Abstract
Artificial intelligence (AI), particularly foundation and generative models, is reshaping the practice of hepatology through enhanced knowledge synthesis, quantitative and reproducible analysis of multimodal data, and personalized clinical decision support. This narrative review examines the transition from task-specific discrimination AI to large language models (LLMs), multimodal foundation models, and agentic AI. We synthesize evidence from original and validation studies, clinical evaluations, and benchmark studies, as well as expert reviews and regulatory frameworks across metabolic dysfunction-associated steatotic liver disease, chronic hepatitis B, cirrhosis and portal hypertension, hepatocellular carcinoma, and liver transplantation. LLMs can convert free-text notes into structured data, summarize longitudinal electronic health records, support patient education, and retrieve guideline-based information. Retrieval-augmented generation and agentic AI may improve traceability and workflow support, but current evidence is largely retrospective or proof-of-concept. In digital pathology and imaging, discriminative AI has enabled more quantitative and reproducible histologic scoring and biomarker analysis. Pathology and multimodal foundation models offer transferable representations, report generation, and cross-modal reasoning, but hepatology-specific validation remains limited. Key risks include hallucination, automation bias, domain shift across centers and devices, and inequities due to under-representation of patient subgroups. We outline the future directions for safe AI model deployment based on multimodal foundation models, prospective and federated evaluation, lifecycle governance, and continuous monitoring for performance, calibration, and equity. Most generative AI applications in hepatology remain at the proof-of-concept stage, and rigorous prospective validation with human-in-the-loop oversight is required before clinical integration.
Keywords: Generative artificial intelligence, Large language models, Supervised machine learning, Deep learning
INTRODUCTION
In hepatology, clinical decision-making requires meticulous use of longitudinal laboratory trajectories, virological markers, imaging, endoscopy, histology, comorbid metabolic risk and medications, and evolving guidelines to manage metabolic dysfunction-associated steatotic liver disease (MASLD), chronic hepatitis B virus (HBV), cirrhosis and portal hypertension, hepatocellular carcinoma (HCC), and liver transplantation. Increasing multimorbidity and expanding indications for non‑invasive tests, surveillance, and therapeutic options make individualized care more complex, particularly across multidisciplinary teams [1]. Hepatology continues to face important unmet needs, including earlier detection of advanced fibrosis, scalable risk‑stratification tools, more reproducible clinical trial endpoints, and effective patient education, all within a setting of limited workforce and resources.
Generative artificial intelligence (AI) has emerged at a time when both data volume and clinical expectations are increasing. Breakthroughs in transformer architectures, self‑supervised pretraining, and scalable computation have produced state-of-the-art foundation models. Foundation models are large-scale pretrained neural networks that acquire general-purpose representations and can subsequently be adapted to downstream applications with minimal task-specific supervision via fine-tuning or prompting [2]. In particular, large language models (LLMs) can transform free‑text notes into structured variables, summarize longitudinal records, and link patient presentation to evidence and guidelines via retrieval-augmented generation (RAG). Vision foundation models move beyond manual feature pipelines toward scalable, generalizable encoders that support quantitative histology, biomarker discovery, and multimodal reasoning across slides and reports.
Earlier discriminative AI has shown strong performance in focused hepatology tasks, such as predicting fibrosis stage or HCC risk [3-6]. These models are usually developed for narrowly defined tasks, depend on task-specific labels, and often show limited generalizability across clinical settings, imaging platforms, pathology workflows, and patient populations [7]. Generative AI and foundation models offer a broader framework for integrating longitudinal, multimodal, and often fragmented data in hepatology. Through self-supervised pretraining on large-scale heterogeneous datasets, generative models enable more flexible learning across multimodal clinical inputs and may reduce key limitations of discriminative AI, including their reliance on expert-labelled data and narrow, task-specific representations. Still, broader pretraining does not guarantee generalizability, and performance must be externally validated across clinical settings, data platforms, workflows, and patient populations before implementation. At the same time, the complexity and high-stakes nature of hepatology expose key limitations of generative AI, including hallucination, inherent bias from training corpora, domain shift, and poor traceability of unsupported recommendations. Therefore, the outputs of generative models must be interpretable, traceable to evidence, and validated across diverse patient populations before clinical implementation.
This review provides a focused summary of AI and generative models in hepatology, moving beyond prior reviews of conventional discriminative models to examine the emerging paradigm of generative models. Discriminative AI is discussed, illustrating the transition from task-specific prediction models to foundation models and generative AI approaches. By integrating applications and highlighting evolving frameworks such as RAG and AI agents, we aim to contextualize generative AI within real-world hepatology workflows. This narrative review synthesizes evidence across MASLD/metabolic dysfunction-associated steatohepatitis (MASH), chronic HBV, cirrhosis and portal hypertension, HCC, and liver transplantation, highlighting current applications and remaining gaps in AI and generative models. We further emphasize the unique challenges of hepatology and outline key requirements for translating rapid technical advances into safe, equitable, and clinically meaningful impact.
FUNDAMENTALS OF GENERATIVE AI AND FOUNDATION MODELS
Unlike discriminative AI, which predicts labels for existing data, generative AI models the underlying probability distribution of data itself, synthesizing coherent outputs that follow the original data distributions. Table 1 shows the glossary of key terms in generative AI. Especially, LLMs exhibit emergent intelligence benefits from scaling laws, unlocking new levels of performance and generalization in clinical applications [8].
Table 1.
Glossary of key terms in generative AI
| Key term | Explanation |
|---|---|
| Agentic AI | A design framework where AI systems exhibit goal-directed autonomy, coordinating reasoning, planning, tool execution, and multi-step workflows to accomplish complex tasks with minimal human supervision, such as reviewing longitudinal liver tests, retrieving clinical guidance, checking elastography or imaging results, and drafting a management summary for clinician review. |
| AI agent | A software system that combines an AI model with tools, memory, and planning to act on hepatology tasks, for example extracting cirrhosis complications from the EHR, ordering relevant guideline passages through retrieval, calculating MELD-Na or fibrosis scores, and flagging missing HCC surveillance while keeping a human clinician in the loop. |
| Chain-of-thought | A prompting technique that asks an LLM to reason through intermediate clinical steps before producing an answer, mimicking human problem-solving to improve accuracy on complex reasoning tasks. In clinical use, the reasoning should be summarized and evidence-linked rather than exposed as unsupported hidden logic. |
| Diffusion model | A generative model that creates images by iteratively denoising random noise. In hepatology, diffusion models may support synthetic liver imaging, histology augmentation, lesion segmentation, or privacy-preserving data generation, but require careful validation to avoid unrealistic pathology or biased imaging features. |
| Few-shot learning | Adapting model behavior by giving a few examples in the prompt, such as examples of how to extract ascites, variceal bleeding, encephalopathy, antiviral therapy, or MASLD risk factors from hepatology clinic notes without additional model training. |
| Fine-tuning | Further training a pretrained model on hepatology-specific datasets, such as liver clinic notes, radiology reports, pathology reports, transplant records, or guideline question-answer pairs, to improve performance on specialized tasks. |
| Foundation model | A large-scale pretrained model that can be adapted to many hepatology applications, including LLMs for clinical notes and guidelines, vision models for ultrasound/CT/MRI or digital pathology, and multimodal models linking reports, laboratory trends, imaging, and histology. |
| Generative AI | A subset of AI focus on generating new content, such as text, images or synthetic data, by learning patterns from existing data. |
| Hallucination | The generation of plausible but factually incorrect, inconsistent, or unsupported content by a generative model, such as inventing a guideline recommendation, misstating HCC surveillance intervals, or citing non-existent evidence; a critical challenge for reliability in clinical applications. |
| LLM | A deep learning model trained on massive text datasets to understand and generate human-like language, capable of tasks like answering questions, writing and reasoning. |
| Multimodal AI | Models capable of processing and generating content across multiple data modalities (e.g., text, images, audio) within a unified framework. |
| Prompt | The input text or instruction given to a generative model to guide its output; effective prompting is key to obtaining desired results. |
| Prompt engineering | The practice of designing and refining prompts to optimize AI responses for accuracy, relevance, and creativity. |
| RAG | A technique that enhances LLM responses by fetching relevant external information from a knowledge base before generation. In hepatology, RAG can connect patient-specific questions to up-to-date guidelines and institutional protocols to reduce unsupported responses. |
| Reasoning model | A specialized AI model optimized for complex, multi-step problem-solving. Unlike standard LLMs, it uses internal chain-of-thought processes to plan and self-correct before answering. |
| Reinforcement learning from human feedback | A training approach in which human preferences guide model behavior. For hepatology, feedback from hepatologists, transplant clinicians, radiologists, pathologists, nurses, and patients could help align outputs with clinical accuracy, safety, clarity, and patient-centered communication. |
| Transformer architecture | A neural network using self-attention mechanisms to process sequential data in parallel, forming the foundation of most modern LLMs and generative models. |
| VLM | A type of multimodal AI that jointly processes images and text. In hepatology, VLMs may connect ultrasound, CT, MRI, endoscopy, or digital pathology images with radiology/pathology reports and clinical context for tasks such as lesion description, steatosis or fibrosis assessment, and HCC workup support. |
| Zero-shot learning | The ability of a model to perform a task for which it was not explicitly trained, relying solely on its general knowledge and the provided prompt. |
AI, artificial intelligence; CT, computed tomography; EHR, electronic health records; HCC, hepatocellular carcinoma; LLM, large language model; MASLD, metabolic dysfunction-associated steatotic liver disease; MELD-Na, model for end-stage liver diseasesodium; MRI, magnetic resonance imaging; RAG, retrieval-augmented generation; VLM, vision-language model.
