Artificial intelligence applications for managing metabolic dysfunction-associated steatotic liver disease: Current status and future prospects

Abstract

The incidence and prevalence of metabolic dysfunction-associated steatotic liver disease (MASLD) have continued to increase in recent years, making it one of the most common chronic liver diseases worldwide. MASLD is highly comorbid with obesity, type 2 diabetes, cardiovascular disease, and chronic kidney disease, posing a serious threat to public health and creating a significant medical and socioeconomic burden. Despite advances in research, current clinical practice still faces considerable challenges in early screening, risk stratification, prognostic prediction, and long-term therapeutic monitoring. Recent advances in artificial intelligence (AI) have provided transformative opportunities to address these challenges. AI has demonstrated unique advantages in imaging interpretation, multiomics integration, electronic health record analysis, and remote health management, significantly improving the accuracy and efficiency of the noninvasive diagnosis, individualized risk stratification, precision therapy, and dynamic disease monitoring of MASLD. In this mini-review, the latest advances in AI applications for MASLD diagnosis and management are systematically summarized, and a forward-looking perspective on the role of AI in enabling the next generation of smart health care systems for MASLD is offered, with the aim of providing theoretical and practical guidance for the clinical management of this disease.

Key Words: Metabolic dysfunction-associated steatotic liver disease; Artificial intelligence; Multiomics integration; Early screening; Risk stratification; Disease monitoring; Machine learning; Clinical decision support

Core Tip: Artificial intelligence (AI) is rapidly advancing the management of metabolic dysfunction-associated steatotic liver disease (MASLD). Because AI can integrate imaging, multiomics, and electronic health record data, the use of AI enables earlier diagnosis, more accurate risk stratification, and personalized treatment. This mini-review summarizes recent progress and future directions for applying AI to improve MASLD care and outcomes.

INTRODUCTION

Metabolic dysfunction-associated steatotic liver disease (MASLD), previously known as nonalcoholic fatty liver disease (NAFLD), is a chronic liver disorder closely linked to metabolic syndrome[1]. The diagnostic criteria for MASLD have shifted away from the traditional “exclusion-based” definition of NAFLD, which requires ruling out significant alcohol consumption or other specific liver diseases, to a “compatibility-based” standard centered on the evidence for metabolic dysfunction. This shift emphasizes the positive association between MASLD and metabolic abnormalities, such as obesity, type 2 diabetes (T2D), hypertension, and dyslipidemia[1]. The global increase in MASLD prevalence is linked to the increasing rates of obesity, T2D, hypertension, and dyslipidemia, especially in the context of unhealthy lifestyles and increasingly sedentary behavior[2-4]. Currently, approximately 38% of adults and 7%-14% of children and adolescents are affected by MASLD, and the prevalence in adults is projected to exceed 55% by 2040[5,6]. In China, the increasing prevalence of MASLD is particularly rapid, with onset occurring at progressively younger ages[5,7,8]. Cardiovascular disease is the leading cause of death in MASLD patients, and MASLD itself is being increasingly recognized as an independent risk factor for cardiovascular events, in addition to its well-established role in metabolic syndrome[6,9-11].

Despite its growing burden, management remains challenging given the highly variable clinical course and only partially understood mechanisms of progression to metabolic dysfunction-associated steatohepatitis (MASH), fibrosis, and cirrhosis[8,12]. Routine screening relies on serologic markers, risk scores, and imaging [e.g., ultrasound, computed tomography (CT)], which have limited sensitivity for mild steatosis and early fibrosis and are affected by operator and equipment variability[13-15]. Elastography (e.g., FibroScan) enables the quantification of steatosis and fibrosis, but its accuracy decreases in patients with obesity, ascites, or advanced fibrosis[16]. Magnetic resonance imaging (MRI) provides precise noninvasive quantification of liver fat but is hampered by high costs and limited accessibility[14,15]. Although liver biopsy remains the “gold standard”, its invasive nature, cost, and sampling variability make it unsuitable for large-scale screening[17]. Given the considerable heterogeneity in metabolic background, clinical features, and disease progression among MASLD patients, the traditional “one-size-fits-all” management approach is increasingly inadequate, underscoring the need for more precise stratification, dynamic risk prediction, and comprehensive intelligent management tools.

