GPT-4 also recommended cartilage resurfacing treatment. However, doing so in a septic joint is contraindicated and medical malpractice26.
GPT-4 was similarly unaware of the patient’s overall situation after knee dislocation. Even though the surgical treatment recommendations for multi-ligament knee injuries were plausible, a potential concomitant popliteal artery injury was not mentioned. It occurs in around 10% of knee dislocations and may dramatically alter treatment2.
Remarkably, we did not find signs of so-called “hallucinations”, i.e., GPT-4 “inventing” facts and confidently stating them. Even though speculative at this stage, the absence of such hallucinations may be due to the substantial and highly specific information provided in the prompt (i.e., the entire MRI report per patient) and our straightforward prompting strategy compared to more suggestive promptings of other studies16.
No patient is treated on the basis of the MR images or the MRI report. Nonetheless, using real-patient (anonymized) MRI reports rather than artificial data, increases our study’s applicability and impact.
However, while GPT-4 offered treatment recommendations, it is crucial to understand that it is not a replacement for professional medical evaluation and management. The accuracy of its recommendations is largely contingent upon the input’s specificity, correctness, and reasoning, which is typically not how a patient would phrase the input and prompt the tool. Therefore, LLMs, including GPT-4, should be used as supplementary resources by healthcare professionals only, as they provide critical oversight and contextual judgment. Optimally, healthcare professionals know a patient’s constitution and circumstances to provide effective, safe, and nuanced diagnostic and treatment decisions. Consequently, we caution against the use of GPT-4 by laypersons for specific treatment suggestions.
Along similar lines, integrating LLMs into clinical practice warrants ethical considerations, particularly regarding medical errors. First and foremost, their use does not obviate the need for professional judgment from healthcare professionals who are ultimately responsible for interpreting the LLM’s output. As with any tool applied in the clinic, LLMs should only assist (rather than replace) healthcare professionals. However, the safe and efficient application of LLMs requires a thorough understanding of their capabilities and limitations. Second, developers must ensure that their LLMs are rigorously tested and validated for clinical use and that potential limitations and errors are communicated, necessitating ongoing performance monitoring. Third, healthcare institutions integrating LLMs into their clinical workflows should establish governance structures and procedures to monitor performance and manage errors. Fourth, the patient (as a potential end-user) must be made aware of the potential for hallucinations and erroneous and potentially harmful advice. Our study highlights the not-so-theoretical occurrence of harmful advice—in that case, we advocate a framework of shared responsibility. The healthcare professional is immediately responsible for patient care if involved in alleged malpractice. Simultaneously, LLM developers and healthcare institutions share an ethical obligation to maximize the benefits of LLMs in medicine while minimizing the potential for harm. While there is no absolute safeguard against medical errors, informed patients make informed decisions—this applies to LLMs as to any other health resource utilized by patients seeking medical advice.
Importantly, LLMs, including GPT-4, are currently not approved as medical devices by regulatory bodies. Therefore, LLMs cannot and should not be used in the clinical routine. However, our study indicates that the capability of LLMs to make complex treatment recommendations should be considered in their regulation.
Moreover, the recent advent of multimodal LLMs such as GPT-4Vision (GPT-4V) has highlighted the (potentially) vast capacities of multimodal LLMs in medicine. In practice, the text prompt (e.g., original MRI report) could be supplemented by select MR images or additional clinical parameters such as laboratory values. Recent literature evidence studying patients in intensive care confirmed that models trained on imaging and non-imaging data outperformed their counterparts trained on only one data type.27 Consequently, future studies are needed to elucidate the potentially enhanced diagnostic performance as well as the concomitant therapeutic implications.
When evaluating the original MRI report (in German) and its translated version (in English), we observed them to be excellently aligned regarding accuracy, consistency, fluency, and context. This finding is confirmed by earlier literature, indicating an excellent quality of GPT-4-based translations, at least for high-resource European languages such as English and German28. Inconsistent taxonomies in MRI reports may be problematic for various natural language processing tasks but did not affect the quality of report translations in this study.
Our study has limitations. First, we studied only a few patients, i.e., ten patients each for the shoulder and knee. Consequently, our investigation is a pilot study with preliminary results and lacks a solid quantitative basis for statistical analyses. Consequently, no statistical analysis was attempted based on our dataset. Second, to enhance its depth and relevance to clinical scenarios, GPT-4’s predictions need to be more specific. Additional ‘fine-tuning’ and domain-specific training using medical datasets, clinical examples, and multimodal data may enhance its robustness and specificity as well as its overall value as a supplementary resource in healthcare. Third, the patient spectrum was broad. A more thorough performance assessment would require substantially more patients with rare conditions and subtle findings to be included. Fourth, treatment recommendations were qualitatively judged by two experienced orthopedic surgeons. Given the excellent level of inter-surgeon agreement, we consider the involvement of two surgeons sufficient, yet involving three or more surgeons could have strengthened the outcome basis even further. Fifth, the tendency of GPT-4 to give generic and unspecific answers and to err on the side of caution rendered it challenging to assess its adherence to guidelines or best practices exactly. Sixth, we used a standardized and straightforward way of prompting GPT-4. After more extensive modifications of these prompts, outcomes may be different.
In summary, common conditions and associated treatment recommendations were well handled by GPT-4, whereas the quality of the treatment recommendations for rare and more complex conditions remains to be studied. Most treatment recommendations provided by GPT-4 were largely consistent with the expectations of the evaluating orthopedic surgeons. The schematic approach used by GPT-4 often aligns well with the typical treatment progression in orthopedic surgery and sports medicine, where conservative treatments are usually attempted first, and surgical intervention is considered only after the failure of conservative treatments.