Generative AI development in medicine transforms general-purpose knowledge into specialized, clinically actionable intelligence (Fig. 1). This pipeline involves pretraining foundation models on massive datasets, domain adaptation through fine-tuning, and reinforcement learning from human feedback, downstream tailoring using techniques such as prompt engineering or RAG [9-11]. Lastly, embedding the generative models within an agentic AI framework enables autonomous, goal-directed clinical reasoning, allowing AI agents to plan, execute, and iterate over complex tasks by interacting with external tools and data sources [12].
Figure 1.

The proposed pipeline of generative AI development in medicine. The primary technical schema of generative AI development in medicine is designed to transform broad, general-purpose knowledge into specialized, clinically actionable intelligence. This pipeline involves pretraining foundation models to establish general abilities, domain adaptation through fine-tuning, downstream deployment, and grounding using techniques such as prompt engineering or RAG, and embedding the generative models within an agentic AI framework. AI, artificial intelligence; APIs, application programming interfaces; CLIP, contrastive language-image pretraining; CoT, chain-ofthought; DICOM, digital imaging and communications in medicine; EHRs, electronic health records; GPT, generative pre-trained transformer; MAE, masked autoencoders; MoE, mixture-of-experts; RAG, retrieval-augmented generation; RLHF, reinforcement learning from human feedback; SFT, supervised fine-tuning.
LLMs built upon the transformer architecture use self-attention to capture long-range dependencies across input sequences [13]. Reasoning models, including GPT-5 (OpenAI, San Francisco, CA, USA), DeepSeek-R1 (DeepSeek, Hangzhou, Zhejiang, China), and Gemini 3 (Google LLC, Mountain View, CA, USA), exhibit multi-step inference and complex instruction following [14,15], making them a central component of agentic AI systems. Figure 2 shows the coevolution of general LLMs and domain-specific medical models. Benchmark evaluations reported that general-purpose LLM, GPT-4, achieved over 90% accuracy on MedQA benchmark derived from the United States medical licensing examination [16]. Continual pretraining on large-scale medical domain corpora developed medical foundation LLMs, such as Me-LLaMA and Med-PaLM 2 [17,18]. Alternatively, some medical foundation models are trained from scratch on massive biomedical text to more deeply embed medical expertise (e.g., PubMedBERT, NYUTron) [19,20]. These medical foundation LLMs outperformed general-purpose LLMs on domain tasks, matching expert-level human performance on task-specific datasets [21].
Figure 2.

Evolution of general LLMs and domain‑specific medical models. The timeline illustrates the co‑evolution of general‑purpose LLMs (lower branch) and domain‑specific medical models (upper branch), starting from the Transformer architecture. BERT, bidirectional encoder representations from transformers; GPT, generative pre-trained transformer; LLM, large language model; MedQA, medical question answering; OSCE, objective structured clinical examination; PCPs, primary care physicians; USMLE, United States medical licensing examination.
Vision-language models (VLMs) jointly process and generate visual and textual clinical information, comprising a vision encoder for extracting image features, an LLM decoder for text generation, and a cross-modal alignment module. They support image captioning, visual question answering, and text-to-image generation [22]. An evolution from discriminative to generative AI in digital pathology has been observed. Targeted at narrowly specific tasks, such as tumor detection, grading, and tissue segmentation, deep learning models (convolutional neural networks [CNNs], and vision transformers) trained using supervised learning strategy have achieved high accuracy [23,24]. However, these models lack generalizability to different diseases or tissue types, and the training data commonly requires extensive manual annotation. In contrast, self-supervised pretraining on large-scale, unlabeled whole-slide images has produced pathology foundation models, such as UNI and Virchow2, which learned transferable visual representations across diverse tissue types, staining protocols, and scanning platforms, and showed strong performance across a variety of downstream retrospective tasks and benchmarks [25]. Moreover, multimodal pathology foundation models align whole-slide images with clinical text and pathology reports, enabling conversational cross-modal interaction, for example, text-prompt slides retrieval, abnormality localization, and pathology report generation [26,27].
Although foundation models have demonstrated broad capabilities in text generation, image analysis, and multimodal reasoning, most supporting evidence remains benchmark-based, retrospective, or derived from general biomedical tasks. Hepatology-specific evidence is still limited, particularly for prospective, real-world validation and clinical outcomes assessment.
AI AGENT AND CLINICAL DECISION SUPPORT SYSTEM
Clinical decision support systems (CDSS) bridge the clinical observations (e.g., electronic health records, EHRs) and clinical knowledge at the point of care. Powered by LLMs, particularly reasoning models, AI agents may drive a transformative advance in clinical decision support [28,29]. Figure 3 shows a proposed workflow of generative AI in hepatology and applications across the spectrum of liver diseases.
Figure 3.

Generative AI applications across the spectrum of liver diseases. The pipeline in blue illustrates the proposed general workflow: embedding knowledge from multimodal data into liver-specific foundation models and building agentic systems that execute specific clinical actions. The green boxes depict the current or preliminary applications across the spectrum of liver diseases. AI, artificial intelligence; CT, computed tomography; DICOM, digital imaging and communications in medicine; EHR, electronic health record; HBV, hepatitis B virus; HCC, hepatocellular carcinoma; LLMs, large language models; MASLD, metabolic dysfunction-associated steatotic liver disease; MASH, metabolic dysfunction-associated steatohepatitis; MRI-PDFF, magnetic resonance imaging proton density fat fraction; VLMs, vision-language models.
AI agents can be characterized by the following key features [12]. First, environmental perception; through natural language processing (NLP) and application programming interface calls, the agent can understand and interpret its operating environment, including user inputs, multimodal clinical data, dynamic data streams (e.g., physiological monitoring), and outputs from external tools. Second, reasoning and planning; under user-defined clinical objectives, LLMs synthesize patient data with medical knowledge and up-to-date evidence-based guidelines and perform multilevel clinical reasoning and task planning. Third, action; the agent can automatically use tools and be deeply engaged in clinical workflows, such as triggering risk alerts, generating diagnoses and treatment recommendations, scheduling examinations, and dynamically updating follow-up plans. Lastly, reflection and iteration; utilizing real-world evidence, for example, adoption rates of clinical recommendations and patient outcomes, as well as reinforcement learning strategy, the agent can dynamically refine its inherent decision-making logic and continuously calibrate and evolve itself during clinical practice. Consequently, AI agents may extend traditional rule-based CDSS by supporting more flexible information synthesis, guideline retrieval, task coordination, and clinical workflow assistance.
Early proof-of-concept studies reported promising performance of AI agents in selected clinical tasks [30]. In an oncology setting, Ferber et al. [31] developed an AI agent powered by GPT-4, autonomously orchestrated segmentation models, pathology foundation models, external tools (calculators and web search), and RAG to generate evidence-based evaluations supporting clinical decision-making. Among 20 gastrointestinal oncology cases, including HCC, the agent used tools with 87.5% accuracy, and generated correct clinical conclusions in 91.0% of cases. These performances should be interpreted cautiously, as they were derived from only 20 curated cases without external validation.
LLM-based AI agents demonstrated promising benefits for patient education and self-management by supporting care instructions [32]. Studies reported that LLM-based AI agents accurately answer patients’ questions over 90% of the time, covering complications, treatment, and common inquiries [33,34]. Another study found that 4 LLMs outperformed web-based forum users in answering laboratory test-related questions [35]. Among cancer patients, LLMs can also support their self-management through emotional, informational, and motivational assistance [36].
Several barriers hinder real‑world clinical use of AI agents. First, existing evaluation benchmarks and metrics emphasize classification accuracy and automated technical metrics, offering limited insight into the performance of generative outputs on clinical tasks [37,38]. Evaluation frameworks tailored to generative outputs remain underdeveloped for medicine and are even scarcer in hepatology [39]. Second, the inherent hallucination is another concern, wherein LLMs can generate plausible but factually incorrect content [40]. This affects different aspects, including but not limited to fabricated references, incorrect staging or treatment recommendations, and automation bias. A recent reference-integrity audit of 2.5 million biomedical papers found that the fabrication rate was more than 12 times greater in 2025 than that in 2023 [41]. A study of 104 prompts showed that the rate of hallucinated treatment recommendations from ChatGPT-3.5 ranged from 5% to 25% in different cancer types [42]. The risk of hallucination is amplified by automation bias, in which clinicians defer to confident but erroneous machine outputs. A randomized clinical trial introducing deliberate errors in ChatGPT-4o’s suggestions for physicians trained in AI literacy showed a reduction in accuracy of 14% compared to the error-free suggestion group, suggesting substantial automation bias when exposed to erroneous LLM recommendations [43]. Mitigation strategies are required, including mandatory citation verification, human-in-the-loop confirmation for high-risk recommendations, and prospective post-deployment safety monitoring. Third, like discriminative AI, generative AI is also susceptible to bias from its training data, including the underrepresentation of minority populations and rare diseases, undermining reliability and equity. Understanding sources of AI bias in clinical scenarios enables a proactive approach to mitigate its effect [44]. All these aspects pose a critical safety concern given the high-stakes nature of medical decision-making in hepatology. However, early evidence indicates that domain‑specific fine‑tuning and RAG grounded in hepatology guidelines can improve factuality and task performance, suggesting a practical pathway toward safer deployment [18,45].
