Artificial intelligence in metabolic dysfunction-associated steatotic liver disease: Transforming diagnosis and therapeutic approaches

Abstract

Metabolic dysfunction-associated steatotic liver disease (MASLD) is an increasingly prevalent condition associated with hepatic complications and cardiovascular and renal events. Given its significant clinical impact, the development of new strategies for early diagnosis and treatment is essential to improve patient outcomes. Over the past decade, the integration of artificial intelligence (AI) into gastroenterology has led to transformative advancements in medical practice. AI represents a major step towards personalized medicine, offering the potential to enhance diagnostic accuracy, refine prognostic assessments, and optimize treatment strategies. Its applications are rapidly expanding. This article explores the emerging role of AI in the management of MASLD, emphasizing its ability to improve clinical prediction, enhance the diagnostic performance of imaging modalities, and support histopathological confirmation. Additionally, it examines the development of AI-guided personalized treatments, where lifestyle modifications and close monitoring play a pivotal role in achieving therapeutic success.

Key Words: Metabolic dysfunction-associated steatotic liver disease; Artificial intelligence; Machine learning; Deep learning; Ultrasonography; Digital pathology; Hepatocellular carcinoma; Precision medicine

Core Tip: Artificial intelligence (AI) is redefining the clinical approach to metabolic dysfunction-associated steatotic liver disease (MASLD). In diagnosis, it enhances the detection of steatosis and fibrosis beyond the limits of conventional tools. For prognosis, AI accurately stratifies risk and anticipates complications, consistently demonstrating superior performance. In treatment, it enables personalized interventions and accelerates drug development. By integrating multimodal data, including clinical, imaging, histopathological, and molecular information, AI transforms fragmented data into actionable insights, establishing itself as a cornerstone for the future of MASLD management.

INTRODUCTION

Metabolic dysfunction-associated steatotic liver disease (MASLD) is the leading cause of chronic liver disease worldwide, and its prevalence is steadily rising[1]. The nomenclature shifted from nonalcoholic fatty liver disease (NAFLD) to MASLD in 2023 to better reflect the pathophysiology of the disease and its strong association with cardiometabolic syndrome[2]. MASLD is now a major public health concern, with an estimated global prevalence of 38.8%. This prevalence varies geographically (e.g., 55.3% in Europe, 33.6% in Asia, and 36.0% in North America) and according to the diagnostic method used[3,4]. In high-risk populations with comorbidities such as diabetes mellitus and obesity, rates increase substantially reaching 65%-75%[3,4].

MASLD is defined by hepatic steatosis in the presence of at least one cardiometabolic risk factor[5]. The disease encompasses a spectrum from simple steatosis to steatohepatitis, fibrosis, and cirrhosis[6], which can lead to serious complications[7]. The most critical outcomes are cirrhosis and hepatocellular carcinoma (HCC). An estimated 10%-25% of patients progress to advanced fibrosis or cirrhosis[6,8]. A European cohort study identified MASLD as a significant predictor of cirrhosis [hazard ratio (HR) = 4.73], with risk markedly higher in patients with steatohepatitis (HR = 22.67)[9]. Due to its high prevalence, MASLD is now the fastest-growing indication for liver transplantation in Western countries[10]. Consequently, the proportion of HCC cases attributable to MASLD has surpassed those linked to viral hepatitis or alcohol[11]. MASLD and its advanced form, metabolic dysfunction-associated steatohepatitis (MASH), are recognized as leading etiologies of HCC. MASLD-associated HCC accounts for up to 59% of cases[12], with 35% occurring in patients without cirrhosis[13]. Reported HRs for HCC are 3.15 for MASLD and 8.02 for MASH[9]. Although targeted surveillance strategies have been proposed, the cost-effectiveness of screening, particularly in patients without cirrhosis, remains debated[14].

The effect of MASLD extends beyond the liver due to shared pathogenic mechanisms, such as insulin resistance, endothelial dysfunction, systemic inflammation, and oxidative stress, that underpin its strong association with cardiovascular disease (CVD)[15]. MASLD is closely linked with type 2 diabetes, obesity, and atherogenic dyslipidemia[16,17]. Epidemiological studies consistently identify MASLD as an independent risk factor for major cardiovascular events, including coronary artery disease, myocardial infarction, stroke, and atrial fibrillation[18]. Consequently, CVD is the leading cause of death in patients with MASLD[16]. A meta-analysis reported an HR of 1.45 for overall cardiovascular events in patients with MASLD, which rose to 2.5 in those with MASH and advanced fibrosis, even after adjusting for other cardiometabolic risk factors[19].