Artificial intelligence (AI) offers unique advantages for handling high-dimensional, heterogeneous data, uncovering latent patterns, and generating individualized predictions[18-20]. Its roles across hepatology, encompassing imaging, histopathology, noninvasive tests, and predictive modeling, have also been recognized and emphasized[21,22]. Beyond imaging and laboratory data, AI can leverage longitudinal weight trajectories, medication histories, and large-scale electronic health records to identify high-risk individuals who might otherwise be missed by conventional screening approaches. This provides innovative solutions to the challenges of high heterogeneity, slow progression, and fragmented data inherent to MASLD management[23,24]. Empowered by AI, MASLD management is evolving toward a closed-loop model of “early identification, timely intervention, and continuous monitoring” (Figure 1), laying a robust foundation for intelligent liver disease management and offering a scalable template for future digital chronic disease care systems.

Figure 1 Artificial intelligence applications across the full management spectrum of metabolic dysfunction-associated steatotic liver disease. Diagnosis: Artificial intelligence (AI) models integrate biochemical indicators (alanine aminotransferase, aspartate aminotransferase, glucose, homeostatic model assessment of insulin resistance, triglycerides, high-density lipoprotein cholesterol) and multimodal imaging data (ultrasound, magnetic resonance imaging) for automatic detection and quantification of hepatic steatosis and fibrosis. Deep learning improves diagnostic accuracy, standardization, and efficiency, enabling non-invasive and early identification of metabolic dysfunction-associated steatotic liver disease. Risk stratification: Machine-learning algorithms combine clinical, imaging, and metabolic variables to predict the probability of advanced fibrosis, cirrhosis, or extrahepatic complications. AI-based risk maps provide individualized stratification for optimized referral and management pathways. Prognostic prediction: AI fusion models integrate longitudinal clinical and lifestyle data to forecast disease trajectories and long-term outcomes. Continuous learning frameworks (e.g., recurrent or transformer networks) enable dynamic prediction of fibrosis progression, hepatocellular carcinoma, and cardiovascular events. Therapeutic decision-making and treatment monitoring: AI-assisted systems analyze treatment responses and generate personalized lifestyle or pharmacologic recommendations. By incorporating wearable-device and mobile-health data, the platform enables real-time monitoring, adaptive feedback, and closed-loop management, improving adherence and long-term efficacy. AI: Artificial intelligence; ALT: Alanine aminotransferase; AST: Aspartate aminotransferase; GLU: Glucose; HCC: Hepatocellular carcinoma; HDL-C: High-density lipoprotein cholesterol; HOMA-IR: Homeostatic model assessment of insulin resistance; MASLD: Metabolic dysfunction-associated steatotic liver disease; MRI: Magnetic resonance imaging; TG: Triglycerides.

APPLICATIONS OF AI IN THE DIAGNOSIS OF MASLD

Given that MASLD is often clinically silent yet progressive, early detection and accurate risk stratification are critical for preventing fibrosis, cirrhosis, and hepatocellular carcinoma (HCC)[3,13]. Biopsy and traditional diagnostic approaches face the well-known constraints of invasiveness, scalability, and variability[17,25,26]. The rise of AI offers new avenues for noninvasive, efficient, and standardized diagnostics, especially in the realms of image interpretation, serological modeling, and multiomics data integration[24]. Automated, AI-driven systems for MASLD diagnosis and staging have been validated across multiple studies and cohorts and often outperform conventional tools in terms of accuracy, sensitivity, and specificity, especially when clinical, imaging, and histopathological data[24] are integrated into multimodal decision-support models[27-31].