RAG is particularly relevant to hepatology because many clinical decisions depend on rapidly evolving guidelines, local protocols, and patient-specific longitudinal data. In a RAG workflow, an LLM retrieves relevant information from trusted sources, including society clinical guidelines, institutional HCC surveillance protocols, transplant program manuals, or prior similar cases, and generates evidence-based responses. Exploratory studies proposed preliminary LLM frameworks in liver diseases using RAG and guidelines. Ge et al. [45] developed LiVersa, an LLM using RAG and guidelines from the American Association for the Study of the Liver, while Kresevic et al. [9] integrated hepatitis C virus (HCV) guidelines using RAG into CDSS to support treatment recommendations for HCV patients. However, these performances were reported based on a few expert-drafted clinical cases, and more rigorous evaluations are required. In a hepatology RAG system, curated knowledge bases, version control of guidelines, and source citation are important to ensure the accuracy and usefulness in practice.
APPLICATIONS IN MASLD/MASH
AI‑assisted scoring of liver histology in MASH
Traditionally, pathologists have scored various features of liver histology on a semiquantitative scale to facilitate comparisons across studies and the interpretation of clinical trials [46]. This approach has gained prominence with drug development for MASH, as both the U.S. Food and Drug Administration (FDA) and the European Medicines Agency (EMA) have accepted histological responses, specifically the resolution of MASH and improvements in fibrosis, at interim analyses as the basis for considering a drug for accelerated approval [47].
To demonstrate these histological responses, clinical trials need to enroll a narrow spectrum of patients with “at-risk” MASH, defined as the presence of MASH (with a MASLD activity score of 4 or higher, including at least 1 point each in steatosis, lobular inflammation, and hepatocyte ballooning) alongside F2–F3 fibrosis. These requirements have created substantial challenges in drug development. One key issue is the very low intrarater and interrater agreement in histological assessments, which is particularly poor for ballooning—unfortunately, a defining feature of MASH. Consequently, two pathologists may disagree on whether a patient is eligible for a clinical trial or whether they have achieved a histological response. In a secondary analysis of the EMMINENCE study, 46% of patients deemed eligible by one pathologist were considered not eligible by one or more other pathologists [48]. Importantly, when two serial biopsies were evaluated, the interrater variability was compounded, resulting in kappa values as low as around 0.4 for both MASH resolution and fibrosis improvement.
To mitigate this problem, current clinical trials typically involve histological assessments conducted by at least two pathologists, with a third pathologist consulted in the case of discrepant results. Although this approach improves consistency to some extent, the need for histological evaluations by multiple pathologists and the efforts involved in resolving discrepancies further increase the cost and time required for drug development.
Imaging analysis is at the forefront of AI applications in medicine. Several platforms, including PathAI, Histoindex, FibroNest, and MorphoQuant, have already demonstrated superior reproducibility and accuracy in scoring histology compared to human pathologists [49]. These platforms utilize scanned images of stained slides as well as second harmonic generation/two-photon excitation on unstained slides. The latter approach can theoretically further reduce variability introduced during staining; however, head-to-head comparisons among these platforms are largely lacking.
Regulatory qualification of AIM-MASH
In 2025, AIM-MASH (formerly AIM-NASH by PathAI) became the first AI tool qualified by the FDA under the Drug Development Tool Qualification programmes and EMA under the Qualification Options of novel methodologies for medicine development to evaluate the histological severity of MASH in clinical trials, based on secondary analyses of clinical trial data [23,50-53]. AIM-MASH demonstrated comparable accuracy to consensus scoring by human pathologists for histological activity and fibrosis stage, while showing significantly higher consistency. Unlike the ordinal Nonalcoholic Steatohepatitis Clinical Research Network scoring system, AIM-MASH scores each histological feature on a continuous scale, providing more granular data that highlights changes often missed by the crude ordinal scale.
Despite this exciting development, several key questions remain. Firstly, no studies have yet demonstrated that AI combined with human consensus scoring is superior to AI alone. It raises the question of whether clinical trials still require liver biopsies to be evaluated by multiple pathologists. If human pathologists do not enhance scoring accuracy, their role may shift to excluding unexpected concomitant liver diseases, a task that AI could potentially also handle in the future.
Secondly, while model refinement is common in AI, regulatory approval typically relies on a “finalized” model. Regulators have indicated that major changes could necessitate further qualification, but defining what constitutes a “major change” is still pending.
Thirdly, the field should not be dominated by a single model. Multiple AI digital pathology platforms will likely emerge, making cross-platform comparisons essential. Moreover, while AI can detect smaller changes in histology [54], defining what constitutes clinically meaningful improvement remains critical.
Finally, since current AI platforms are trained against the imperfect reference of human histological scoring, pursuing unbiased model training using clinical outcomes as the reference standard may yield more reliable tools [55,56].
Physics-aware imaging AI in MASLD/MASH
Although biopsy remains a histological gold standard, its invasive nature and susceptibility to sampling error preclude its routine use in all patients with MASLD/MASH. Non-invasive imaging remains of critical clinical importance. A growing body of literature now explores the utility of physics-aware imaging AI mapping quantitative biomarkers to enhance these modalities. The main strength of this physics-aware AI paradigm is that known physical principles, such as MRI signal generation, X-ray scattering simulations, and energy conservation laws, can be incorporated directly into AI, enhancing models’ scientific consistency, generalization, and interpretability, as well as their superior performance in terms of resolution and speed [57].
In ultrasonography (US), applications diverge between physics-aware quantitative ultrasound (QUS) and B-mode image analysis. Vianna et al. [58] applied a deep learning algorithm to B-mode US to detect and grade steatosis, achieving performance comparable to radiologists. Similarly, AI-enhanced QUS techniques demonstrated high agreement with magnetic resonance imaging proton density fat fraction (MRI-PDFF) [59,60], indicating an expanding role for this primary, accessible modality in the assessment of MASLD. Computed tomography (CT) and MRI have achieved methodological maturity. Utilizing hepatosplenic attenuation derived from deep learning, a contrast-enhanced volumetric CT tool displayed satisfactory performance in categorized assessment of hepatic steatosis [61], showing its potential in opportunistic steatosis assessment. CT and MRI employ deep learning and radiomic methods for automated organ segmentation, which facilitates hepatic volumetry and improves reproducibility of biomarkers [62]. The application of AI-facilitated MRI for PDFF and R2* quantification has also been validated, with a strong correlation with magnetic resonance spectroscopy [63], confirming its robustness for hepatic fat and iron quantification. However, before clinical translation, this technology warrants external validation across different devices and populations, as well as refined harmonization of acquisition parameters.
APPLICATIONS IN CHRONIC HBV INFECTION
The management of chronic HBV infection draws on serial laboratory parameters, medication history, diagnostic and procedural codes, and family and psychosocial history [64]. While a significant portion of this data is stored as structured datasets, some clinical information is documented as freetext medical notes. LLMs, exemplified by ChatGPT-3.5, have been shown to outperform two conventional NLP methods [65] in extracting pathological classifications from unstructured pathology notes with 89% overall accuracy [66]. Such promising findings support the feasibility of using LLMs to transform unstructured healthcare data into structured formats, enhancing data analytics and research capability, ultimately contributing to improved risk prediction, informed clinical decision-making, and effective allocation of healthcare resources [20]. LLMs tailored for medical language can be fine-tuned for a variety of clinical and operational predictive tasks. These include, but are not limited to, significant clinical outcomes such as hepatic events, HCC, mortality, readmission rates, and length of hospital stay.
LLMs may also improve the efficiency of clinical operations by enabling clinical text summarization and timely delivery of relevant information to the appropriate healthcare providers within clinical care workflow. Clinical text summarization derived from EHRs facilitates the streamlining of various tasks, including radiology reports, patient inquiries, progress notes, and doctor–patient dialogues. Summaries generated by LLMs are comparable to, and in some cases exceeding, the quality of summaries produced by medical experts, without introducing an increased risk of summarization errors that could potentially result in adverse medical outcomes [21]. Accurate interpretation and subsequent integration of clinical guidelines through LLMs into clinical management systems enhances the effectiveness of executing guidelines in daily clinical practice [9].
To date, HBV-specific generative AI evidence remains sparse. Important clinical questions in chronic HBV infection, including HCC and decompensation prediction, interpretation of quantitative serum HBsAg, HBV DNA, or HBV RNA trajectories, antiviral treatment monitoring, and reactivation risk during immunosuppression, still rely mainly on conventional risk scores or discriminative machine learning based on structured EHR. For generative AI to add value in HBV care, future studies should integrate laboratory trajectories, antiviral exposure, family history, imaging, fibrosis assessment, and guidelines, and should be externally validated in Asian and other HBV-endemic populations.