Beyond hepatic and cardiovascular complications, MASLD is also associated with chronic kidney disease (CKD) and extrahepatic malignancies. A meta-analysis showed a moderate increase in the risk of stage ≥ 3 CKD over long-term follow-up (HR = 1.43), independent of traditional risk factors. Subsequent studies have confirmed even stronger associations in patients with advanced fibrosis[20,21]. Furthermore, MASLD is associated with a near doubling of the risk of extrahepatic cancers, particularly those of the gastrointestinal tract and also of the lungs, breast, and urogenital systems. This association persists even after adjustment for confounders such as age, obesity, diabetes, and smoking[6,14,22].

Given the updated diagnostic framework and growing recognition of the clinical and pathophysiological complexity of MASLD, innovative strategies are urgently needed to improve patient management. Current diagnostic and prognostic tools often lack accuracy, reproducibility, or accessibility, particularly for early identification of patients at risk of disease progression or extrahepatic complications. In this context, artificial intelligence (AI) has emerged as a promising approach capable of integrating and analyzing large, heterogeneous datasets to refine diagnosis, predict outcomes, and guide personalized treatment strategies. This review provides a comprehensive overview of the advances in AI applications for MASLD, highlighting their clinical implications, current limitations, and future directions in the pursuit of precision hepatology.

A narrative review was conducted using a non-systematic search in PubMed, Scopus, and Google Scholar. Search terms included “MASLD”, “NAFLD”, “artificial intelligence”, “machine learning”, “deep learning”, “radiomics”, “digital pathology”, “liver fibrosis”, and “hepatocellular carcinoma”. Articles were selected by the authors based on relevance, recency, and clinical impact, with priority given to publications from 2020 to 2025.

AI IN MEDICINE

AI refers to computational systems designed to simulate human cognitive processes, such as learning, reasoning, and decision-making, through the analysis of large and complex datasets. Its application in medicine dates back several decades, initially focused on expert systems and clinical decision support, and has since expanded with advances in machine and deep learning (DL). In contemporary medical practice, AI encompasses a wide spectrum of modalities, each with distinct mechanisms and utilities that extend from diagnosis to treatment personalization, enabling tailored implementation across diverse clinical contexts[23,24]. Table 1 summarizes the principal AI modalities currently applied in medicine, highlighting their core mechanisms and general clinical applications as a framework for understanding their potential in complex diseases. In MASLD, these modalities are increasingly applied to support early diagnosis, risk stratification, and longitudinal monitoring through the integration of clinical, imaging, histopathological, and molecular data[25], thereby fostering more accurate and individualized patient assessment (Figure 1). This integration represents a key step toward precision management in hepatology. The following sections describe the principal applications of AI across the comprehensive management of MASLD.

Figure 1 Artificial intelligence applications across the clinical spectrum of metabolic dysfunction-associated steatotic liver disease. The figure illustrates how artificial intelligence tools are integrated at key stages of metabolic dysfunction-associated steatotic liver disease (MASLD) management. A: In clinical diagnosis, supervised machine learning (ML) models (e.g., Random Forest, XGBoost, support vector machine, logistic regression) predict MASLD, metabolic dysfunction-associated steatohepatitis, and significant fibrosis using clinical and laboratory data; B: Imaging-based deep learning (DL) methods, such as convolutional neural network and ResNet, support the detection of steatosis and the staging of fibrosis; C: In histopathology, DL combined with digital pathology enables the reproducible quantification of hepatic features using whole-slide imaging; D: Prognostic models based on supervised and unsupervised ML predict hepatocellular carcinoma, major adverse cardiovascular events, renal progression, and mortality; E: In treatment Large Language Model personalize lifestyle interventions. ML and DL facilitate novel drug discovery by integrating clinical, omics, imaging, and free-text data. Figure created by the authors using royalty-free and open-license content. Components were obtained from Servier Medical Art (CC BY 4.0), NIAID BioArt Source (free for educational and scientific use), and SciDraw (CC BY 4.0), with original creators credited in the source files.