Deep learning algorithms, particularly convolutional neural networks (CNNs), can perform automatic segmentation, feature extraction, and quantitative analysis of both imaging and histological data, substantially increasing the standardization and efficiency of MASLD diagnostics[32]. On ultrasound, the use of AI-based texture analysis improves the detection of mild-to-moderate steatosis and interobserver consistency over manual reads[33,34]. On CT, AI can quantify liver attenuation for opportunistic screening during routine or emergent scans[35]. Furthermore, multimodal residual networks support automated liver localization and region-of-interest (ROI) extraction, enabling high-throughput, scalable pipelines. With respect to MRI, the proton density fat fraction (PDFF) remains the noninvasive reference for fat quantification[32]. Furthermore, multimodal residual networks support automated liver localization and ROI extraction, enabling high-throughput, scalable pipelines. With respect to MRI, the PDFF remains the noninvasive reference for fat quantification[36]. Additionally, tools such as LiverColor utilize machine learning to analyze smartphone-captured liver images, greatly increasing the accuracy of steatosis quantification and offering innovative solutions in settings such as transplant assessment[37].

In terms of pathology, AI systems can automatically segment and quantify steatosis, ballooning, inflammation, and fibrosis networks on biopsy slides and improve interobserver agreement[38]. Translational work, such as an AI-based measurement (AIM) tool for scoring MASH (formerly known as nonalcoholic steatohepatitis) histology, shows excellent reproducibility and alignment with expert consensus for histologic endpoints in MASH trials and yields continuous scores that correlate with outcomes[39].

Because single-modality inputs rarely capture disease complexity, multimodal integration has become central to robust AI diagnostics. Compared with single-modality approaches, contemporary models that combine MRI-PDFF, magnetic resonance elastography (MRE)-derived stiffness, ultrasound elastography, laboratory markers [e.g., alanine aminotransferase (ALT)/aspartate aminotransferase (AST)], and lifestyle/behavioral data achieve significantly higher areas under the curves (AUCs) for differentiating MASLD from MASH and predicting advanced fibrosis (≥ F2)[40-42]. Deep learning architectures such as ResNet and Transformer can simultaneously process both structured (e.g., laboratory results) and unstructured (e.g., imaging) data, paving the way for comprehensive, closed-loop AI diagnostic platforms that span data collection, processing, modeling, and clinical decision support, ultimately advancing diagnostic automation and standardization[43]. Future platforms will incorporate explainable AI (XAI) (e.g., Grad-CAM) and causal modeling frameworks to increase clinicians’ trust in model reasoning and accelerate clinical adoption[24]. For instance, AI-driven non-contrast MRI models using T1WI and T2FS sequences have demonstrated high accuracy in staging hepatic fibrosis[44], and in animal studies, the combination of second-harmonic generation/two-photon excitation fluorescence imaging with AI systems has enabled dynamic, automated assessment of steatosis patterns in MASLD models[45].

APPLICATIONS OF AI IN RISK STRATIFICATION AND PROGNOSTIC PREDICTION FOR MASLD

MASLD is characterized by marked clinical heterogeneity, with disease progression trajectories shaped by a complex interplay of metabolic background, inflammatory status, and genetic factors[3,46]. Consequently, the development of precise and dynamic risk stratification and prognostic prediction models is essential for enabling individualized treatment, optimal resource allocation, and improved long-term outcomes in MASLD patients[47,48]. AI can integrate heterogeneous signals from clinical, imaging, and multiomics sources without manually predefined hypotheses, enabling automated discovery and weighting of risk factors and improving risk assessment accuracy and efficiency[21,22,49].

In the construction and validation of risk stratification models, AI typically employs feature selection and preprocessing [e.g., ALT, AST, platelet count, body mass index (BMI), homeostatic model assessment of insulin resistance, liver stiffness] with demographic and comorbidity information. Algorithms such as random forest, XGBoost, support vector machines, and neural networks are widely utilized for model training[50,51]. Among these, XGBoost has achieved the highest AUCs for MASLD subtyping and MASH prediction[23,50]. However, high AUCs do not guarantee robustness or transportability. For example, in the construction of the LiverRisk score for long-term liver outcomes, multiple algorithms (including XGBoost) were evaluated, yet a simpler linear model was ultimately selected because of its superior calibration and reproducibility[52,53], underscoring the risks of overfitting and dataset dependence and the need for external validation and calibration assessment[21,22]. Notably, NASHMap, a machine learning model based on 14 parameters, achieved areas under the receiver operating characteristic curve of 0.82 and 0.76 in the NIDDK and Optum datasets, respectively, and identified a substantial number of NASH cases that were missed in routine clinical practice[54].

