Artificial intelligence in liver disease: Current status and future direction

Abstract

Chronic liver disease is a leading cause of death globally, primarily owing to liver cirrhosis and hepatocellular carcinoma. Early diagnosis and effective treatment are critical for curative therapy. By integrating imaging data, multiomics data, clinical test results, and electronic health records, artificial intelligence (AI) and machine learning algorithms are increasingly being developed to improve the diagnosis, prognosis, and treatment-related decision-making of liver disease. Notable AI models include CatBoost, ALADDIN (mAchine Learning ADvanceD fibrosis and rIsk MASH Novel predictor), GigaTIME, and METABOLISM. Additionally, AI supports donor organ quality assessment, outcome prediction, and the optimization of drug delivery and therapeutic efficacy. This mini-review summarizes the evidence of AI applications in enhancing liver disease diagnosis, prognosis, treatment, and research. However, many AI models remain difficult to interpret or explain and face challenges such as data bias and limited generalizability across regions. Furthermore, clinical integration of AI requires robust multicenter validation, data interpretability, ethical compliance, and adequate healthcare infrastructure.

Key Words: Artificial intelligence; Machine learning; Liver disease; Diagnosis and prognosis; Therapy

Core Tip: Chronic liver disease is a leading cause of disease-related mortality worldwide. Artificial intelligence, including machine learning and deep learning algorithms, is increasingly being applied to the diagnosis, prognosis, and prediction of treatment outcomes in chronic liver disease, to prevent progression to cirrhosis and hepatocellular carcinoma. Although limitations exist, integrating artificial intelligence into clinical workflows can help reduce errors and facilitate the extraction of critical information from large electronic health record datasets and complex diagnostic images.

INTRODUCTION

Liver disease is the 11^th leading cause of death worldwide, accounting for 2 million deaths annually. Liver disease-related mortality is mainly caused by cirrhosis and hepatocellular carcinoma (HCC)[1]. Chronic liver diseases, including alcohol-associated liver disease (ALD)[2], metabolic dysfunction-associated steatosis liver disease (MASLD)[3], and viral hepatitis[4], are common contributors to liver cirrhosis and HCC[5]. Surgical ablation is a curative treatment for early-stage HCC[6]. For patients with liver cirrhosis or early-stage HCC who are not eligible for surgical resection and for those with late-stage HCC, liver transplantation is the only curative option. However, it is not feasible for all patients owing to the associated high costs, liver disease-related comorbidities, limited donor organ availability, and other factors[7,8]. Therefore, early diagnosis of liver cirrhosis and HCC, as well as effective treatment of chronic liver diseases, are critically important for reducing liver disease-related mortality.

Recently, artificial intelligence (AI) techniques have been increasingly applied to the diagnosis and treatment of various diseases, including liver diseases[9,10]. These approaches include AI-assisted interpretation of diagnostic images, machine learning (ML) algorithms for predicting disease progression, and large-scale data analysis. Proceeding step by step from diagnosis to prognosis and then to treatment and research methodologies, this mini-review provides a comprehensive overview of AI applications in liver diseases, which focus on imaging and digital data analysis, novel biomarker prediction, and emerging algorithms. AI and ML algorithms have also been reviewed for predicting treatment outcomes in liver disease, followed by their applications in liver disease research.

AI APPLICATIONS IN LIVER DISEASE DIAGNOSIS

In this section, we discuss AI applications in the diagnosis of the most prevalent chronic liver diseases, including MASLD or its progressive form, metabolic dysfunction-associated steatohepatitis (MASH), liver fibrosis, and HCC.