Table 1 Types of artificial intelligence, application models, and their utility in medicine.

Type of artificial intelligence		Basic functioning	Medical utility
ML		Algorithms that learn patterns from structured data	Predictive diagnosis, risk analysis, disease classification, support for clinical decision-making, and selection of relevant variables in large clinical datasets
Classical ML methods	RF	Learning algorithm based on the construction of multiple DT, incorporating randomization to improve accuracy
	GBM	Technique that sequentially trains multiple DT, correcting the errors of the previous tree using gradients
	XGBoost	Optimized version of GBM that incorporates regularization, tree pruning, and parallel processing to enhance speed and performance
	SVM	Algorithm that identifies the optimal hyperplane that separates classes by maximizing the margin
	Logistic regression	Linear statistical model that estimates the probability of a binary event using the logistic function
DL		A subtype of ML that uses deep neural networks to process large volumes of data in order to identify patterns	Interpretation of unstructured data, including medical images (radiology and histopathology), omics and genomic data, and clinical text. It also facilitates information extraction from medical records and supports predictive analytics for clinical outcomes
DL subtypes	CNN	Uses convolutions to detect spatial patterns in structured data. Comprised of convolutional and pooling layers
	Transformer	Model based on attention mechanisms that enables parallel processing of entire sequences, capturing complex relationships among words or data
	MLP	Feedforward neural network with one or more hidden layers. Each neuron applies a nonlinear activation function to learn complex representations
NLP		Algorithms that comprehend and process human language in clinical texts	Extraction of information from electronic health records, analysis of medical notes, medical chatbots
Unsupervised learning		Algorithms that identify patterns or groupings in unlabeled data, capable of detecting subclusters, outliers, or low-dimensional data representations	Detection of disease subtypes, clustering of patients with similar clinical profiles
Reinforcement learning		Algorithms that learn through trial and error using feedback	Optimization of personalized treatments, sequential decision-making, such as drug dosing

Data adapted from Yu et al[23], Lidströmer et al[24], and Žigutytė et al[25]. ML: Machine learning; RF: Random Forest; DT: Decision trees; GBM: Gradient boosting machine; XGBoost: Extreme gradient boosting; SVM: Support vector machine; DL: Deep learning; CNN: Convolutional neural network; MLP: Multilayer perceptron; NLP: Natural language processing.

AI APPLICATIONS IN THE DIAGNOSIS OF MASLD

Current diagnostic approach

The current diagnostic approach for MASLD follows a sequential process. Since most patients are asymptomatic or present with nonspecific symptoms such as fatigue or vague abdominal discomfort, detection relies on active case-finding in individuals with cardiometabolic risk factors, elevated liver enzymes, or incidental findings of steatosis on imaging studies[26]. Given that the degree of fibrosis is the primary predictor of adverse outcomes, current diagnostic algorithms prioritize its early detection[27]. To this end, a sequence of non-invasive tests is recommended, starting with biochemical markers to estimate the risk of fibrosis, which may be complemented by radiologic assessments of liver mechanical properties to quantify stiffness. Finally, in selected cases or when non-invasive test results are inconsistent, a liver biopsy may be performed to confirm MASLD or rule out other differential diagnoses[1].

AI in the integration of biomarkers and clinical data

Currently, the use of biomarkers and various scoring systems constitutes the first step in predicting steatohepatitis and fibrosis due to their accessibility. Among the most commonly used tools are the fibrosis-4 index (FIB-4), the NAFLD fibrosis score, and the aspartate aminotransferase-to-platelet ratio index (APRI), which have demonstrated moderate to good performance, especially in detecting advanced fibrosis or cirrhosis. However, these tools have notable limitations, including low positive predictive value, lack of consensus on cutoff thresholds, and limited validation across diverse populations[28]. Moreover, relying solely on a small number of clinical variables may be inadequate to address a complex pathophysiological process that requires a multivariable approach[29].

In response to these limitations, multiple machine learning (ML) algorithms have been developed to integrate a wide range of clinical information, including demographic variables (e.g., age, sex, race, geographical region), electronic health record data (diagnostic codes, clinical notes, unstructured data), and serum biomarker levels. These models aim not only to detect MASLD or MASH but also to estimate the degree of steatosis and fibrosis[30]. Algorithms such as Random Forest (RF), Extreme gradient boosting (XGBoost), and neural networks have shown high accuracy in predicting steatosis and significant fibrosis, often outperforming traditional scoring systems[31].

