Large Language Models in MASLD: The New Era of Generative Artificial Intelligence-Augmented Clinical Practice

Abstract

In a rapidly transforming healthcare landscape, generative artificial intelligence (GenAI) is rapidly emerging as a powerful ally, especially in the management of complex diseases such as metabolic dysfunction-associated steatotic liver disease (MASLD) [1]. At the core of this revolution are large language models (LLMs), such as OpenAI's GPT-4 or Meta's Llama-3, which are showing the potential to reshape how healthcare professionals interact with patients and their data [1, 2] (Figure 1). These tools have unprecedented potential, not only to improve the accuracy of diagnosis, but also to reach patients in underserved areas and provide accessible, timely, and patient-centred support [1, 2]. In particular, the field of digestive diseases has been very prolific in terms of LLM-oriented research, with a recent systematic review defining baseline ChatGPT accuracy ranging from 6.4% to 91.4% when applied to gastroenterology and hepatology queries in the form of simple text-comprehension and response-generation tasks without the use of real-world patient data [3]. However, the recent study by Wu et al. represents a step forward in LLM-oriented clinical research, demonstrating the potential role of OpenAI's GPT (i.e., GPT-3.5, GPT-4, and GPT-4 Vision) in diagnosing MASLD using real-world patient data, including both electronic health records and ultrasound images, and showing comparable diagnostic accuracy to established tools such as the fatty liver index (FLI) and the United States FLI (USFLI) [4]. Therefore, with further refinement, LLMs could quickly and accurately assess patients, ultimately accelerating diagnosis and enabling healthcare providers to make faster, more informed decisions. This potential is further supported by previous validations in machine learning, where algorithms have already demonstrated their capacity to enhance diagnostic processes, streamlining patient assessments and decision-making with impressive accuracy [5]. In addition, we can envision LLMs acting as virtual companions for MASLD patients, guiding them through lifestyle adjustments and answering questions in plain language [6, 7].

In addition to textual interactions, LLMs have demonstrated the ability to process and analyse sophisticated medical data, including images of liver biopsy specimens. A recent exploratory study demonstrated the potential of LLMs for histological diagnosis and fibrosis staging in metabolic dysfunction-associated steatohepatitis (MASH) [8]. In particular, ChatGPT-4 has demonstrated superior performance in the interpretation of histological images, achieving an accuracy of 87.5% compared to 38.3% for Google Bard [8]. The results of this preliminary study suggest that ChatGPT-4, which is accessible even in low- and middle-income countries, could significantly improve cost-effectiveness and resource allocation by aiding in the preliminary staging of MASH, particularly in contexts with limited pathology capabilities.

However, while these advancements bring optimism, the path forward is not without obstacles. It is crucial to remain aware that LLMs are prone to generating plausible but false information (i.e., hallucinations) and that there is a significant risk of patient data leakage of personal health information [2]. The problem of misinformation is of utmost importance in healthcare due to the risk of potential patient’ harm. To address the first issue, the medical community is exploring strategies to link LLM responses to evidence-based medicine through the ingestion of guidelines via retrieval-augmented generation (RAG) or supervised fine-tuning (SFT), both of which have significantly improved the accuracy of LLM responses in the medical field [2, 9]. For the privacy concern, one potential solution is the use of open-source models that can be hosted locally or, alternatively, strengthening and supporting joint ventures between hospital systems and medical companies (such as the collaboration between Epic and OpenAI) [10]. Accessibility remains another significant issue, particularly in rural or underdeveloped regions that may lack the necessary technology or internet connectivity. To overcome these barriers, healthcare systems must invest in education and infrastructure to ensure that these ground-breaking tools are available to all who need them.

These challenges are particularly important to address for conditions such as MASLD, which is anticipated to become increasingly prevalent and for which therapeutic options are limited [1]. Effective management of MASLD relies heavily on lifestyle changes, an area where LLMs could play a crucial role by providing accessible, personalised advice on diet, exercise, and other health behaviours. However, to fully realise this potential, issues of accuracy, privacy, and accessibility need to be addressed. With continued progress in these areas, LLMs could become transformative tools in the management of high-burden, lifestyle-dependent diseases. By providing accurate, secure, and easily accessible support, LLMs offer a path toward a more preventive, patient-centred model of care that empowers both healthcare providers and patients to achieve better long-term outcomes.

As we envision the future, the growing capabilities of GenAI hold the promise to revolutionise patient care, paving the way for unprecedented improvements in outcomes and creating a healthcare system that is more effective, impactful, and deeply patient-centred. While we are only at the beginning of this journey, continued collaboration and innovation will enable AI to play a defining role in the future of healthcare, transforming the way we support and empower both patients and providers.

N.P. declares speaker honorarium from Gilead, AlfaSigma, Advanz Pharma, Gore. J.M.S. declares consultant honorary from Akero, Alentis, Alexion, Altimmune, Astra Zeneca, 89Bio, Bionorica, Boehringer Ingelheim, Gilead Sciences, GSK, HistoIndex, Ipsen, Inventiva Pharma, Madrigal Pharmaceuticals, Kríya Therapeutics, Lilly, MSD Sharp & Dohme GmbH, Nordic Bioscience, Northsea Therapeutics, Novartis, Novo Nordisk, Pfizer, Roche, Sanofi, Siemens Healthineers, Research; speaker honorarium from AbbVie, Boehringer Ingelheim, Gilead Sciences, Ipsen Novo Nordisk, Madrigal Pharmaceuticals, Worldwide Clinical Trials, Stockholder options: Hepta Bio. M.G. declares no COIs.