MASLD and MASH

MASLD, previously known as non-alcoholic fatty liver disease (NAFLD), has become the most common type of chronic liver disease and can progress to liver cirrhosis and HCC. The diagnosis of MASLD typically involves blood tests, imaging scans, and histological analysis of liver biopsy samples[11]. Hematological markers in the blood or serum are commonly applied to assess liver function by measuring enzyme levels such as alanine aminotransferase (ALT) and aspartate aminotransferase (AST), metabolic markers such as A1c and lipids (cholesterols and triglycerides), and serologic tests to exclude viral hepatitis. Imaging techniques provide noninvasive methods for accessing hepatic fat and fibrosis, including ultrasound-based elastography and magnetic resonance imaging (MRI)[11].

With the increasing availability of multiomics data, researchers have developed ML-based approaches for biomarker discovery to improve the diagnosis of MASLD or MASH. Using multiomics data, Zhang et al[12] applied the least absolute shrinkage and selection operator (LASSO) regression method to establish a Triglyceride-Glucose Index × A Body Shape Index (ABD-LTyG) predictive model for diagnosing high-risk MASH (Table 1). This model incorporated markers including AST, body mass index, total bilirubin, vitamin D, triglyceride-glucose index, and biomarkers such as lectin galactoside-binding soluble 3 binding protein and triggering receptor expressed on myeloid cells 2 (TREM2) identified from single-cell RNA sequencing data[12].

Table 1 Examples of artificial intelligence applications in liver disease diagnosis.

Disease	Model	Algorithms	Advantages	Disadvantages	Ref.
MASH	ABD-LTyG	LASSO regression	Multiomics data (proteomics and transcriptomics); outperforms traditional assessments such as FIB-4, NFS, and APRI	Reliance on specific biomarkers such as TREM2, IL18BP, and LGALS3BP; validation in a relatively small cohort; complex components	Zhang et al[12], 2025
Fibrosis	ALADDIN	Random forest, gradient boosting machines, and XGBoost	Uses common laboratory parameters with or without VCTE, superior to FIB-4, SAFE, and LiverRisk scores, and noninferior to the FAST Score; includes a user-friendly web-based calculator	Data interpretation lacks expert consensus; limitations in extrapolation of this model to all patients with MASLD	Alkhouri et al[22], 2026
Fibrosis	LiverAID	Basic ML algorithms, including random forest, elastic net, bagging classification trees, and support vector machine	Effective for ruling out significant liver fibrosis; outperforms conventional blood-based indices, including FIB-4, Forns index, and APRI; validated in a large population	Regional study; internal validation only	Blanes-Vidal et al[27], 2022
HCC	PLAN-B	Gradient-boosting machine (GBM)	Validated in two independent cohorts with large, multicenter populations; provides better discrimination than previous models (PAGE-B, modified PAGE-B, REACH-B, and CU-HCC)	Limited generalizability across ethnicities and viral genotypes; potential selection bias; lacks metabolic risk factors and liver stiffness data	Kim et al[31], 2022

ALADDIN: MAchine Learning ADvanceD fibrosis and rIsk MASH Novel predictor; APRI: Aspartate aminotransferase to platelet ratio index; FAST: FibroScan-aspartate aminotransferase; FIB-4: Fibrosis-4; GMB: Gradient-boosting machine; IL18BP: Interleukin-18 binding protein; LGALS3BP: Galectin-3-binding protein; LASSO: Least absolute shrinkage and selection operator; NFS: NAFLD fibrosis score; SAFE: Steatosis-associated fibrosis estimator score; TREM2: Triggering receptor expressed on myeloid cells 2; VCTE: Vibration-controlled transient elastography; XGBoost: EXtreme Gradient Boosting; PAGE-B: Platelet, age, gender–hepatitis B; REACH-B: Risk estimation for hepatocellular carcinoma in chronic hepatitis B; CU-HCC: Chinese University hepatocellular carcinoma.