For example, a study by Yu et al[32] evaluated various ML models using data from the National Health and Nutrition Examination Survey (NHANES) cohort (n = 13436), incorporating demographic variables, comorbidities, body composition, and laboratory tests. Most models demonstrated excellent predictive performance [area under the receiver operating characteristic curve (AUROC) > 0.80], with RF achieving the highest accuracy [AUROC: 0.928; 95% confidence interval (CI): 0.918-0.937] and external validation. This model outperformed conventional clinical indices such as the Framingham Steatosis Index, Fatty Liver Index, and Hepatic Steatosis Index, which all showed AUROCs below 0.75. Similarly, a study conducted within the United States Veterans Affairs health system involving over 4 million patients reported an AUROC of 0.83 for MASLD prediction using the RF model. Notably, the model identified 514997 at-risk patients who had previously been classified as healthy[33].

Regarding fibrosis detection and staging, McTeer et al[34] utilized data from the European NAFLD Registry and employed an XGBoost model incorporating routine clinical variables and basic laboratory parameters. The model achieved AUROCs of 0.89 for MASH risk, 0.85 for significant fibrosis, and 0.99 for cirrhosis. The inclusion of non-routine variables (e.g., ferritin, FibroScan, or protein levels) did not significantly improve the model performance. Similarly, Liu et al[35] evaluated the performance of an XGBoost model in detecting MASH and significant fibrosis (≥ F2), reporting AUROCs of 0.670 (95%CI: 0.530-0.811) for MASH and 0.713 (95%CI: 0.611-0.815) for fibrosis, both superior to traditional scores such as APRI, FIB-4, and the NASH index, which exhibited limited predictive power.

The number of studies exploring AI-based prediction of MASLD using clinical variables and biomarkers has increased significantly in recent years. Table 2 summarizes the most recent studies focused on this application[36-44].

Table 2 Studies on artificial intelligence for predicting metabolic dysfunction-associated steatotic liver disease, steatohepatitis, and fibrosis based on clinical data.

Ref.	Sample size	Machine learning type	Comparator	Reference standard	Classification categories	Model performance	Additional information
Qin et al[36], 2023	n = 14439 general population	SVM; RF	None	Color Doppler ultrasound (3.5-MHz, expert-interpreted)	MASLD diagnosis	AUC: SVM 0.85, RF 0.852; Acc: SVM 0.81, RF 0.78
Dabbah et al[37], 2025	Training: n = 618 MASLD; Validation: n = 540	XGBoost	FIB-4; NFS	Elastography ≥ 9.3 kPa/Biopsy ≥ F3	Advanced fibrosis	AUC 0.91; Sen 91%; Spe 76%	AUC; FIB-4 0.78; NFS 0.81
Nabrdalik et al[38], 2024	n = 2000 with DMT2	MLR	None	Ultrasonography plus metabolic criteria	MASLD diagnosis	AUC 0.84; Sen 75%; Spe 79%	Unsupervised ML was applied to identify a cluster of patients at high risk for MASLD
Njei et al[39], 2024	n = 5281	XGBoost	FIB-4; APRI; NFS; BARD	FibroScan-AST score (≥ 0.35/≥ 0.67)	High-risk MASH	AUC 0.95; Sen 82%; Spe 91%	AUC: FIB-4 0.50; NFS 0.54; BARD 0.39; APRI 0.50
Yang et al[40], 2024	n = 14913	LGBM; XGboost; RF	None	Transient elastography (CAP, LSM)	MASLD diagnosis	AUC; LGBM 0.90; XGboost 0.89; RF 0.89	The SHAP method was applied to enhance model interpretability
Boullion et al[41], 2025	n = 15560	RF	None	Transient elastography CAP ≥ 238 dB/m (steatosis)/LSM ≥ 7 kPa (fibrosis)	MASLD diagnosis Fibrosis	Acc; Steatosis: 79.5%; Fibrosis: 86.07%
Wakabayashi et al[42], 2025	n = 463	SVM; XGBoost; LR	FIB-4; APRI	Liver biopsy	Significant fibrosis (≥ F2)	AUC; SVM 0.88; LR 0.87; XGB 0.85	AUC: FIB-4 0.88; APRI 0.85
Xiong et al[43], 2025	Training n = 522; Validation n = 224	XGBoost	APRI; FIB-4	Liver biopsy	Advanced fibrosis	AUC 0.917	AUC; APRI 0.73; FIB-4 0.752
Zhu et al[44], 2025	n = 10007	LR; XGBoost	None	Transient elastography (CAP)	MASLD diagnosis	AUC; LR 0.79; XGBoost 0.79	The NHANES dataset was used as an external validation cohort

