Open AI in Transplantation: A Friend or a Foe?

Abstract

Transplantation is a rapidly changing field, with novel technologies and innovations emerging at an unprecedented pace. However, the implementation of new technologies demands careful deliberation, weighing their potential benefits against any unforeseen risks. In recent years, the transplant community has witnessed a rapid growth of artificial intelligence (AI). Chat Generative Pre-Trained Transformer (ChatGPT), a prototype of a large language model (LLM), is one of the most widely used AI products by patients and health professionals. LLM is a class of foundation models trained on vast, publically available data sets, including medical literature, clinical guidelines, patient-centered resources, observational studies, and clinical trials and has the ability to distill and synthesize different forms of information into digestible, understandable, and actionable insights.1 Through Open AI interfaces such as ChatGPT, patients and healthcare professionals can access this information in real time. Improved accessibility to this information can empower transplant recipients to become better informed about their care, enabling a more personalized approach to managing posttransplant complications, medication adherence, preventive measures such as vaccinations and cancer screenings, and enhancing their knowledge of self-management strategies. However, the reliability, accuracy, completeness, and validity of the information provided by these platforms in transplantation have not been appropriately addressed and assessed.2 Effective integration of LLMs in transplantation care requires rigorous validation to ensure the information provided to patients and caregivers is accurate, supported by quality and robust evidence and aligned with current clinical practice and guideline recommendations. Additionally, the key concerns regarding data security and equity in access to this information must also be addressed.3 In this issue of Transplantation, Xu et al4 conducted a survey by 7 health professionals in a single center to assess the accuracy, completeness, and potential harms associated with responses of ChatGPT to patient questions derived from diverse internet sources. A total of 20 questions were retrieved from social media. The questions focused on topics such as immunosuppression, risk of acute rejection, and posttransplant longer-term care. Expert health professionals subsequently ranked their responses using 3-point and 5-point Likert scales. Their responses were benchmarked against several landmark Kidney Disease Improving Global Outcomes guidelines, including the living kidney donors,5 kidney transplant candidates,6 and care of the kidney transplant recipient guidelines.7 Overall, the responses were positive and had minimal harm, with most questions scoring an average of 4 of 5 in the overall rating. However, notable deficiencies were identified. For example, the clarity and completeness of responses regarding sensitive topics, such as sexual and reproductive health or contraceptive use, were insufficient, raising concerns about the potential spread of misinformation. Additionally, health professionals found some inconsistencies in responses of ChatGPT related to posttransplant care, and there were occasional issues with reproducibility and reliability. Although these AI models can help patients to navigate through complex medical systems and equip them with relevant data so that they can process intricate medical evidence and support clinical decision-making, the information provided by these models requires appropriate oversight to ensure the accuracy and validity of the guidance provided. These gaps highlighted the need for caution and further refinement in using LLMs for medical advice, particularly for complex or nuanced topics. More importantly, many of these LLMs are trained using clinical data from high-income countries; therefore, existing LLMs may not be applicable and generalizable in low-resource settings. Also, the integration of the LLM model in low to lower-middle-income countries may pose many challenges due to the lack of technical expertise, digital knowledge and literacy, and infrastructure. Underrepresentation of data and information from diverse populations may further accentuate the health disparities observed in low and lower-middle-income countries. Notwithstanding the limitations of this current study, it serves as a crucial reminder that integrating LLMs into transplantation care requires vigilant oversight, adherence to data privacy regulations such as the 2012 Data Privacy Act, and robust governance to safeguard patient information. To truly harness the potential of AI, we need frameworks that ensure these technologies are not only cutting edge but also uphold the highest standards of data safety, transparency, and inclusivity. Responsible AI in transplantation is not just about advancing technology but also about building trust and ensuring equitable, reliable care for all.

Ähnliche Arbeiten