Artificial intelligence and machine learning in hepatology: Revolutionizing diagnosis and treatment

Abstract

Artificial intelligence (AI) has emerged as a powerful tool in the field of hepatology, offering new opportunities to enhance diagnosis, risk stratification, and therapeutic decision-making. AI models have demonstrated improved performance compared with conventional methods by integrating complex multimodal datasets. These models have shown promising results across a broad spectrum of liver disease management. Despite these advances, unmet needs remain a significant barrier to the full integration of AI models into clinical practice. Future progress will depend on developing interpretable, well-validated AI systems supported by multicentre collaborations and robust data infrastructure. In this comprehensive review, we summarise the current applications of AI in hepatology, highlight areas of significant clinical promise, and outline the challenges and future directions necessary for safe, equitable, and effective integration into routine practice.

Key Words: Artificial intelligence; Machine learning; Deep learning; Liver transplant; Hepatocellular carcinoma; Cirrhosis

Core Tip: Artificial intelligence (AI) is transforming hepatology by improving disease detection, prognostication, and treatment planning. Machine learning and deep learning models outperform traditional tools by integrating imaging, laboratory, histological, and clinical data. However, challenges, including dataset bias, lack of interpretability, and limited prospective validation, still hinder routine clinical use. This review outlines current applications, limitations, and future directions, highlighting how AI could support more precise, equitable, and personalized liver care.

INTRODUCTION

Hepatology, the field focused on liver disease, reflects the complexity of modern clinical practice. Accurate diagnosis, prognosis, and treatment require integrating multiple forms of diagnostic data. Patients typically undergo detailed history-taking, physical examination, laboratory investigations, and imaging, with liver biopsy used when needed. Each modality provides valuable information, but synthesizing these data can be difficult, and diagnostic uncertainty remains common even for experienced clinicians[1].

Artificial intelligence (AI) allows computers to learn from complex data and address real-world challenges, both in medicine and other fields, often achieving performance comparable to or exceeding that of humans. Rather than relying on explicit expert instructions, AI systems learn patterns and decision-making directly from example data. Two main subsets of AI have been continuously evolving over the last decade, providing a foundation for AI applications in the field of medicine. Firstly, machine learning (ML), which does not require specialized hardware but depends on “handcrafted features” defined by experts. For example, in hepatology imaging, this means clinicians manually selecting measurable image characteristics such as tumor size, shape, symmetry, or contrast intensity to predict outcomes. These predefined variables are then processed by algorithms like random forests (RFs), which handle structured data well. This handcrafted feature-based approach is commonly known as “radiomics” or “classical radiomics”[2]. The goal of ML is to create mathematical models that help machines make decisions or predictions without explicit programming. Numerous ML approaches, such as support vector machines (SVM), artificial neural networks (ANNs), regression trees, and classification, appear to be used in a variety of medical research[3].

Secondly, deep learning (DL), which has expanded rapidly over the past decade due to advances in algorithms, hardware, and the availability of large datasets. Although related to traditional machine-learning methods, DL models contain far more trainable parameters, allowing them to capture complex patterns in data such as medical images[4]. Common DL architectures in medicine include ANNs for image and time-series analysis and image processing[5]. Unlike classical ML, DL does not rely on predefined, handcrafted features; instead, networks automatically identify features relevant to clinical outcomes[1]. Consequently, DL typically outperforms handcrafted approaches and now dominates AI research in hepatology. Nonetheless, the boundary between these methods is not strict, as some studies combine DL-derived features with handcrafted variables. Overall, ML and DL can either automate expert-level data interpretation or detect subtle, clinically meaningful patterns beyond human perception[6]. Graphical illustration can be found in Figure 1.

Figure 1 End-to-end pipeline of data processing and prediction in hepatology artificial intelligence models. This figure illustrates the end-to-end pipeline by which artificial intelligence processes multimodal clinical data in hepatology. Diverse inputs, including clinical notes, laboratory tests, imaging (computed tomography, magnetic resonance imaging, ultrasound, positron emission tomography), histopathology, and omics, are aggregated and cleaned. Cleaned datasets are then analyzed using machine learning and deep learning algorithms to enable automated feature extraction. Feature extraction encompasses both handcrafted variables (e.g., liver stiffness, liver imaging reporting and data system category, radiomic texture metrics, fibrosis stage) and automatically learned features from deep networks. These computationally derived features feed predictive models that support downstream clinical applications such as hepatocellular carcinoma detection, digital pathology interpretation, risk stratification, grading of portal hypertension, prediction of variceal bleeding or ascites infection, and post-transplant graft or patient survival. The workflow highlights how artificial intelligence integrates complex datasets into actionable outputs that enhance precision hepatology. CT: Computed tomography; MRI: Magnetic resonance imaging; PET: Positron emission tomography; USS: Ultrasound; ML: Machine learning; DL: Deep learning; HCC: Hepatocellular carcinoma.