For prognostic prediction, deep learning models using MRI-PDFF, T1 mapping, and elastography can automatically extract subtle hepatic features, enabling earlier risk identification than traditional visual assessment does[44,55]. Time series neural networks such as T-LSTM and Transformer-XL have further advanced the field by overcoming the limitations of static risk models, making it possible to predict risk continuously across different disease stages and support dynamic patient monitoring[56,57]. Multimodal AI fusion approaches can integrate transcriptomic, lipidomic, and metabolomic data with clinical, behavioral, and lifestyle information to help identify windows of elevated fibrosis risk and optimal intervention timing[58,59]. Moreover, the continuous input of real-world lifestyle data via wearable devices [e.g., smartwatches, continuous glucose (GLU) monitors, and smart scales] allows AI to detect metabolic imbalances and impending adverse outcomes at an early stage[60]. Overall, increasing model generalizability and leveraging multisource data integration are key to realizing the full potential of AI in MASLD risk stratification and prognosis.

AI-ASSISTED THERAPEUTIC DECISION-MAKING AND TREATMENT MONITORING IN MASLD

The management of MASLD is inherently long-term, dynamic, and highly individualized. Traditional treatment relies largely on lifestyle interventions and empiric pharmacotherapy, but the significant heterogeneity in treatment response and the lag in conventional monitoring often impede precise management[9]. The advent of AI has introduced new opportunities and tools for personalized therapeutic decision-making, outcome prediction, dynamic monitoring, and long-term disease management in MASLD[61-63].

Multimodal models that combine liver stiffness, MRI-PDFF, glycemia, lipids, BMI, and comorbidities can generate individualized response predictions and quantify how fibrosis modulates expected benefit, thereby informing choices that pertain to exercise, diet, and pharmacotherapy[27,64]. AI can also analyze pharmacogenomic data, such as CYP450 genotypes, to predict drug metabolism and adverse reactions, enabling dose individualization and a more accurate selection of suitable patients for novel therapeutics[65]. For dynamic management, AI is used not only to optimize the initial plan but also to adapt therapy over time. Data from wearable devices (steps, heart rate, sleep, etc.), when combined with digital dietary logs, enable the dynamic detection of physical inactivity or glycemic excursions, further strengthening individual health support[66]. With respect to treatment monitoring, AI overcomes the limitations of traditional intermittent biochemical or imaging assessments by providing more sensitive, continuous, and personalized outcome predictions[67].

Deep CNNs and other AI models can detect subtle changes in hepatic fat distribution and fibrosis on imaging, enabling the early prediction of treatment response; these models have been successfully applied in MRE and PDFF analyses for pharmacological efficacy evaluation. AI can also integrate longitudinal data on blood GLU, weight, and liver function for daily tracking of therapeutic efficacy, automatically generating clinical recommendations on the basis of trends in body weight and transaminases. By analyzing historical drug response data and genomic data, AI is also able to predict the risk of adverse drug reactions and refine monitoring priorities. Because MASLD patients often require long-term or lifelong follow-up, AI systems offer particular advantages in terms of remote monitoring and early warning. AI can generate adaptive follow-up schedules based on a patient’s therapeutic response, metabolic control, and lifestyle factors, shortening intervals for high-risk individuals and extending them for low-risk individuals, thus optimizing resource utilization[68]. AI-enabled remote health monitoring platforms now allow comprehensive tracking of weight, activity, diet, and mood, with abnormal trends triggering automated follow-up reminders or physician alerts, supporting a seamless “data collection – dynamic evaluation – AI feedback – remote intervention” closed-loop model. The incorporation of federated learning technology into multicenter follow-up platforms enables privacy-preserving joint modeling of follow-up data across institutions, facilitating standardization and traceability, while blockchain integration can increase the security and reliability of model updates. With long-term data tracking, AI can map individualized disease trajectories and trigger automated intervention or further evaluation whenever a transition to fibrosis or HCC is detected. Additionally, AI combined with natural language processing and large generative models (such as ChatGPT) can provide patients with personalized health education, dietary advice, and emotional support, substantially improving adherence and empowering proactive self-management[20,69,70].