Rusu et al[13] identified enrichment of extracellular matrix (ECM) gene expression based on differentially expressed gene analysis of liver transcriptomic data, as well as complement-related proteins from circulating proteomic data. This group further developed an ML model demonstrating that proteins involved in the complement cascade, apolipoproteins and lipoproteins, and coagulation-related proteins are potential markers of MASH[13]. Cao et al[14] applied two ML methods, including LASSO and support vector machine recursive feature elimination, to refine efferocytosis-related biomarkers (e.g., TREM2 and T cell immunoglobulin and mucin domain-containing 4), thereby improving the diagnosis of MASH. Xu et al[15] developed an ML-based stepwise diagnostic strategy to identify high-risk MASLD using a random forest model to predict liver fibrosis stage (F ≥ 2), followed by a multilayer perceptron to identify patients with liver fibrosis (F ≥ 2) and MASLD (NAFLD Activity Score - NAS ≥ 5). In addition, Stuart et al[16] developed a natural language processing algorithm to analyze electronic health records to identify patients with MASLD, achieving a positive predictive value exceeding 93%.

Liver fibrosis

Liver fibrosis represents both an injury process and a repair response to chronic liver diseases and can progress to liver cirrhosis and HCC. Noninvasive tests, such as the fibrosis-4 score (FIB-4)[17], the NAFLD fibrosis score (NFS)[18,19], and the AST-to-platelet ratio[20], are commonly used to assess liver fibrosis. Using categorical gradient boosting machines (CatBoost), lightweight ML models incorporating predetermined features, such as aminotransferases, metabolic syndrome components, and body mass index, along with up to two additional variables (3-ureidopropionate and alpha-ketoglutarate), accurately detected MASH with liver fibrosis stages F2–F3[21]. These models outperformed conventional noninvasive tests, including NFS and FIB-4.

Alkhouri et al[22] developed the ALADDIN (mAchine Learning ADvanceD fibrosis and rIsk MASH Novel predictor) model to predict liver fibrosis (≥ F2) in patients with MASH using routine laboratory tests, with or without vibration-controlled transient elastography (VCTE). When combined with VCTE[22], the model outperformed VCTE alone, the FibroScan-AST score[23], and the Agile 3+ score[24], as assessed by area under the curve comparisons. Without VCTE[22], the model also demonstrated superior performance compared with FIB-4, the steatosis-associated fibrosis estimator[25], and the LiverRisk scores[26]. However, it has certain limitations (Table 1).

LiverAID models[27], which incorporate patient demographic information, clinical data, and serum markers as input parameters, demonstrated a superior ability to discriminate significant liver fibrosis compared with conventional blood-based indices, including FIB-4, the Forns index, and the AST-to-platelet ratio. The major limitation of these models is that data validation was performed internally. Yin et al[28] applied ML models using selected hepatic and splenic features in a computed tomography-based radiomics analysis for liver fibrosis staging, demonstrating superior performance compared with models using hepatic features alone. This model applied computed tomography-based AI techniques in staging liver fibrosis. However, this was a retrospective, single-center study with a limited number of patients, restricting the potential applicability of this model.

HCC

HCC is the most common type of primary liver cancer and can arise from multiple etiologies, including MASH, ALD, and viral hepatitis[29,30]. AI and ML techniques are increasingly used in liver cancer diagnosis. Kim et al[31] developed the Predicting Liver Associated Negative Biases (PLAN-B) model (Table 1) using a gradient-boosting machine algorithm to predict the development of HCC in Korean and Caucasian patients with chronic hepatitis B. This AI model outperformed previous models, including PAGE-B (age, gender, and platelet count), modified PAGE-B (age, gender, platelet count, and albumin level), REACH-B (age, gender, ALT, hepatitis B e antigen, and hepatitis B virus DNA), and CU-HCC (age, albumin, bilirubin, hepatitis B virus DNA, and cirrhosis)[31]. Lin et al[32] applied ML (XGBoost) to develop the Stiffness Measurement-Assisted Risk Tool for Hepatocellular Carcinoma score, using clinical features such as etiology, age, ALT, alkaline phosphatase, creatinine, platelet count, gender, hypertension, and liver stiffness measurements to stratify HCC risk across chronic liver diseases. Nishida et al[33] developed three convolutional neural network models using B-mode ultrasound image datasets of liver tumors, which reduced diagnosis errors in liver cancer in Japan.