In practice, classical ML and DL are complementary rather than competing approaches in hepatology. Tree-based ML models and regularized regression perform particularly well on structured tabular data, such as laboratory tests, clinical scores, and waitlist variables, and are often preferable when datasets are moderate in size or when model interpretability is a priority (e.g., risk prediction for cirrhosis mortality or transplant outcomes). In contrast, DL architectures such as convolutional neural networks (CNN)s and residual networks excel in high-dimensional image and histology data, underpinning most advances in automated hepatocellular carcinoma (HCC) detection on computed tomography (CT)/magnetic resonance imaging (MRI), radiomics, and computational pathology. Hybrid approaches that combine DL-derived imaging features with ML models trained on clinical variables are increasingly common, and may provide the best performance for complex tasks such as transcatheter arterial chemoembolization (TACE) response prediction or portal hypertension grading[1,7,8].

This comprehensive review provides insights into the recent advances in clinical application of AI technology in diagnostics and therapeutic strategies in the field of hepatology, with special focus on HCC and liver transplant medicine.

AI in liver cirrhosis and its complications

For interventional radiology specialists, diagnosing and treating cirrhosis and its primary consequence, portal hypertension, is a significant issue. Due to the intricacy of imaging results and the presence of characteristics that coincide with those of other liver disorders, early diagnosis is frequently challenging. Additionally, conventional imaging modalities, including MRI, CT, and Doppler ultrasonography, have each shown shortcomings in various circumstances.

AI-driven prediction tools have demonstrated promising potential in clinical applications related to liver diseases. Chang et al[9] conducted an evaluation of several ML models, specifically focusing on logistic regression, regression, RF algorithms, as well as ANNs for the purpose of predicting the stages of fibrosis in a cohort of 1370 metabolic dysfunction-associated steatotic liver disease (MASLD) patients. The study indicated that these models demonstrated a marked improvement in predictive accuracy when compared to traditional methods such as Fibroscan and the fibrosis (FIB)-4 test.

DL has been utilized on MRI data to distinguish between alcohol-related cirrhosis and cirrhosis from other causes. Luetkens et al[10] showed that ResNet50 delivered the best results in a group of 465 patients, achieving 75% accuracy and an area under the curve (AUC) of 0.82. Additionally, Wei et al[11] used Pyradiomics software to extract radiomic features from MRI scans of 280 patients with chronic liver disease, thereby facilitating accurate staging of fibrosis and grading of inflammation.

Anushiravani et al[12] developed the FIB-6 index in a different study using RF algorithms, and it showed a high sensitivity and specificity in excluding cirrhosis in hepatitis C virus (HCV) patients. Following validation, the FIB-6 score outperformed other non-invasive indices in a sample of 2472 biopsy-proven MASLD patients from Egypt, Iran, Saudi Arabia, Greece, Turkey, and Oman. The FIB-6 index performs well, but its applicability across different aetiologies and geographical areas is constrained by its reliance on biopsy-proven data from specific cohorts.

Beyond cross-sectional fibrosis staging, AI tools are increasingly being explored for early detection and longitudinal monitoring of MASLD. ML-based scores derived from routinely available laboratory and clinical parameters (e.g., FIB-6 and other non-invasive indices) can be recalculated over time to track fibrosis progression or regression in large metabolic cohorts[9,12]. Similarly, radiomics and DL models applied to ultrasound, CT, and MRI may enable repeated, fully non-invasive assessment of steatosis, fibrosis, and portal hypertension without the need for serial biopsies, supporting population-level screening and follow-up strategies. A sequential algorithm combining FIB-4 and ultrasound DL model showed higher diagnostic accuracy and improved referral management for all-cause advanced liver fibrosis compared with FIB-4 or the DL model alone, which provides a reliable fibrosis screening tool[13]. Another study by Sun et al[14] investigated liver fibrosis progression in MASLD patients and the risk of renal decline. The investigators used an automated quantitative liver fibrosis assessment (qFibrosis) technique to investigate the temporal changes in regional liver fibrosis, which provides vital insights into the role of AI technology in the quantification of liver fibrosis progression in this cohort.

Using trichrome-stained liver biopsy samples from individuals with cirrhosis brought on by MASLD, Bosch et al[15] created an ML model. Clinically significant portal hypertension (CSPH) was identified by this model, with AUCs of 0.85 and 0.76 during training and testing, respectively. In a similar vein, Yu et al[16] used an automated CT radiomics hepatic venous pressure gradient quantitative model, which outperformed other non-invasive methods, such as elastography, in determining the severity of portal hypertension. While Reiniš et al[17] used ML models based on common lab parameters, such as bilirubin, platelet count, and international normalized ratio, to predict CSPH in patients with compensated cirrhosis, Noureddin et al[18] used ML to develop a scoring system based on septa, nodules, and fibrosis derived from liver biopsies in MASLD-related cirrhosis, achieving a high predictive accuracy for CSPH.