AI-driven therapeutic decision-making and outcome monitoring are transforming MASLD management from passive, experience-based approaches to active, intelligent systems, laying a robust technical foundation for comprehensive, lifelong liver disease care. By bridging advanced predictive modeling with pragmatic decision-support tools, AI not only holds promise as a frontier technology but also as a practical clinical aid ready for near-term translation, laying the groundwork for comprehensive, lifelong, and multidisciplinary liver disease care for MASLD.

CHALLENGES AND LIMITATIONS OF AI IN CLINICAL PRACTICE FOR MASLD

Despite the tremendous promise of AI applications in MASLD – demonstrated by advances in early diagnosis, risk stratification, individualized therapeutic decision-making, and long-term follow-up – there remain significant barriers to widespread clinical implementation. Foremost among these is the persistent challenge of data quality and heterogeneity. Clinical records for MASLD patients frequently suffer from missing information, inconsistent labeling, and non-standardized data formats; differences in imaging equipment and protocols, variability in laboratory units, and gaps in follow-up data can directly compromise model training and performance[71,72]. In addition, limited model interpretability poses a serious obstacle to clinical trust and adoption. While many deep learning models outperform traditional methods in predictive tasks, their “black box” nature makes it difficult for clinicians to trace or understand the underlying reasoning – especially in key areas such as steatosis detection and fibrosis staging. When models lack clear justification for their predictions, physicians are less likely to rely on AI-supported decisions[73]. Although XAI techniques such as Local Interpretable Model-Agnostic Explanations and SHaply Additive exPlanation can visualize the contribution of variables, their effectiveness in managing complex multimodal data remains limited, and most AI systems still provide “results as conclusions” in a unidirectional fashion, which does not meet clinicians’ needs for stepwise reasoning and erodes trust in AI-assisted decision-making[74]. Ethical and privacy concerns are also prominent, as MASLD-related electronic medical records, imaging, and multiomics data are highly sensitive and subject to strict regulations. Another key limitation is the lack of high-quality prospective studies; the majority of AI research in this field still relies on retrospective datasets, which constrains the ability to make causal inferences and evaluate real-world effectiveness. In summary, only with advances in data standardization, improved model interpretability, robust privacy and ethical frameworks, greater diversity in training samples, and support from prospective studies can AI achieve broad, reliable clinical translation in MASLD.

CONCLUSION

In conclusion, the use of AI is driving the transformation of MASLD management from traditional experience-based approaches to a data-driven and intelligent precision medicine paradigm. At present, AI has significantly increased the efficiency and scientific rigor of key aspects of MASLD care, including early noninvasive screening, disease classification, risk prediction, individualized therapy, and longitudinal monitoring. The integration of multimodal data, real-time monitoring with wearable devices, and federated learning is laying a solid foundation for a closed-loop, intelligent management system that encompasses the entire disease course. However, the large-scale clinical adoption of AI in MASLD still faces challenges related to data heterogeneity, model interpretability, and ethical and privacy issues. Moving forward, it will be essential to build high-quality, multicenter, standardized data platforms, increase model transparency and explainability, establish comprehensive compliance frameworks, and ensure secure, collaborative data sharing across regions. Only then can AI progress beyond algorithmic proof-of-concept to achieve meaningful clinical application. As AI technologies continue to merge with MASLD research, they are positioned to become an indispensable engine for precision subtyping, adaptive disease management, and remote health interventions, ultimately facilitating the development of a smart, personalized, and lifecycle-oriented liver disease care ecosystem.