APPLICATIONS IN CIRRHOSIS AND PORTAL HYPERTENSION
Multiple studies focused on the adoption of traditional discriminative AI for the diagnosis of cirrhosis and portal hypertension, which mitigates the need for invasive procedures such as liver biopsy and hepatic venous pressure gradient measurement in making the diagnosis. Using multimodal parameters, including clinical and radiomics data, enables the identification of gastroesophageal varices and clinically significant portal hypertension, as well as the prediction of hepatic decompensation and liver-related mortality [67,68]. Generative AI in this area is relatively confined to the diagnosis of cirrhosis. Several retrospective or exploratory studies have examined the use of LLMs in cirrhosis-related tasks. One retrospective study evaluated GPT-4 for determining whether cirrhosis and its complications were the cause of hospitalization, and the results suggested that GPT-4 with zero-shot prompts had a robust accuracy of 88–99% for the identification of cirrhosis and its complications, with chart review being considered the gold standard [69]. This enhanced the identification of appropriate cohorts from purely diagnosis code-based classification and potentially served as a standard where a labor-intensive chart review may be unnecessary. Beyond cohort identification, another comparative evaluation study assessed four LLMs (ChatGPT-4o, Claude-3.5 Sonnet, Gemini 1.5 Pro, and Llama-3.1) in responding to questions related to cirrhosis, and found that all models were comprehensive with the generated answers, and that they carried self-correction ability to improve the accuracy of initial answers with simple prompts [70]. Though an earlier exploratory study showed that the performance of LLMs (using ChatGPT-3.5 as the tool) may differ in answering different domains of questions, such as poorer performance in preventive medicine and specifying certain treatment durations [71]. Nonetheless, LLMs may help identify patients with cirrhosis and its complications for better risk stratification in clinical practice and cohort identification in research. The role of LLMs in developing patient educational materials on cirrhosis and portal hypertension has also been shown, and this may assist in the clinical care for patients with cirrhosis [72]. Further studies focusing on acute-on-chronic liver failure risk prediction and multimodal prediction of decompensation, particularly through the integration of hemodynamic measurements, radiomics, and clinical data, will be required to advance the use of generative AI in cirrhosis and portal hypertension.
Generative AI and drug discovery for liver fibrosis
Drug development in hepatology, particularly for liver fibrosis, has been hindered by complex disease biology, limited translational models, and historically high clinical-trial failure rates. Heterogeneous metabolic, inflammatory, and regenerative pathways also complicate target identification and validation.
LLM-driven “co-scientist” systems can synthesize biomedical literature, integrate multimodal biological data, and generate mechanistic hypotheses at scale. They can iteratively propose, refine, and prioritize therapeutic targets and identify opportunities for drug repurposing [73]. Hepatology provided early evidence of biological validation. A Gemini-based multi-agent co-scientist system identified epigenetic targets that were assessed in multi-lineage human hepatic organoids, where they suppressed fibrosis and promoted liver regeneration [74]. This illustrates an early closed-loop workflow integrating AI-generated hypotheses with experimental testing.
AI-generated hypotheses still require rigorous experimental and clinical validation, with safeguards against bias, poor reproducibility, and overinterpretation of associative findings. Future work should focus on integrating generative AI with high-quality biological models and prospective validation frameworks to translate these advances into clinically effective therapies.
APPLICATIONS IN HCC
Digital pathology of HCC has been studied, spanning from lesion diagnosis and characterization to biomarker prediction and prognostication [75]. Using different AI modalities, such as deep learning on whole-slide images, HCC can be differentiated from other liver pathologies, such as dysplastic nodules or intrahepatic cholangiocarcinoma, with high diagnostic performance [76,77]. Despite heterogeneity in methodologies across studies, most reported an accuracy and area under the receiver operating characteristic curve (AUROC) >80% using pathomics in distinguishing key histological features, such as tumor differentiation and microvascular invasion, which can be subjective and prone to interobserver variability among pathologists [78-81]. With available data from resectable to unresectable HCC receiving immunotherapy, the ability of pathomics alone or in combination with clinical parameters to predict therapeutic response and recurrence has also been suggested, though the accuracy and AUROC (0.6–0.9) vary across studies [82-85]. Moreover, using pathomics to build CNNs allows identification of genetic mutations from histopathology slides, facilitating future development in precision oncology [79,86].
Generative multimodal models, such as VLMs, have also been explored. A multicenter retrospective study in China utilized preoperative radiological imaging to develop Hepa-PathGPT, which exhibited robust tumor segmentation (mean Intersection over Union 0.883±0.007; Dice 0.934±0.006) and accuracy of 0.697±0.024 for six pathological markers. It allowed generation of structured pathology reports to provide a real-time, non-invasive pathological prediction [85].
Generative AI may also support clinical decision-making for HCC, but careful validation, workflow integration, and safeguards against hallucinations and bias are essential [87]. Yang et al. [88] compared LLM-based treatment recommendations for untreated HCC with physician decisions and guideline-based strategies in real-world registry data. This study reported low overall concordance (26.8–32.7%). Concordance was associated with better survival only in early-stage HCC and with worse survival in advanced-stage disease, suggesting that LLMs should remain supervised adjuncts for relatively straightforward, guideline-concordant scenarios.
APPLICATIONS IN LIVER TRANSPLANTATION
Generative AI solutions have been developed to enable personalized clinical decision support along the liver transplant journey, from pre- to post-transplant [89]. Important steps along this journey include referral for transplant, evaluation for transplantation, discussions around candidacy, donor-recipient matching, and post-transplant care. Generative AI solutions have been proposed to help both transplant clinicians and patients along their journey.
ChatGPT has been examined for patient questions during the transplant journey, which is paved by significant uncertainty and anxiety [90]. Among 1,281 AI-generated responses to patient inquiries, although 95.3% of patient users reported a high degree of satisfaction, response accuracy was only 80%, suggesting the necessity of RAG grounded in guidelines and local transplant program manuals [91]. A LLM framework was developed by Robitschek et al. [92] to extract social determinants of health and psychosocial risk factors contributing to disparities in access to liver transplantation. These factors collectively accounted for 94.6% of the differences in access to transplant seen.
Transplantation candidacy decisions can be inconsistent, affected by recency bias, and swayed by the loudest voice in the room. Hasjim et al. [93] developed a multi-agentic framework based on a national retrospective cohort, with each agent representing different experts on the committee (AI hepatologist, AI transplant surgeon, AI social worker, and AI cardiologist) reviewing over 8,400 candidates for liver transplantation and arriving at a consensus decision around acceptance versus decline for transplant. This decision was accompanied by a rationale, and using cosine similarity analysis for each AI agent, they were able to delineate the factors that impacted each agent’s decision. The consensus decision of the multi-agentic committee was 92% accurate in predicting patients who would derive a 1-year survival post-transplant, representing the appropriateness of patient acceptance for transplantation. Although this work is still a proof of concept, it paves the way for future implementation of a multi-agentic framework for more objective decision-making along the transplant patient journey.
FUTURE DIRECTIONS: MULTIMODAL, FEDERATED, AND REAL‑WORLD EVALUATION
The more decisive issue in generative AI in hepatology is the data reality. Clinical decisions integrate EHR history, laboratory trajectories, ultrasound, CT or MRI, endoscopy, histology, and sometimes omics. Yet complete cross-modality datasets are uncommon. Endoscopy and ultrasound data are inconsistently archived, and radiology protocols vary across centers. Models trained on narrow or single-center data may therefore be fragile [94].
Multimodal foundation models offer a plausible solution. Pretraining on large, partially observed datasets can learn reusable representations that support several downstream tasks. In hepatology, these may include surveillance triage, fibrosis progression prediction, prognosis, treatment response, and trial matching. Reusing fewer core models may also simplify updating and auditing, and improve consistency across related endpoints.
Generalization is limited by cross-center heterogeneity, fragmented infrastructure, population imbalance, scarce benchmarks, and restrictions on centralized data pooling [95]. Federated and swarm learning enable joint training while keeping patient-level data local, but they do not correct biased sampling. In hepatology, representativeness should be assessed against the intended population and use case, accounting for variation in viral etiology, disease stage, ancestry and geography, scanner and staining protocols, and clinical setting. Although pathology foundation models such as UNI and Virchow2 were pretrained on large, heterogeneous datasets, published studies do not establish adequate representation of Asian or HBV-associated populations. This evidence gap is particularly important for HBV-associated HCC because models developed or calibrated in populations with different etiologies and risk profiles may not transfer reliably. Equity should also be prespecified because models developed on biased samples can reduce generalizability and perpetuate disparities [96].