A neural network developed by Hou et al[19] outperformed traditional indices in predicting the 1-year risk of esophagogastric variceal haemorrhage based on 12 independent risk factors. Similarly, Chen et al[20] assessed ENDOANGEL, a deep CNN trained on 14718 endoscopic images, which outperformed endoscopists in several crucial areas, including detecting gastro-oesophageal varices and predicting rupture risk.

By providing non-invasive and customized methods, ML models have the potential to improve ascites management. To identify patients with infected ascites, Würstle et al[21] created a model using 34 widely accessible clinical data points. This model showed strong negative predictive values across a range of pre-test probabilities, suggesting it may be a reliable non-invasive alternative to paracentesis. Similarly, Hatami et al[22] used normal clinical data to predict ascites grade using ML models, such as k-nearest neighbours (KNNs), SVMs, and neural networks. With an accuracy of 94%, the KNN model outperformed the others, demonstrating ML’s capacity to predict the severity of ascites accurately.

Overall, these studies demonstrate the growing impact of AI in addressing key diagnostic and prognostic challenges in cirrhosis and portal hypertension. ML and DL models consistently outperform traditional tools by integrating complex imaging, laboratory, and clinical data to improve fibrosis staging, detect CSPH, predict variceal bleeding, and guide ascites management. Although further validation across diverse populations is needed, AI-driven approaches show strong potential to enhance non-invasive assessment, reduce reliance on invasive procedures, and support more precise, personalized care in hepatology.

AI in autoimmune liver diseases

Autoimmune liver diseases are chronic conditions of the liver and biliary system, thought to arise from autoimmune mechanisms. They include autoimmune hepatitis (AIH), primary biliary cholangitis (PBC), and primary sclerosing cholangitis (PSC). Progress in drug development particularly for PSC and AIH has been limited by their low prevalence, unclear causes, and significant heterogeneity among patients who meet the same diagnostic criteria[23]. Nonetheless, high-throughput DNA and RNA sequencing technologies, digital pathology, and digital radiology are also gradually entering this overlooked field. An increasing volume of experimental and clinical data is becoming accessible in the field[24], which requires dedicated analytical pipelines capable of handling big data.

Histopathology

Histopathology slides contain vast information that has historically been underutilized. Their digitization through whole slide imaging (WSI) forms the basis of digital pathology[25], which, together with ML and DL, has advanced into the field of computational pathology[26]. This approach enables data extraction from digital slides for clinical and research purposes, offering applications in diagnostics (automation, decision support, telemedicine) and research (pathogenesis, prognostication, risk stratification)[23]. In hepatology and rare disorders such as autoimmune liver diseases, WSI may improve diagnosis, prognosis, and treatment decisions[27]. However, challenges remain, including lack of universal data standards compared to radiology[1], variability in slide quality due to staining, section thickness, and artifacts[28], as well as batch effects across institutions or time points[29].

Radiology

The anticipated integration of AI into medical practice, along with the production of extensive high-quality imaging data, positions radiology as a crucial contributor to precision medicine[30,31]. To this end, two groups of AI-based techniques can be mentioned: Radiomics, which relies on ML, and DL systems (based on neural networks)[1]. Radiomics has emerged as a high-throughput computing technique that enables the extraction of large amounts of quantitative features from medical imaging, mainly CT, MRI, and positron emission tomography (PET)[32]. We can speculate that the ultimate goal of both techniques is the combination of radiological data with clinical and laboratory data and potentially other omics, to develop more accurate predictive models that incorporate a wider spectrum of disease-related features[7,33].

There is evidence supporting the reliability of magnetic resonance cholangiopancreatography (MRCP) + metrics as a non-invasive tool for differentiating paediatric PSC from paediatric AIH[34,35]. In addition, MRCP + parameters hold prognostic value, as proven by their strong correlation with validated biochemical and semi-quantitative MRCP-based risk scoring systems[36,37].

As regards AIH, promising data employing Liver MultiScan technology have been recently presented. Following the evidence that multiparametric MRI (mpMRI) using iron-corrected T1 relaxation maps provides an accurate, non-invasive quantitative biomarker of liver fibrosis and inflammation, recent works have shown that mpMRI, when applied to AIH patients, has a better performance in detecting residual disease activity than serological biomarkers[38,39].

