The role of artificial intelligence language models in dermatology: Opportunities, limitations and ethical considerations

Abstract

Dear Editor, As the fields of artificial intelligence (AI) and machine learning continue to permeate the landscape of medical research and practice, it is vital to appraise the role these technologies can play in transforming patient care.1 In the realm of dermatology, large language models (LLMs) like Chat Generative Pre-Trained Transformer (ChatGPT) have shown the capacity to retrieve, analyse and summarise scientific information for efficient and informed decision-making accurately and instantly. ChatGPT can also personalise information outputs to incorporate patient factors which may facilitate a high quality of patient care.2 Such tools have the potential to offer invaluable assistance, particularly in the context of skin cancers like basal cell carcinoma (BCC) where its commonality makes it a common patient complaint which requires non-urgent excision much of the time. In light of this, we posed seven BCC-related questions to ChatGPT (Figures 1-3 and Figures S1–S4) where its responses were evaluated for accuracy, informativeness, and user-friendliness by two senior plastic surgeons with extensive skin cancer expertise. When asked about the dermatoscopic features of nodular BCC, ChatGPT responded accurately identified the most common features but missed important characteristics such as blue-grey globules and hypopigmented areas.3 The primary concern with ChatGPT, as outlined in Figure 2, pertains to its potential to generate non-existent references. This poses a significant ethical challenge for researchers and calls into question the credibility of its use. When asked about the histopathological features of metastatic risk for BCC, ChatGPT distinguished between high-risk and low-risk subtypes of BCC and correctly identified size >2 cm, perineural invasion and lymphovascular invasion as indicators of metastasis.4 When asked about follow-up for a micronodular, infiltrative BCC excised 0.1 mm from the deep margin, ChatGPT recognised the close margin but failed to acknowledge that infiltrative subtypes and micronodular histopathology are associated with higher recurrence rates.5 As such, it suggested clinical monitoring as opposed to best practice which recommends consideration of re-excision and radiotherapy. When probed about the management of another similar case, ChatGPT stated dependence on high-risk features to guide management, but failed to mention pertinent factors such as perineural invasion and lesion recurrence. Patient scenarios were given for the other three questions. In the fifth scenario, ChatGPT correctly identified BCC as the most likely diagnosis but did not consider other skin cancers such as SCC and atypical melanoma. In the sixth prompt, the response suggested a skin check once a year for a high-risk individual who may require more frequent checks. In the last question, ChatGPT declined to provide direct medical advice and recommended consulting a healthcare professional. These responses highlight the benefits and limitations of using AI tools like ChatGPT for the diagnosis and management of BCC. While ChatGPT generated accurate and pertinent information, its lack of nuanced detail and reference to outdated studies demonstrate the complexity of the principles underlying BCC management. Considering its advice did not adhere to best practice in some instances—for example, recommending prolonged intervals between skin checks for high-risk patient populations—this makes AI an unethical partner in patient care in its current form. Such shortcomings highlight the need for continuous refining of AI models, particularly when management decisions are led by individual patient factors rather than established guidelines. This is crucial, as the ethical and legal implications of using LLMs like ChatGPT in patient care could be enormous if it produced errors with serious consequences. Despite disclaimers stating that it is not a substitute for professional medical advice, its rising popularity and accessibility may encourage its misuse as a healthcare tool among patients. Therefore, it is imperative to regulate and scrutinise AI technologies to ensure its responsible use in the medical field. Despite this, the manner of information retrieval presents a strong advantage over time-consuming manual searching of textbooks and online literature, and customisable responses for parameters like word count and number of references is an attractive attribute for facilitating education at the level of junior medical doctors and educating patients with new diagnoses. Its accessibility for disadvantaged or remote populations where specialist care is unavailable or expensive may prompt the dissemination of high-quality medical information which could motivate populations who would normally reject seeing doctors. If future versions incorporated image analysis of skin cancers alongside information refinement and specific training of dermatological guidelines into the algorithm, these LLMs could have bright futures in dermatology in the long term. For now though, these technologies should be kept at arm's reach of doctors and have a black box warning in the interests of ensuring high-quality patient care in dermatology. There are no acknowledgements to declare. The authors received no financial support for the research, authorship, and/or publication of this article. All authors declare that they have no conflicts of interest. Figure S1 Figure S2 Figure S3 Figure S4 Please note: The publisher is not responsible for the content or functionality of any supporting information supplied by the authors. Any queries (other than missing content) should be directed to the corresponding author for the article.

Ähnliche Arbeiten