Evaluation has to match the stakes. Retrospective discrimination is not enough for models used across sites and over time. Prospective multicenter validation should assess calibration, clinical utility, human-AI interaction, and risk-of-bias appraisal, as well as patient outcomes, workflow efficiency, and costs [97,98]. Transparent and reproducible reporting should follow the emerging frameworks, including transparent reporting of a multivariable model for individual prognosis or diagnosis (TRIPOD)-LLM for studies developing or evaluating LLMs [99], the standard protocol items: recommendations for interventional trials (SPIRIT)-AI extension, and consolidated standards of reporting trials (CONSORT)-AI for prospective clinical trials of AI interventions [100,101]. Embedding these frameworks would help move hepatology from ad hoc adoption toward accountable, auditable clinical deployment.
Achieving trustworthy, accessible, and well-integrated AI systems also requires lifecycle governance. Performance can change as patient mix, practice patterns, and data quality evolve. Ongoing monitoring is necessary, and concise summaries such as model cards can support clinical oversight. This lifecycle governance should be anchored to existing regulatory instruments. For example, the FDA’s predetermined change control plan framework provides a mechanism to pre-specify permissible modifications and validation [102]; the European Union AI Act classifies AI systems by risk, emphasizes risk management, data governance, transparency, and human oversight, and imposes corresponding legal obligations [103]. Accountability should also remain explicit as AI-enabled decision support disperses influence across vendors, cloud providers, hospitals, regulators, and clinicians, particularly as systems become more autonomous.
CONCLUSIONS
A practical roadmap of generative AI implementation begins by prioritizing use cases that tackle current bottlenecks, including cohort identification, registry automation, decision support aligned with clinical guidelines, reproducible histology endpoints in trials, and scalable surveillance triage. For each use case, the intended user, hepatologist, pathologist, radiologist, or allied clinician, and clinical setting should be explicitly defined, because a tool embedded in a point-of-care workflow demands different design, latency, and explainability than one supporting retrospective review or multidisciplinary tumor board deliberation. The next step is to align data governance, privacy safeguards, and the regulatory context of use with the intended deployment. Systems should be engineered for safety by design. They should use RAG over trusted local sources, communicate uncertainty clearly, restrict permissions for tool use, and enable human‑in‑the‑loop escalation for critical decisions.
Prospective evaluation should assess discrimination, calibration, bias, robustness across centers and devices, workflow effects, patient-reported outcomes, and health‑economic endpoints. Successful deployment also requires multidisciplinary teams that include hepatologists, liver transplant and oncology specialists, pathologists, radiologists, nurses, data scientists, and medical ethicists with hospital information technology and governance. These teams should define accountability and maintain model documentation and post-market surveillance.
Impact and safety should be monitored using pre‑specified metrics, including errors prevented, time saved, equity gaps, and clinical outcomes improvement. Teams should also define “stop rules” for unacceptable failures. With rigorous measurement and shared stewardship, generative AI can augment clinical expertise, reduce unwarranted variation, and improve liver-disease care.
Abbreviations
- AI
artificial intelligence
- AUROC
area under the receiver operating characteristic curve
- CDSS
clinical decision support systems
- CNNs
convolutional neural networks
- CT
computed tomography
- EMA
European Medicines Agency
- EHRs
electronic health records
- FDA
U.S. Food and Drug Administration
- HBV
hepatitis B virus
- HCC
hepatocellular carcinoma
- HCV
hepatitis C virus
- LLMs
large language models
- MASH
metabolic dysfunction-associated steatohepatitis
- MASLD
metabolic dysfunction-associated steatotic liver disease
- MRI-PDFF
magnetic resonance imaging proton density fat fraction
- NLP
natural language processing
- QUS
quantitative ultrasound
- RAG
retrieval-augmented generation
- US
ultrasonography
- VLMs
vision-language models
Footnotes
Authors’ contribution
All authors were responsible for the writing plan, content, drafting, and critical revision of the manuscript for important intellectual content.
Acknowledgements
During the preparation of this manuscript, the authors used ChatGPT (OpenAI GPT-5.5) to assist with language refinement, improve clarity, and presentation. All AI-assisted content was critically reviewed and revised by the authors, who take full responsibility for the accuracy and integrity of the final manuscript.
Conflicts of Interest
Jimmy Che-To Lai has served as an advisory committee member for Gilead Sciences and Boehringer Ingelheim, and a speaker for Gilead Sciences. Vincent Wai-Sun Wong served as a consultant or advisor for AbbVie, AstraZeneca, Boehringer Ingelheim, Echosens, Eli Lilly, Gilead Sciences, Merck, Novartis, Novo Nordisk, Pfizer, Roche Diagnostics, and TARGET PharmaSolutions; a speaker for Abbott, AbbVie, AstraZeneca, Echosens, Gilead Sciences, and Novo Nordisk; an independent non-executive director for Furui; and a co-founder of Illuminatio Medical Technology. He has received a research grant from Gilead Sciences. Grace Lai-Hung Wong has served as an advisory committee member for AstraZeneca, Gilead Sciences, GlaxoSmithKline Pharmaceuticals and Janssen, and as a speaker for Abbott, AbbVie, Ascletis, Bristol-Myers Squibb, Echosens, Gilead Sciences, Janssen, and Roche. She has also received a research grant from Gilead Sciences. Mamatha Bhat has received investigator-initiated grants from Novo Nordisk, Merck, Paladin, Roche, Eisai, Natera, and Astra- Zeneca. Terry Cheuk-Fung Yip has served as an advisory committee member for Altimmune, Gilead Sciences, and GSK, and as a speaker for Echosens and Gilead Sciences. The other authors declare that they have no competing interests.
REFERENCES
- 1.Friedman SL, Sanyal AJ. The future of hepatology. Hepatology. 2023;78:637–648. doi: 10.1097/HEP.0000000000000389. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 2.Teo ZL, Thirunavukarasu AJ, Elangovan K, Cheng H, Moova P, Soetikno B, et al. Generative artificial intelligence in medicine. Nat Med. 2025;31:3270–3282. doi: 10.1038/s41591-025-03983-2. [DOI] [PubMed] [Google Scholar]
- 3.Peng N, Song SJ, Hui VW, Lai JC, Wong GL, Wong VW, et al. Foundations of artificial intelligence in hepatology: what a clinician needs to know. J Clin Exp Hepatol. 2026;16:103183. doi: 10.1016/j.jceh.2025.103183. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 4.Anteby R, Klang E, Horesh N, Nachmany I, Shimon O, Barash Y, et al. Deep learning for noninvasive liver fibrosis classification: a systematic review. Liver Int. 2021;41:2269–2278. doi: 10.1111/liv.14966. [DOI] [PubMed] [Google Scholar]
- 5.Rodriguez LA, Schmittdiel JA, Liu L, Macdonald BA, Balasubramanian S, Chai KP, et al. Hepatocellular carcinoma in metabolic dysfunction-associated steatotic liver disease. JAMA Netw Open. 2024;7:e2421019. doi: 10.1001/jamanetworkopen.2024.21019. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 6.Shin H, Hur MH, Song BG, Park SY, Kim GA, Choi G, et al. AI model using CT-based imaging biomarkers to predict hepatocellular carcinoma in patients with chronic hepatitis B. J Hepatol. 2025;82:1080–1088. doi: 10.1016/j.jhep.2024.12.029. [DOI] [PubMed] [Google Scholar]
- 7.Esteva A, Robicquet A, Ramsundar B, Kuleshov V, DePristo M, Chou K, et al. A guide to deep learning in healthcare. Nat Med. 2019;25:24–29. doi: 10.1038/s41591-018-0316-z. [DOI] [PubMed] [Google Scholar]
- 8.Omiye JA, Gui H, Rezaei SJ, Zou J, Daneshjou R. Large language models in medicine: the potentials and pitfalls: a narrative review. Ann Intern Med. 2024;177:210–220. doi: 10.7326/M23-2772. [DOI] [PubMed] [Google Scholar]
- 9.Kresevic S, Giuffrè M, Ajcevic M, Accardo A, Crocè LS, Shung DL. Optimization of hepatological clinical guidelines interpretation by large language models: a retrieval augmented generation-based framework. NPJ Digit Med. 2024;7:102. doi: 10.1038/s41746-024-01091-y. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 10.Ge J, Chen IY, Pletcher MJ, Lai JC. Prompt engineering for generative artificial intelligence in gastroenterology and hepatology. Am J Gastroenterol. 2024;119:1709–1713. doi: 10.14309/ajg.0000000000002689. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 11.Ng KKY, Matsuba I, Zhang PC. RAG in health care: a novel framework for improving communication and decision-making by addressing LLM limitations. NEJM AI. 2025;2:AIra2400380. doi: 10.1056/AIra2400380. [DOI] [PubMed] [Google Scholar]
- 12.Liu F, Niu Y, Zhang Q, Wang K, Dong Z, Wong IN, et al. A foundational architecture for AI agents in healthcare. Cell Rep Med. 2025;6:102374. doi: 10.1016/j.xcrm.2025.102374. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 13.Vaswani A, Shazeer N, Parmar N, Uszkoreit J, Jones L, Gomez NA, et al. Attention is all you need. arXiv 1706.03762 [Preprint] doi: 10.48550/arXiv.1706.03762. 2017 [cited 2026 Jul 8]. Available from: [DOI]
- 14.Guo D, Yang D, Zhang H, Song J, Wang P, Zhu Q, et al. DeepSeek-R1 incentivizes reasoning in LLMs through reinforcement learning. Nature. 2025;645:633–638. doi: 10.1038/s41586-025-09422-z. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 15.Wang S, Hu M, Li Q, Safari M, Yang X. Capabilities of GPT-5 on multimodal medical reasoning. arXiv 2508.08224 [Preprint] doi: 10.48550/arXiv.2508.08224. 2025 [cited 2026 Jul 8]. Available from: [DOI]