Genetics

From a genetic perspective, autoimmune liver diseases are complex traits[40] with several genetic interactive relationships are expected in this context. Although genome wide associated study has already been developed for these autoimmune liver diseases, sequencing data are still lacking. Additionally, single-nucleotide polymorphisms (SNPs) within the human leukocyte antigen region play a pivotal role in determining the genetic risks in this cohort[41-43]. Given the complexity and large amount of data that can be obtained from population genetics, supervised ML can provide a valuable tool to handle these data efficiently and build up a polygenic model that can accurately classify cases and controls based on their SNPs.

Moreover, considering different phenotypic features, ML could potentially generate a predictive model that incorporates both genotypic and phenotypic features in a single equation, thereby enhancing predictive power and further understanding of the influence of specific genetic variants on disease expression patterns[44]. It is worth mentioning that these models require training data sets with both internal and external validations to avoid overfitting and also test the model’s replication.

AI IN HCC

HCC is an extremely deadly liver cancer. It is the most prevalent primary liver cancer in adults, which ranks as the third most frequent cause of cancer-related mortality globally. The mainstay of treatment for HCC is surgery, which includes both liver transplantation and resection. For tiny tumors, delamination or removal is another therapy option. Furthermore, intra-arterial therapy and chemotherapy can somewhat regulate the illness[45].

AI in HCC diagnostics

The reduction of diagnostic variability, the reallocation of healthcare resources, and the optimization of data analysis remain unmet needs in HCC management and represent the three main benefits of AI in HCC diagnosis. A combination of radiological, histological, and cytological factors is used to diagnose HCC[46].

Abdominal ultrasound

B-mode ultrasound’s sensitivity for identifying HCC is only 46%-63%[47]. To address this, several recent studies have examined whether AI frameworks can enhance ultrasound’s diagnostic accuracy in this context. A CNN model was trained using ultrasound data from 3487 patients in a large multi-center study[48]. It achieved 87% [95% confidence interval (CI): 84.3%-89.6%] and 75% (95%CI: 71.7%-78.3%) detection rates in an internal validation and external validation cohorts, respectively, on the task of classification among HCCs, cysts, haemangiomas, focal fatty sparing, and focal fatty infiltration. In the external validation set, the specificity and negative predictive value for HCC diagnosis were 94.4% (95%CI: 92.8%-96.0%) and 97.4% (95%CI: 96.2%-98.5%), respectively.

The potential of AI as a screening tool for alpha-fetoprotein (AFP)-negative HCCs was demonstrated by a DL model described by Zhang et al[49]. A total of 305 B-mode ultrasound images of HCC and focal nodular hyperplasia were used to train the CNN model (Xception)[50], and 102 B-mode ultrasound images were used to evaluate the model. Both the training and testing data sets showed heterogeneity in HCC staging, lesion size, echogenicity, and liver function.

A DL algorithm applied to liver lesions detected by contrast-enhanced ultrasound (CEUS) could improve the sensitivity, specificity, and overall accuracy of CEUS for detecting HCC, as Guo et al[51] recently showed. Others have improved the identification of uncertain focal liver lesions by using AI to add additional pattern recognition classifiers to CEUS DCNN algorithms[52]. Nevertheless, the majority of earlier CEUS research has had limited sample sizes and lacked external validation cohorts or standardized imaging data (to test the generalizability of models across populations).

CT

A DL model described by Wang et al[53] reported promising results in detection of patients with HCC derived from CT data with an AUC of 0.887 (95%CI: 0.855-0.919) regarding internal data set and 0.883 (95%CI: 0.855-0.911) for an outside data collection[53].

Yasaka et al[54] used contrast-enhanced CT imaging and a three-layer ANN model to classify liver masses into five categories: A (HCC, cholangiocarcinoma, or metastases), B (other malignant tumours), C (indeterminate lesions, including dysplastic nodules or early HCC and other benign masses), D (haemangiomas), and E (cysts).

An automatic recurrence detection method based on early tumor manifestation, CT behavior, baseline tumor load/mass quantification, and follow-up was proposed by Vivanti et al[55]. This method showed a greater percentage of true positives in identifying tumor relapse, with an accuracy of 86%.

MRI

A neural network (NN) algorithm developed by Hamm et al[56] was able to classify MRI liver lesions with 92% sensitivity, 98% specificity, and 92% overall accuracy. Zhang et al[57] investigated an automated method for segmenting multi-parameter MRIs in 20 HCC patients and showed that it was possible to avoid the laborious procedure of creating MRI-based features by hand. Further research has developed an automated classification method that classifies hepatic lesions as adenoma, cyst, haemangioma, HCC, and metastasis using additional MRI sequences, risk factors, and patient clinical data with a sensitivity/specificity of 0.80/0.78, 0.93/0.93, 0.84/0.82, 0.73/0.56, and 0.62/0.77, respectively[58].