AI APPLICATION IN LIVER DISEASE PROGNOSIS

Disease prognosis is critically important for guiding treatment decisions and care planning. AI offers transformative applications in this area. In this section, we present some examples. Feng et al[34] applied different ML algorithms, including ExtraTrees, XGBoost, LightGBM, and GradientBoosting, to predict HCC recurrence using clinical blood biomarkers, MRI features, and postoperative pathological data. By integrating serum heat shock protein 90α levels in patients with unresectable HCC treated with transarterial chemoembolization (TACE), AI models were developed to predict overall survival, with the StepCox (forward) plus random survival forest model demonstrating better performance[35]. Wu et al[36] evaluated five ML models, namely Cox regression, LASSO, Decision Tree, Random Survival Forests, and XGBoost, to improve overall survival prediction accuracy. Among these, Random Survival Forests achieved the highest concordance index and outperformed the other models, as validated using receiver operating characteristic curves, decision curve analysis, and calibration curves[36].

ML algorithms such as random forest can be used to predict early mortality in patients with ALD- or MASLD-associated HCC. Serum markers, including serum albumin and C-reactive protein-to-albumin ratio, showed greater relative importance than traditional cancer staging systems, such as Barcelona Clinic Liver Cancer stage and TNM classification, for predicting early cancer mortality[37]. Suárez et al[37] developed AI-based models using digitized hematoxylin and eosin (HE)-stained slides to identify outcome-related molecular profiles associated with poor prognosis and microvascular invasion. These models employed a transformer-based multiple-instance learning algorithm within a five-fold cross-validation framework.

AI AND ML IN THE TREATMENT OF LIVER DISEASES

Clinical decision-making requires complex data analysis and specialized expertise. AI applications for analyzing complex data can improve clinical decision-making and reduce unintended objective errors[38]. With appropriate validation, AI applications in clinical practice can further improve decision-making accuracy[39].

However, many challenges remain. In this section, we review the roles of AI and ML in enhancing the treatment of liver diseases.

Treatment response prediction

Shi et al[40] developed a novel superpixel-based graph attention network (IMCSGAT) to predict responses to combination therapy with programmed death-1 blockade and TACE in patients with HCC. This model was developed based on high-dimensional imaging mass cytometry data, enabling the identification of spatial interactions among different cell types within the dynamic tumor microenvironment (TME). Imaging mass cytometry simultaneously quantified 40 protein expressions in tissues from 43 patients with HCC, increasing the k value of the model, which was determined by the number of superpixels in the cell image type. Therefore, the IMCSGAT model has relatively high computational complexity to accurately estimate interactions among different TME components[40].

Using this model, the study found that patients with HCC who responded to the combination treatment exhibited highly enriched cell interactions between macrophages and natural killer cells or T cells. These findings were further validated in a murine HCC model[40]. The IMCSGAT model also outperformed the Barcelona Clinic Liver Cancer stage and may aid in treatment selection during HCC progression.

In contrast, a systematic review and meta-analysis showed that AI-based handcrafted radiomics and deep learning models only have potential utility in predicting TACE outcomes in patients with HCC. However, no significant differences were observed between the two models in assessing treatment outcomes, even when clinical data were incorporated. This inconsistency may be caused by the heterogeneity among the included studies[41].

TME modeling

The tumor immune microenvironment (TIME) plays an essential role in cancer therapy. Recently, a multimodal AI framework, termed GigaTIME, was developed to convert hematoxylin and eosin (HE) staining images to virtual multiplex immunofluorescence images for population-based TIME analysis. GigaTIME was pre-trained on 40 million cells using paired HE and multiplex immunofluorescence slides (multi-channel protein profiling) from the same tissues. It enabled the characterization of the spatial and combinatorial protein activation and expression patterns. The GigaTIME model incorporates multiple biological markers to improve predictions of cancer staging, survival prognosis, and patient stratification. Using this approach, patients can be further stratified into different subgroups based on disease stage and survival outcome trajectories according to the identified biomarkers[42].