- 16.Nori H, Lee YT, Zhang S, Carignan D, Edgar R, Fusi N, et al. Can generalist foundation models outcompete special-purpose tuning? Case study in medicine. arXiv 2311.16452 [Preprint] doi: 10.48550/arXiv.2311.16452. 2023 [cited 2026 Jul 8]. Available from: [DOI]
- 17.Xie Q, Chen Q, Chen A, Peng C, Hu Y, Lin F, et al. Medical foundation large language models for comprehensive text analysis and beyond. NPJ Digit Med. 2025;8:141. doi: 10.1038/s41746-025-01533-1. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 18.Singhal K, Azizi S, Tu T, Mahdavi SS, Wei J, Chung HW, et al. Large language models encode clinical knowledge. Nature. 2023;620:172–180. doi: 10.1038/s41586-023-06291-2. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 19.Gu Y, Tinn R, Cheng H, Lucas M, Usuyama N, Liu X, et al. Domain-specific language model pretraining for biomedical natural language processing. ACM Trans Comput Healthcare. 2022;3:2. [Google Scholar]
- 20.Jiang LY, Liu XC, Nejatian NP, Nasir-Moin M, Wang D, Abidin A, et al. Health system-scale language models are all-purpose prediction engines. Nature. 2023;619:357–362. doi: 10.1038/s41586-023-06160-y. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 21.Van Veen D, Van Uden C, Blankemeier L, Delbrouck JB, Aali A, Bluethgen C, et al. Adapted large language models can outperform medical experts in clinical text summarization. Nat Med. 2024;30:1134–1142. doi: 10.1038/s41591-024-02855-5. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 22.Lu MY, Chen B, Williamson DFK, Chen RJ, Zhao M, Chow AK, et al. A multimodal generative AI copilot for human pathology. Nature. 2024;634:466–473. doi: 10.1038/s41586-024-07618-3. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 23.Taylor-Weiner A, Pokkalla H, Han L, Jia C, Huss R, Chung C, et al. A machine learning approach enables quantitative measurement of liver histology and disease monitoring in NASH. Hepatology. 2021;74:133–147. doi: 10.1002/hep.31750. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 24.Calderaro J, Ghaffari Laleh N, Zeng Q, Maille P, Favre L, Pujals A, et al. Deep learning-based phenotyping reclassifies combined hepatocellular-cholangiocarcinoma. Nat Commun. 2023;14:8290. doi: 10.1038/s41467-023-43749-3. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 25.Campanella G, Chen S, Singh M, Verma R, Muehlstedt S, Zeng J, et al. A clinical benchmark of public self-supervised pathology foundation models. Nat Commun. 2025;16:3640. doi: 10.1038/s41467-025-58796-1. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 26.Ding T, Wagner SJ, Song AH, Chen RJ, Lu MY, Zhang A, et al. A multimodal whole-slide foundation model for pathology. Nat Med. 2025;31:3749–3761. doi: 10.1038/s41591-025-03982-3. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 27.Chen RJ, Ding T, Lu MY, Williamson DFK, Jaume G, Song AH, et al. Towards a general-purpose foundation model for computational pathology. Nat Med. 2024;30:850–862. doi: 10.1038/s41591-024-02857-3. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 28.Kraljevic Z, Bean D, Shek A, Bendayan R, Hemingway H, Yeung JA, et al. Foresight-a generative pretrained transformer for modelling of patient timelines using electronic health records: a retrospective modelling study. Lancet Digit Health. 2024;6:e281–e290. doi: 10.1016/S2589-7500(24)00025-6. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 29.Acosta JN, Falcone GJ, Rajpurkar P, Topol EJ. Multimodal biomedical AI. Nat Med. 2022;28:1773–1784. doi: 10.1038/s41591-022-01981-2. [DOI] [PubMed] [Google Scholar]
- 30.Gorenshtein A, Omar M, Glicksberg BS, Nadkarni GN, Klang E. AI agents in clinical medicine: a systematic review. medRxiv [Preprint] doi: 10.1101/2025.08.22.25334232. 2025 [cited 2026 Jul 8]. Available from: [DOI]
- 31.Ferber D, El Nahhas OSM, Wölflein G, Wiest IC, Clusmann J, Leßmann ME, et al. Development and validation of an autonomous artificial intelligence agent for clinical decision-making in oncology. Nat Cancer. 2025;6:1337–1349. doi: 10.1038/s43018-025-00991-6. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 32.Aydin S, Karabacak M, Vlachos V, Margetis K. Large language models in patient education: a scoping review of applications in medicine. Front Med (Lausanne) 2024;11:1477898. doi: 10.3389/fmed.2024.1477898. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 33.Almagazzachi A, Mustafa A, Eighaei Sedeh A, Vazquez Gonzalez AE, Polianovskaia A, Abood M, et al. Generative artificial intelligence in patient education: ChatGPT takes on hypertension questions. Cureus. 2024;16:e53441. doi: 10.7759/cureus.53441. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 34.Caglar U, Yildiz O, Meric A, Ayranci A, Gelmis M, Sarilar O, et al. Evaluating the performance of ChatGPT in answering questions related to pediatric urology. J Pediatr Urol. 2024;20:26.e1–26.e5. doi: 10.1016/j.jpurol.2023.08.003. [DOI] [PubMed] [Google Scholar]
- 35.He Z, Bhasuran B, Jin Q, Tian S, Hanna K, Shavor C, et al. Quality of answers of generative large language models versus peer users for interpreting laboratory test results for lay patients: evaluation study. J Med Internet Res. 2024;26:e56655. doi: 10.2196/56655. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 36.Alanezi F. Examining the role of ChatGPT in promoting health behaviors and lifestyle changes among cancer patients. Nutr Health. 2025;31:739–748. doi: 10.1177/02601060241244563. [DOI] [PubMed] [Google Scholar]
- 37.Jin D, Pan E, Oufattole N, Weng W, Fang H, Szolovits P. What disease does this patient have? A large-scale open domain question answering dataset from medical exams. Appl Sci. 2021;11:6421. [Google Scholar]
- 38.Papineni K, Roukos S, Ward T, Zhu WJ. In: Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics. Isabelle P, Charniak E, Lin D, editors. Philadelphia (PA): Association for Computational Linguistics; 2002. Bleu: a method for automatic evaluation of machine translation; pp. 311–318. [Google Scholar]
- 39.Bedi S, Cui H, Fuentes M, Unell A, Wornow M, Banda JM, et al. Holistic evaluation of large language models for medical tasks with MedHELM. Nat Med. 2026;32:943–951. doi: 10.1038/s41591-025-04151-2. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 40.Thirunavukarasu AJ, Ting DSJ, Elangovan K, Gutierrez L, Tan TF, Ting DSW. Large language models in medicine. Nat Med. 2023;29:1930–1940. doi: 10.1038/s41591-023-02448-8. [DOI] [PubMed] [Google Scholar]
- 41.Topaz M, Roguin N, Gupta P, Zhang Z, Peltonen LM. Fabricated citations: an audit across 2·5 million biomedical papers. Lancet. 2026;407:1779–1781. doi: 10.1016/S0140-6736(26)00603-3. [DOI] [PubMed] [Google Scholar]
- 42.Chen S, Kann BH, Foote MB, Aerts HJWL, Savova GK, Mak RH, et al. Use of artificial intelligence chatbots for cancer treatment information. JAMA Oncol. 2023;9:1459–1462. doi: 10.1001/jamaoncol.2023.2954. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 43.Qazi IA, Ali A, Khawaja AU, Akhtar MJ, Sheikh AZ, Alizai MH. Automation bias in large language model–assisted diagnostic reasoning among physicians trained in AI literacy — a randomized clinical trial. NEJM AI. 2026;3:AIoa2501001. doi: 10.1056/AIoa2501001. [DOI] [PubMed] [Google Scholar]
- 44.Tejani AS, Ng YS, Xi Y, Rayan JC. Understanding and mitigating bias in imaging artificial intelligence. Radiographics. 2024;44:e230067. doi: 10.1148/rg.230067. [DOI] [PubMed] [Google Scholar]
- 45.Ge J, Sun S, Owens J, Galvez V, Gologorskaya O, Lai JC, et al. Development of a liver disease-specific large language model chat interface using retrieval-augmented generation. Hepatology. 2024;80:1158–1168. doi: 10.1097/HEP.0000000000000834. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 46.Kleiner DE, Brunt EM, Van Natta M, Behling C, Contos MJ, Cummings OW, et al. Design and validation of a histological scoring system for nonalcoholic fatty liver disease. Hepatology. 2005;41:1313–1321. doi: 10.1002/hep.20701. [DOI] [PubMed] [Google Scholar]
- 47.Loomba R, Ratziu V, Harrison SA, NASH Clinical Trial Design International Working Group Expert panel review to compare FDA and EMA guidance on drug development and endpoints in nonalcoholic steatohepatitis. Gastroenterology. 2022;162:680–688. doi: 10.1053/j.gastro.2021.10.051. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 48.Davison BA, Harrison SA, Cotter G, Alkhouri N, Sanyal A, Edwards C, et al. Suboptimal reliability of liver biopsy evaluation has implications for randomized clinical trials. J Hepatol. 2020;73:1322–1332. doi: 10.1016/j.jhep.2020.06.025. [DOI] [PubMed] [Google Scholar]
- 49.Ratziu V, Hompesch M, Petitjean M, Serdjebi C, Iyer JS, Parwani AV, et al. Artificial intelligence-assisted digital pathology for non-alcoholic steatohepatitis: current status and future directions. J Hepatol. 2024;80:335–351. doi: 10.1016/j.jhep.2023.10.015. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 50.Pulaski H, Harrison SA, Mehta SS, Sanyal AJ, Vitali MC, Manigat LC, et al. Clinical validation of an AI-based pathology tool for scoring of metabolic dysfunction-associated steatohepatitis. Nat Med. 2025;31:315–322. doi: 10.1038/s41591-024-03301-2. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 51.Iyer JS, Juyal D, Le Q, Shanis Z, Pokkalla H, Pouryahya M, et al. AI-based automation of enrollment criteria and endpoint assessment in clinical trials in liver diseases. Nat Med. 2024;30:2914–2923. doi: 10.1038/s41591-024-03172-7. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 52.U.S. Food and Drug Administration (FDA) FDA qualifies first AI drug development tool, will be used in ‘MASH’ clinical trials. FDA web site, < https://www.fda.gov/drugs/drug-safetyand-availability/fda-qualifies-first-ai-drug-development-toolwill-be-used-mash-clinical-trials>. Accessed 8 Jul 2026.