PET

In order to evaluate the results of fluorine 18 fluorodeosyglucose PET/CT, Preis et al[59] employed a NN to examine hepatic intake of fluorodeosyglucose 18F in conjunction with patient data and clinical information. Higher sensitivity and specificity were found to identify liver cancer that was not readily apparent. This technique can assist the radiologist with PET analysis.

HCC histology

A residual CNN was developed to reduce manual processing in HCC classification using 592 hematoxylin and eosin-stained histopathology slides. The model showed excellent performance, achieving 87.9% accuracy on an external validation set, with 87.8% pixel-level accuracy and 98.7% slide-level accuracy in internal testing[60]. Using prior histological photos of HCC, others described how a deep CNN can automatically diagnose HCC, distinguish between normal and malignant tissue, and find important biological predictors[61].

AI IN EVALUATING HCC THERAPEUTICS

Surgical resection

Ji et al[62] created predictive models for relapse following excision surgery for evaluating contrast-enhanced CT images using radiomic techniques; these models had a C-index value of 0.633-0.699. These models could be used to provide a personalized risk assessment for each patient’s HCC management. Numerous studies have shown that ML methods are useful in evaluating survival following surgical resection[63-66].

TACE

Another automated ML algorithm-based model was developed to forecast TACE response by utilizing a combination of quantitative CT image features and pre-treatment clinical data of the patients[8]. This model achieved a prediction accuracy rate of 74.2% while collaborating on merging the Barcelona clinic liver cancer (BCLC) staging and quantitative imaging features, rather than utilizing the BCLC staging only. Similarly, Peng et al[66] utilized CT scans from 789 individuals across three distinct hospitals to validate a DL model for predicting TACE response[67].

Radiofrequency ablation

An SVM-based prognostic model was developed to predict HCC relapse following radiofrequency ablation, using data from 83 treated patients. The model achieved an AUC of 0.69, with 67% sensitivity and 86% specificity, enabling the identification of individuals at higher risk of recurrence[68].

AI IN LIVER TRANSPLANTATION

Organ allocations

All transplant programs aim to achieve both equity and utility. This ensures that all patients on the waiting list are transplanted promptly based on the severity and urgency of their clinical condition, using suitable donor livers to optimize outcomes for both donor liver and recipient.

Typically, statistical models have been employed to prioritise liver transplant candidates on the waiting lists using disease severity scores such as the model for end-stage liver disease (MELD). However, MELD is a poor predictor of post-transplant survival and is not suitable for all liver conditions, including paediatrics, retransplant cases, severe portal hypertension, and other variants.

This has sparked greater interest in using ML to predict mortality for patients on waiting lists[69]. Applying ML methods to large, longitudinal electronic health record (EHR) data, including a wide range of features, could allow for earlier detection of patients at higher risk of death. It also opens the possibility of incorporating novel approaches to measure clinical complications of cirrhosis, like ascites, hepatic encephalopathy, and frailty[70].

Statistical modelling in liver transplantation also aims to optimize matching between donor livers and recipients through scoring systems that consider donor and recipient characteristics. A practical example is the transplant benefit score (TBS), introduced by the national health services blood and transplant in March 2018. The TBS uses 21 recipient and 7 donor criteria to generate a score, which estimates benefit by comparing waiting list survival with posttransplant survival. Currently, the TBS applies only to adult donation after brain death livers. The algorithm’s clinical performance is regularly reviewed and refined[71].

Basing prediction models on statistical methods assumes linear relationships, which has led to increased interest in ML. So far, the utility of ANNs and RFs has been explored. ANNs are effective at making predictions from large datasets but suffer from being a “black box” in that the large number of variables and layers of computational analysis are not easily understood or accepted by clinicians. RF’s performance is limited when there are too many features, but it performs well in the context of missing data.

Although ANNs have been explored for donor allocation using multi-center transplant registry datasets, their real-time application in clinical practice remains untested[69,70]. Recently, Nagai et al[72] developed a NN model aimed at predicting 90-day mortality (or dropout due to worsening illness) among patients on the liver transplant waitlist. NNs, a subset of ML, can be structured with one (shallow) or multiple (deep) hidden layers between input and output nodes, which improves modelling of complex, non-linear relationships. In their study, the investigators analyzed patients listed in the Organ Procurement and Transplantation Network/United Network for Organ Sharing database between 2002 and 2021. After excluding those who received a transplant within 90 days of listing, the final cohort consisted of 105140 patients with 1540733 observations, split into training, validation, and test datasets in a 60:20:20 ratio. The NN model showed superior performance compared with MELD and MELD-sodium (MELD-Na), with an area under the receiver operating characteristic (AUROC) of 0.936 (95%CI: 0.934-0.937) vs 0.860 (95%CI: 0.858-0.862). These results demonstrate the potential of ML in liver transplantation but also reveal key limitations. Excluding early transplants may bias outcomes, and major predictors largely mirror MELD components. Although adding complications like ascites and encephalopathy improves accuracy, their subjective assessment limits reliability, objective biomarkers could strengthen future models.