Improvement in nanoparticle-targeted delivery systems

A quantitative model has been developed to improve the delivery efficacy of tumor-targeted nanoparticles[43]. This advanced deep learning model incorporates more parameters to evaluate delivery efficacy compared with traditional linear regression models. It expands the integration of ML and AI with molecular dynamics simulations and physiologically based pharmacokinetic modeling in the field of nanomedicine[44].

Organ transplantation

AI has been applied to assess donor liver quality for transplantation[45]. Kue et al[46] found that AI model-identified hepatic steatosis on preimplantation liver frozen sections was positively associated with an increased risk of early allograft dysfunction, progression of liver fibrosis, and respiratory failure. The integration of AI into liver transplantation research enhances international collaboration and improves outcomes for diverse patient populations[47]. In addition, AI-generated synthetic liver transplant waitlist data from the United Network for Organ Sharing database can help protect patient privacy, address data scarcity in clinical hepatology, and accelerate research progress[48].

AI IN THE METHODOLOGY OF RESEARCH STUDIES

Extraction of scientific insights

With the rapid development and accumulation of multiomics data in the research field (e.g., the microbiome field), significant challenges have emerged in extracting scientific insights across diverse and heterogeneous databases. Park et al[49] developed METABOLISM, a microbiome-specialized AI large language model using Low-Rank Adaptation (LoRA)-based parameter-efficient training, to better understand microbiome–liver interactions and associated biological processes. The fine-tuned METABOLISM model improves the extraction of biological insights from large-scale scientific data and demonstrates superior performance compared with general models, such as Gemma-3-12B-IT and ChatGPT-4o[34].

Three-dimensional analysis of organoid

Three-dimensional (3D) cell culture systems mimic the structural and functional complexity of human and animal organs, providing valuable platforms for research studies[50]. However, high-resolution 3D imaging and analysis remain challenging for many researchers. Ong et al[51] developed an AI-based pipeline, named 3DCellScope, for screening 3D organoid morphology and topology. Such AI models facilitate the analysis of 3D images and help reduce objective errors.

CONCLUSION

AI bridges the gaps in integrating complex clinical data, precision medicine, and conventional biomarkers to improve the diagnosis, prognosis, and treatment of liver diseases (Figure 1). AI-based applications in hepatology require multimodal data integration, model interpretability, and rigorous validation to enhance performance[52]. AI is expected to accelerate future research directions. For example, by integrating anthropometric and genetic data with mitochondrial biomarkers, Longo et al[53] applied random forest algorithms to predict the risk of MASLD development. As the liver is a central metabolic organ, AI models have also been developed for the early detection of dyslipidemia using liver chemical shift-encoded MRI fat images[54].

Figure 1  Artificial intelligence bridges gaps in integrating complex clinical imaging data and conventional biomarkers to improve the diagnosis, prognosis, and treatment of liver diseases.

Despite the advantages of AI-based applications in liver disease, many challenges remain. Ethical considerations for AI applications in healthcare include the accuracy of diagnostic and prognostic methods, fairness in treatment, and protection of patient privacy[55]. Healthcare infrastructures, de-identified data sharing systems, independent multicenter validation studies, model generalizability, and ethical compliance heavily influence the clinical implementation of AI[56,57]. The application of multiple AI models can significantly improve diagnostic accuracy and generalizability by integrating imaging data, clinical features, and biomarkers. However, many AI models function as “black boxes”, limiting interpretability and eroding trust[58]. Therefore, improving transparency, explainability, and the training of medical professionals is important for advancing AI applications in healthcare[55,59].