- 53.European Medicines Agency (EMA) Qualification opinion for artificial intelligence-based measurement of non-alcoholic steatohepatitis histology in liver biopsies to determine disease activity in NASH/MASH clinical trials. EMA, Committee for Medicinal Products for Human Use (CHMP) web site, < https://www.ema.europa.eu/en/documents/other/qualification-opinion-artificial-intelligence-based-measurement-nonalcoholic-steatohepatitis-histology-liver-biopsies-determinedisease-activity-nash-mash-clinical-trials_en.pdf>. Accessed 8 Jul 2026.
- 54.Naoumov NV, Brees D, Loeffler J, Chng E, Ren Y, Lopez P, et al. Digital pathology with artificial intelligence analyses provides greater insights into treatment-induced fibrosis regression in NASH. J Hepatol. 2022;77:1399–1409. doi: 10.1016/j.jhep.2022.06.018. [DOI] [PubMed] [Google Scholar]
- 55.Angulo P, Kleiner DE, Dam-Larsen S, Adams LA, Bjornsson ES, Charatcharoenwitthaya P, et al. Liver fibrosis, but no other histologic features, is associated with long-term outcomes of patients with nonalcoholic fatty liver disease. Gastroenterology. 2015;149:389–397.e10. doi: 10.1053/j.gastro.2015.04.043. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 56.Ekstedt M, Hagström H, Nasr P, Fredrikson M, Stål P, Kechagias S, et al. Fibrosis stage is the strongest predictor for disease-specific mortality in NAFLD after up to 33 years of follow-up. Hepatology. 2015;61:1547–1554. doi: 10.1002/hep.27368. [DOI] [PubMed] [Google Scholar]
- 57.Maghsoudi H, Khonche A, Gereami R, Gharebakhshi F. Physics-aware imaging AI for quantitative MASLD biomarker mapping: a systematic review of deep learning and radiomics across ultrasound, CT, and MRI. Abdom Radiol (NY) 2026;51:2919–2931. doi: 10.1007/s00261-025-05317-9. [DOI] [PubMed] [Google Scholar]
- 58.Vianna P, Calce SI, Boustros P, Larocque-Rigney C, Patry-Beaudoin L, Luo YH, et al. Comparison of radiologists and deep learning for US grading of hepatic steatosis. Radiology. 2023;309:e230659. doi: 10.1148/radiol.230659. [DOI] [PubMed] [Google Scholar]
- 59.Kwon H, Kim MG, Oh S, Kim Y, Jung G, Lee HJ, et al. Application of quantitative ultrasonography and artificial intelligence for assessing severity of fatty liver: a pilot study. Diagnostics (Basel) 2024;14:1237. doi: 10.3390/diagnostics14121237. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 60.Jeon SK, Lee JM, Joo I, Yoon JH, Lee G. Two-dimensional convolutional neural network using quantitative US for noninvasive assessment of hepatic steatosis in NAFLD. Radiology. 2023;307:e221510. doi: 10.1148/radiol.221510. [DOI] [PubMed] [Google Scholar]
- 61.Pickhardt PJ, Blake GM, Graffy PM, Sandfort V, Elton DC, Perez AA, et al. Liver steatosis categorization on contrast-enhanced CT using a fully automated deep learning volumetric segmentation tool: evaluation in 1204 healthy adults using unenhanced CT as a reference standard. AJR Am J Roentgenol. 2021;217:359–367. doi: 10.2214/AJR.20.24415. [DOI] [PubMed] [Google Scholar]
- 62.Hu N, Yan G, Tang M, Wu Y, Song F, Xia X, et al. CT-based methods for assessment of metabolic dysfunction associated with fatty liver disease. Eur Radiol Exp. 2023;7:72. doi: 10.1186/s41747-023-00387-0. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 63.Yoo J, Joo I, Jeon SK, Park J, Yoon SH. Utilizing fully-automated 3D organ segmentation for hepatic steatosis assessment with CT attenuation-based parameters. Eur Radiol. 2024;34:6205–6213. doi: 10.1007/s00330-024-10660-4. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 64.Wong GL, Hui VW, Tan Q, Xu J, Lee HW, Yip TC, et al. Novel machine learning models outperform risk scores in predicting hepatocellular carcinoma in patients with chronic viral hepatitis. JHEP Rep. 2022;4:100441. doi: 10.1016/j.jhepr.2022.100441. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 65.Hao T, Huang Z, Liang L, Weng H, Tang B. Health natural language processing: methodology development and applications. JMIR Med Inform. 2021;9:e23898. doi: 10.2196/23898. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 66.Huang J, Yang DM, Rong R, Nezafati K, Treager C, Chi Z, et al. A critical assessment of using ChatGPT for extracting structured data from clinical notes. NPJ Digit Med. 2024;7:106. doi: 10.1038/s41746-024-01079-8. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 67.Khooblall A, Viswanath SE, Khawaja L, Gadani S. Artificial intelligence applications in the diagnosis and management of cirrhosis and portal hypertension: a narrative review. Tech Vasc Interv Radiol. 2025;28:101078. doi: 10.1016/j.tvir.2025.101078. [DOI] [PubMed] [Google Scholar]
- 68.Wang QC, Jiao J, Zhang CQ. Application of artificial intelligence in portal hypertension and esophagogastric varices. World J Gastroenterol. 2025;31:108508. doi: 10.3748/wjg.v31.i24.108508. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 69.Far AT, Bastani A, Lee A, Gologorskaya O, Huang CY, Pletcher MJ, et al. Evaluating the positive predictive value of code-based identification of cirrhosis and its complications utilizing GPT-4. Hepatology. 2025;81:1753–1763. doi: 10.1097/HEP.0000000000001115. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 70.Li Y, Li Z, Li J, Liu L, Liu Y, Zhu B, et al. The actual performance of large language models in providing liver cirrhosis-related information: a comparative study. Int J Med Inform. 2025;201:105961. doi: 10.1016/j.ijmedinf.2025.105961. [DOI] [PubMed] [Google Scholar]
- 71.Yeo YH, Samaan JS, Ng WH, Ting PS, Trivedi H, Vipani A, et al. Assessing the performance of ChatGPT in answering questions regarding cirrhosis and hepatocellular carcinoma. Clin Mol Hepatol. 2023;29:721–732. doi: 10.3350/cmh.2023.0089. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 72.Pradhan F, Fiedler A, Samson K, Olivera-Martinez M, Manatsathit W, Peeraphatdit T. Artificial intelligence compared with human-derived patient educational materials on cirrhosis. Hepatol Commun. 2024;8:e0367. doi: 10.1097/HC9.0000000000000367. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 73.Gottweis J, Weng WH, Daryin A, Tu T, Sirkovic P, Myaskovsky A, et al. Accelerating scientific discovery with Co-Scientist. Nature. 2026;655:487–496. doi: 10.1038/s41586-026-10644-y. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 74.Guan Y, Cui L, Inchai J, Fang Z, Law J, Brito AAG, et al. AI-Assisted drug re-purposing for human liver fibrosis. Adv Sci (Weinh) 2025;12:e08751. doi: 10.1002/advs.202508751. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 75.Ding W, Zhang J, Jin Z, Hua H, Zu Q, Yang S, et al. Artificial intelligence-driven pathomics in hepatocellular carcinoma: current developments, challenges and perspectives. Discov Oncol. 2025;16:1424. doi: 10.1007/s12672-025-03254-z. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 76.Cheng N, Ren Y, Zhou J, Zhang Y, Wang D, Zhang X, et al. Deep learning-based classification of hepatocellular nodular lesions on whole-slide histopathologic images. Gastroenterology. 2022;162:1948–1961. doi: 10.1053/j.gastro.2022.02.025. e7. [DOI] [PubMed] [Google Scholar]
- 77.Kiani A, Uyumazturk B, Rajpurkar P, Wang A, Gao R, Jones E, et al. Impact of a deep learning assistant on the histopathologic classification of liver cancer. NPJ Digit Med. 2020;3:23. doi: 10.1038/s41746-020-0232-8. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 78.Zhang X, Yu X, Liang W, Zhang Z, Zhang S, Xu L, et al. Deep learning-based accurate diagnosis and quantitative evaluation of microvascular invasion in hepatocellular carcinoma on whole-slide histopathology images. Cancer Med. 2024;13:e7104. doi: 10.1002/cam4.7104. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 79.Chen M, Zhang B, Topatana W, Cao J, Zhu H, Juengpanich S, et al. Classification and mutation prediction based on histopathology H&E images in liver cancer using deep learning. NPJ Precis Oncol. 2020;4:14. doi: 10.1038/s41698-020-0120-3. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 80.Lin H, Wei C, Wang G, Chen H, Lin L, Ni M, et al. Automated classification of hepatocellular carcinoma differentiation using multiphoton microscopy and deep learning. J Biophotonics. 2019;12:e201800435. doi: 10.1002/jbio.201800435. [DOI] [PubMed] [Google Scholar]