Although the current “sickest-first” allocation system remains necessary given organ scarcity, future innovations such as bioengineered grafts or xenotransplantation may expand organ availability and enable earlier transplantation, reducing waitlist mortality. In this context, accurate long-term mortality prediction models will become increasingly important. Using data from more than 107000 veterans with cirrhosis, Kanwal et al[69] developed multiple ML models to predict 1-year mortality. Extreme gradient boosting, least absolute shrinkage and selection operator logistic regression, and a simplified model performed similarly, with the simplified model refined into the CiMM, which surpassed MELD-Na (AUROC: 0.78 vs 0.67). However, its dependence on symptom-based variables and underrepresentation of women and minority groups limited generalizability[69].

In a second study, Guo et al[70] evaluated ML models to predict 1-year mortality in 34575 cirrhosis patients from a large transplant centre. They compared deep neural network (DNNs) with the standard logistic regression. The DNN and RF models showed superior performance (AUROCs: 0.85-0.86) compared with logistic regression (0.69). The models drew on 41 predictors, with key features including laboratory markers (alkaline phosphatase, alanine aminotransferase, haemoglobin) and cirrhosis complications identified through billing codes.

Together with prior work, these findings highlight the promise of ML in improving short- and long-term mortality prediction in cirrhosis, potentially reducing waitlist deaths. However, challenges remain around generalizability (given variation in patient populations) and interpretability (making models clinically transparent). Importantly, reliance on subjectively defined complications (e.g., ascites, encephalopathy) risks inconsistency. Future progress may depend on objective surrogate markers in areas where AI could add value, such as electrocardiogram-based DL for portal hypertension[73], stool microbiota for hepatic encephalopathy[74], or neuroimaging for encephalopathy staging[75].

Assessment of transplant eligibility and organ donor quality

Psychosocial assessment is a key part of transplant evaluation, as it helps predict non-adherence, relapse, graft failure, and mortality[76]. The Stanford integrated psychosocial assessment for transplant tool is widely used to standardise this process[77]. Recently, Lee et al[78] developed an AI model using psychosocial variables to predict harmful alcohol use after liver transplantation, with a gradient boosting decision tree achieving superior accuracy (positive predictive value 0.82) compared to existing scoring systems.

The use of AI-enabled models is likely to increasingly support the rapid assessment of liver graft quality in the future. Recently, Cesaretti et al[79] explored this potential by using smartphone images to quantify graft steatosis. They developed a semi-supervised ML model based on an SVM-single instance learning system to detect steatosis levels (> 30%) and compared its results to the standard histopathological assessment. The model achieved 93% sensitivity in identifying steatotic grafts with an accuracy of 89%. This straightforward yet innovative method could enable consistent, real-time organ quality evaluations, aiding transplant surgeons during organ procurement, especially in remote locations where expert pathologists may not be available.

Post-transplant survival and graft performance

Given the shortage of high-quality donor organs, prioritizing recipients with favorable predicted post-transplant survival is increasingly important. Traditional scores like MELD, behaviorally anchored rating, and the donor risk index do not fully account for complex donor-recipient interactions. Machine-learning models overcome these limitations by integrating numerous variables without statistical assumptions and capturing nonlinear relationships. A systematic review found ML models to predict survival after liver transplant more accurately than conventional approaches, though evidence remains limited and direct comparisons to current clinical tools are scarce[80].

Since survival prediction is essentially a binary classification task, most studies have focused on conventional ML approaches. Hoot and Aronsky[81] reported that a Bayesian network model performed best for predicting 90-day mortality using 29 pre-transplant variables, though its validation achieved only modest accuracy (AUROC = 0.681). Similarly, Liu et al[82] applied a RF model based on blood test data. Both studies were limited by relatively small datasets. In contrast, the much larger scientific registry of transplant recipients, containing data from over 150000 United States liver transplant recipients, was used to develop a gradient boosting model that predicted survival up to 9 years post-transplant, although with only moderate accuracy (AUROC = 0.5983)[83].

A range of ML approaches have been used to predict post-transplant mortality, but their performance varies widely due to differences in cohort size, population, algorithms, and reliance on baseline rather than longitudinal data. Most models predict outcomes at fixed time points and inadequately handle follow-up or censoring. New survival-adapted ML methods may address these gaps. Robust, personalized predictions will require large, diverse, multi-center datasets and rigorous external validation. While AI offers promising decision support for transplant assessment, successful integration will depend on collaboration between clinicians and data scientists, ensuring models are accurate, interpretable, and clinically relevant[84].