- 81.Chen Q, Xiao H, Gu Y, Weng Z, Wei L, Li B, et al. Deep learning for evaluation of microvascular invasion in hepatocellular carcinoma from tumor areas of histology images. Hepatol Int. 2022;16:590–602. doi: 10.1007/s12072-022-10323-w. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 82.Shi JY, Wang X, Ding GY, Dong Z, Han J, Guan Z, et al. Exploring prognostic indicators in the pathological images of hepatocellular carcinoma based on deep learning. Gut. 2021;70:951–961. doi: 10.1136/gutjnl-2020-320930. [DOI] [PubMed] [Google Scholar]
- 83.Liu Z, Liu Y, Zhang W, Hong Y, Meng J, Wang J, et al. Deep learning for prediction of hepatocellular carcinoma recurrence after resection or liver transplantation: a discovery and validation study. Hepatol Int. 2022;16:577–589. doi: 10.1007/s12072-022-10321-y. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 84.Zeng Q, Klein C, Caruso S, Maille P, Allende DS, Mínguez B, et al. Artificial intelligence-based pathology as a biomarker of sensitivity to atezolizumab-bevacizumab in patients with hepatocellular carcinoma: a multicentre retrospective study. Lancet Oncol. 2023;24:1411–1422. doi: 10.1016/S1470-2045(23)00468-0. [DOI] [PubMed] [Google Scholar]
- 85.Wang Q, Huang Y, Zhang Y, Zhu Y, Hu P, Xu Y, et al. Digital pathology-based prognostic model for hepatocellular carcinoma: integrating pathomics signatures with clinical parameters for recurrence prediction and biological interpretation. Comput Methods Programs Biomed. 2026;275:109180. doi: 10.1016/j.cmpb.2025.109180. [DOI] [PubMed] [Google Scholar]
- 86.Liao H, Long Y, Han R, Wang W, Xu L, Liao M, et al. Deep learning-based classification and mutation prediction from histopathological images of hepatocellular carcinoma. Clin Transl Med. 2020;10:e102. doi: 10.1002/ctm2.102. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 87.Wiest IC, Bhat M, Clusmann J, Schneider CV, Jiang X, Kather JN. Large language models for clinical decision support in gastroenterology and hepatology. Nat Rev Gastroenterol Hepatol. 2025;22:773–787. doi: 10.1038/s41575-025-01108-1. [DOI] [PubMed] [Google Scholar]
- 88.Yang K, Lee J, Jang JW, Sung PS, Han JW. Evaluating the clinical utility of large language models for hepatocellular carcinoma treatment recommendations: a nationwide retrospective registry study. PLoS Med. 2026;23:e1004855. doi: 10.1371/journal.pmed.1004855. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 89.Deeb M, Gangadhar A, Rabindranath M, Rao K, Brudno M, Sidhu A, et al. The emerging role of generative artificial intelligence in transplant medicine. Am J Transplant. 2024;24:1724–1730. doi: 10.1016/j.ajt.2024.06.009. [DOI] [PubMed] [Google Scholar]
- 90.Kleebayoon A, Wiwanitkit V. Letter to the editor: ChatGPT and unknown clinical questions about liver transplantation. Liver Transpl. 2024;30:E22. doi: 10.1097/LVT.0000000000000281. [DOI] [PubMed] [Google Scholar]
- 91.Umman V, Tosun B, Uygur A, Emre S. Evaluation of the effectiveness, safety, and patient satisfaction of artificial intelligence-based patient education and counseling for both recipients and donors in the preoperative and postoperative phases of organ transplantation. Transplant Proc. 2025;57:1832–1839. doi: 10.1016/j.transproceed.2025.07.001. [DOI] [PubMed] [Google Scholar]
- 92.Robitschek E, Bastani A, Horwath K, Sordean S, Pletcher MJ, Lai JC, et al. A large language model-based approach to quantifying the effects of social determinants in liver transplant decisions. NPJ Digit Med. 2025;8:665. doi: 10.1038/s41746-025-02025-y. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 93.Hasjim BJ, Azarfar G, Lee FG, Diwan TS, Raju S, Gross JA, et al. A multiagent large language model-based system to simulate the liver transplant selection committee: a retrospective cohort study. Lancet Digit Health. 2026;8:100966. doi: 10.1016/j.landig.2025.100966. [DOI] [PubMed] [Google Scholar]
- 94.Qu L, Balachandar N, Zhang M, Rubin D. Handling data heterogeneity with generative replay in collaborative learning for medical imaging. Med Image Anal. 2022;78:102424. doi: 10.1016/j.media.2022.102424. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 95.Clunie DA, Dennison DK, Cram D, Persons KR, Bronkalla MD, Primo HR. Technical challenges of enterprise imaging: HIMSS-SIIM collaborative white paper. J Digit Imaging. 2016;29:583–614. doi: 10.1007/s10278-016-9899-4. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 96.Celi LA, Cellini J, Charpignon ML, Dee EC, Dernoncourt F, Eber R, et al. Sources of bias in artificial intelligence that perpetuate healthcare disparities—a global review. PLOS Digit Health. 2022;1:e0000022. doi: 10.1371/journal.pdig.0000022. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 97.Clusmann J, Balaguer-Montero M, Bassegoda O, Schneider CV, Seraphin T, Paintsil E, et al. The barriers to uptake of artificial intelligence in hepatology and how to overcome them. J Hepatol. 2025;83:1410–1426. doi: 10.1016/j.jhep.2025.07.003. [DOI] [PubMed] [Google Scholar]
- 98.Petersson L, Larsson I, Nygren JM, Nilsen P, Neher M, Reed JE, et al. Challenges to implementing artificial intelligence in healthcare: a qualitative interview study with healthcare leaders in Sweden. BMC Health Serv Res. 2022;22:850. doi: 10.1186/s12913-022-08215-8. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 99.Gallifant J, Afshar M, Ameen S, Aphinyanaphongs Y, Chen S, Cacciamani G, et al. The TRIPOD-LLM reporting guideline for studies using large language models. Nat Med. 2025;31:60–69. doi: 10.1038/s41591-024-03425-5. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 100.Cruz Rivera S, Liu X, Chan AW, Denniston AK, Calvert MJ, SPIRIT-AI and CONSORT-AI Working Group et al. Guidelines for clinical trial protocols for interventions involving artificial intelligence: the SPIRIT-AI extension. Nat Med. 2020;26:1351–1363. doi: 10.1038/s41591-020-1037-7. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 101.Liu X, Cruz Rivera S, Moher D, Calvert MJ, Denniston AK, SPIRIT-AI and CONSORT-AI Working Group Reporting guidelines for clinical trial reports for interventions involving artificial intelligence: the CONSORT-AI extension. Nat Med. 2020;26:1364–1374. doi: 10.1038/s41591-020-1034-x. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 102.U.S. Food and Drug Administration (FDA) Marketing submission recommendations for a predetermined change control plan for artificial intelligence-enabled device software functions: guidance for industry and food and drug administration staff. FDA web site, < https://www.fda.gov/regulatory-information/search-fda-guidance-documents/marketing-submissionrecommendations-predetermined-change-control-planartificial-intelligence>. Accessed 8 Jul 2026.
- 103.European Parliament Council of the European Union Regulation (EU) 2024/1689 of the European Parliament and of the Council of 13 June 2024 laying down harmonised rules on artificial intelligence and amending certain Union legislative acts (Artificial Intelligence Act) European Union web site, < https://eur-lex.europa.eu/eli/reg/2024/1689/oj>. Accessed 8 Jul 2026.