ML models show strong potential for predicting graft failure, helping optimize donor-recipient matching and improve post-transplant outcomes. ANN-based models using early post-transplant clinical and laboratory data have achieved high accuracy (AUROC: 0.90-0.96). Additionally, a RF model combining key donor and recipient variables predicted 30-day graft failure with good performance (AUROC = 0.818)[85]. Moreover, de novo and recurrent liver disease after transplantation can impair graft function and lead to fibrosis. ML and DL tools show promise for early detection, with one ANN accurately predicting significant fibrosis in recipients with recurrent HCV, achieving 100% sensitivity and 79.5% specificity[86].

Transplant oncology

Transplanting the liver for primary or metastatic cancer is possible in carefully selected patients, but success depends on tumor burden, biology, staging, patient fitness, and donor quality. Post-transplant management must balance immunosuppression with vigilant cancer surveillance. Multiple prognostic tools, including the Milan criteria, MORAL score, AFP score, Toronto criteria and Metroticket 2.0, support candidate selection and risk stratification in transplant oncology[84]. AI offers the ability to integrate large datasets for individualized candidate assessment. ML applications using ANNs have already been developed in this field, including tools to predict tumor grade and microvascular invasion, estimate recurrence risk after LT, optimize waiting list allocation, and enable earlier detection of relapse[87].

A 2019 ML model based on the HCC liver database was able to predict 5-year post-LT survival for HCC with greater accuracy than both the MORAL and AFP scoring systems[87]. Bertsimas et al[88] created the optimized prediction of mortality ML model to predict a patient’s 3-month waitlist mortality, balancing efficiency and fairness in access to transplant[89]. This model showed a significant reduction in mortality when optimized prediction of mortality was used vs MELD and MELD-Na scores to determine order placement on the liver waiting list. Other groups have gone on to look at ML modelling for the evaluation of tumor gene expression and even quantitative evaluation of pre- and post-transplant imaging studies[90].

A study found that the CoxNet model achieved the highest concordance in predicting HCC recurrence using diverse clinical and laboratory data[91]. Another study applied a radiomics-based approach in LT recipients who had undergone transarterial chemoembolization (TACE) as a bridging therapy, extracting features from pre-TACE CT scans. A SVM model combining arterial and portal phase data predicted post-transplant recurrence with an AUROC of 0.81[92].

Beyond HCC, AI offers opportunities to build predictive models for a wide range of transplant-associated malignancies, including intra- and extrahepatic cholangiocarcinoma, metastatic colorectal cancer, neuroendocrine tumors, and paediatric hepatobiliary cancers. It can also support decision-making around the use of deceased, split, or living donor grafts in these scenarios. Overall, AI has the capacity to illuminate the complexities of the rapidly evolving field of transplant oncology and improve clinical precision[84].

Future directions of AI in transplant medicine

Advancing precision medicine in liver transplantation will depend on the ability to integrate multiple data streams, including clinical information, imaging, histopathology, and emerging omics technologies. At present, these data sources are siloed, largely due to the absence of standardized approaches for data harmonization. ML and DL provide the computational power needed to combine and interpret these complex datasets, potentially uncovering patterns that are not detectable through traditional analysis[93].

A major obstacle to the adoption of AI tools in clinical practice is the “black box” nature of many ML and DL models, which makes it difficult for clinicians and patients to fully understand how predictions are generated. Confidence in AI systems relies not only on accuracy but also on transparency and interpretability. Bias represents an additional challenge. Models inherit biases from the datasets on which they are trained, including racial, socioeconomic, and institutional disparities. If unaddressed, these biases may inadvertently influence clinical decisions. New debiasing methods, which account for confounding variables during model training, show promise in reducing algorithmic bias. Nevertheless, bias is highly context-specific and may vary across different healthcare settings, highlighting the need for continuous monitoring, recalibration, and validation[94].

A major limitation in liver transplantation is the scarcity of large, prospective, standardized datasets needed for robust AI development. Multi-center collection of granular clinical and biological data is essential to improve model accuracy and generalizability. Progress will require close collaboration between clinicians and data scientists. Ultimately, well-validated AI tools could enhance patient selection, prognostication, and personalized post-transplant care[95].

Barriers to the wide application of AI in hepatology

AI in hepatology faces several key barriers that limit widespread adoption. AI models are highly dependent on large, diverse, and representative datasets; however, like traditional clinical research, they remain vulnerable to bias, often amplified by high-dimensional data and complex DL architectural models[96]. Different types of bias can be recognized in AI models, which could potentially impact their reliability in clinical practice; spectrum bias arises when training data fail to reflect the full heterogeneity of real-world patients, leading to poor generalizability. This was demonstrated in a 90-day post-transplant mortality model that performed well within individual United States, United Kingdom, and Canadian registries but degraded when data were combined[97,98].

Another type of bias that was recognized is overfitting bias, which occurs when a model memorizes noise or local practice patterns rather than true clinical predictors, performing well on its training set but poorly in external validation. For example, a variceal bleeding model might learn center-specific endoscopy practices instead of genuine risk factors, limiting broader applicability[96,99]. Resource-limited settings create further obstacles, including limited digital infrastructure, lack of EHR systems, insufficient computational capacity, and workforce constraints, making it difficult to implement or validate AI tools equitably[100].

Dataset shift occurs when the relationship between a model’s inputs and outputs changes over time, leading to performance decline. In medicine, evolving clinical practices, patient populations, terminology, and AI-driven care can all create this shift. The example of MASLD shows how rapid changes, new terminology, new non-invasive diagnostic tools, and new Food and Drug Administration-approved treatments can make existing predictive algorithms inaccurate or obsolete unless their performance is continuously monitored and updated. The AI community recognizes the problem of performance decay caused by dataset shift, and strategies to maintain models are still being developed. Tools like statistical control charts can track changes in inputs, targets, and performance metrics. Determining when declines in performance meaningfully impact clinical outcomes and when to intervene depends on the model. “Locked” models can be recalibrated, retrained, or revised, though this may be costly. In contrast, continuous learning AI updates itself automatically based on new data[96]. In research studies, investigators should adopt a structured workflow to identify, assess, and mitigate dataset bias. First, dataset audits should summarize case-mix (age, sex, ethnicity, aetiology, disease stage, and centre) to detect systematic under-representation of key subgroups. Second, model performance should be reported not only overall but also stratified by clinically relevant subgroups, using metrics such as AUROC, calibration, and predictive values, in order to identify potential fairness gaps. Third, bias-mitigation strategies including targeted oversampling, sample re-weighting, careful removal of proxy variables for sensitive attributes, and the application of fairness-aware training approaches may help reduce inequities. Finally, rigorous external validation across independent centres and health systems is essential to ensure generalizability and avoid hidden performance failures when models are deployed in new populations[101-103].

Finally, the “black box problem” undermines clinician trust and hampers refinement of AI models. Unlike MELD, which is transparent and easily updated as shown by the transition to MELD-3.0 to reduce sex-based disparities deep models use vast, opaque variable interactions that make bias difficult to detect and correct[104-107]. This also leads to the Clever Hans effect, where models rely on irrelevant features such as sex in PBC prediction, risking inaccurate or inequitable decisions[108-110].

Ethical deployment of AI in hepatology also requires robust data governance and privacy protections, including secure de-identification, controlled data access, and clear consent processes for secondary use of clinical and imaging data. Transparency and explainability are critical for clinician trust; even when models are complex, techniques such as feature importance analyses, saliency maps, and case-based explanations can help clarify why a prediction was made. Responsibility for AI-supported decisions should remain with human clinicians, and institutions must define clear liability frameworks for model failures. In transplantation, there is a particular need to monitor whether AI-guided allocation or risk scores systematically disadvantage specific demographic or socioeconomic groups and to ensure that these tools are used to reduce, not amplify, existing inequities[101,102,111].

Translating AI models into clinical practice

For hepatology AI tools to move beyond proof-of-concept, a structured translational pathway is required. After initial model development and internal validation on retrospective data, robust external validation across centres and scanners should be performed. The next phase typically involves prospective observational studies where model outputs are generated in real time but do not yet influence management. If performance remains acceptable, randomized or pragmatic trials comparing AI-assisted vs standard care can assess clinical benefit, safety, and workflow impact. Successful systems must then undergo regulatory review as software-as-a-medical-device, followed by careful integration into EHRs, endoscopy, or radiology platforms, with clinician education and user-interface optimization. Finally, continuous post-deployment monitoring and recalibration are essential to detect drift and maintain performance over time[30,31,95,96].

CONCLUSION

AI is reshaping hepatology by improving diagnosis, risk stratification, and personalized treatment. By integrating clinical data with imaging, histology, and molecular information, AI can uncover patterns beyond routine assessment and may reduce the need for invasive testing. Yet major hurdles remain. Many models are developed in selective cohorts, are vulnerable to bias and overfitting, and often perform poorly when applied in new clinical settings. Looking ahead, the field should prioritize further research on robust multi-center validation of promising tools, the responsible development of multimodal algorithms that combine imaging, histology, genomics, and EHR data, and the design of interpretable systems that communicate predictions and uncertainty clearly to clinicians. Equally important will be the creation of scalable approaches for resource-limited settings and the establishment of international frameworks for data sharing and benchmarking. Addressing these priorities will be essential for the safe, ethical, and sustainable integration of AI into everyday liver practice, with the ultimate goal of improving outcomes for patients with liver disease.