Acc: Accuracy; APRI: Aspartate aminotransferase-to-platelet ratio index; AST: Aspartate aminotransferase; AUC: Area under the curve; BARD: Body mass index, aspartate aminotransferase/alanine transaminase ratio, and diabetes score; CAP: Controlled attenuation parameter; DMT2: Type 2 diabetes mellitus; EMR: Electronic medical record; FIB-4: Fibrosis-4 index; LGBM: Light gradient boosting machine; LR: Logistic regression; LSM: Liver stiffness measurement; MASH: Metabolic dysfunction-associated steatohepatitis; MASLD: Metabolic dysfunction-associated steatotic liver disease; ML: Machine learning; MLR: Multivariate logistic regression; NFS: Nonalcoholic fatty liver disease fibrosis score; NHANES: National Health and Nutrition Examination Survey; RF: Random Forest; Sen: Sensitivity; SHAP: SHapley Additive exPlanations; Spe: Specificity; SVM: Support vector machine; XGBoost: Extreme gradient boosting.

AI in the detection of steatosis and fibrosis through imaging analysis

To date, the primary imaging modalities used in MASLD diagnosis include abdominal ultrasound, transient elastography (FibroScan^®), magnetic resonance imaging (MRI) - notably magnetic resonance elastography and proton density fat fraction (MRI-PDFF) - and computed tomography (CT)[45]. Ultrasound is the most commonly used initial tool due to its widespread availability and low cost, with good performance in detecting severe steatosis (AUROC 0.93)[46]. However, it has low sensitivity for detecting mild to moderate steatosis, with a detection threshold of approximately 20% fat infiltration, and is subject to significant operator dependence. Elastography allows for the estimation of liver stiffness, facilitating the assessment of fibrosis and enabling risk stratification. Among all imaging techniques, MRI is the most sensitive and specific for detecting steatosis, with MRI-PDFF in particular being the most accurate, achieving 100% sensitivity and 96% specificity. However, its limitations include lower availability, high cost, and susceptibility to artifacts, especially in the presence of hepatic iron overload[45].

Due to its potential, recent studies have increasingly focused on ultrasound techniques enhanced by AI, which have shown significantly improved diagnostic performance for MASLD, allowing for more precise quantification of steatosis, improved fibrosis detection, and a more detailed assessment of inflammation[47,48]. The most frequently used AI models include convolutional neural networks (CNNs) and support vector machines, although other ML algorithms have also been applied[31]. Notable studies include one by Chou et al[49], who analyzed 21,855 B-mode ultrasound images from 2070 patients using a CNN model (ResNet-50 v2) to evaluate and classify steatosis, achieving AUROCs of 0.947, 0.971, and 0.98 for mild, moderate, and severe steatosis, respectively. Another study by Han et al[50] used MRI-PDFF-based steatosis quantification in 204 participants as the reference and applied a unidirectional CNN model to ultrasound images. The model achieved an AUROC of 0.98 (95%CI: 0.94-1.00) for diagnosis and a Pearson correlation coefficient of 0.85 (P < 0.001) for quantitative estimation of hepatic fat fraction.

Similarly, studies focused on liver fibrosis classification have reported promising results. For example, Lee et al[51] applied a CNN model to B-mode ultrasound images, achieving 76.4% accuracy in staging fibrosis (F1 to F4) and an AUROC of 0.857 for diagnosing cirrhosis (F4), outperforming five radiologists. A recent meta-analysis concluded that AI-assisted ultrasound had a sensitivity of 0.97, specificity of 0.98, and an overall AUROC of 0.98[52]. Available evidence suggests that various DL models applied to ultrasound achieve a diagnostic accuracy of 80% or higher for MASLD detection, with robust consistency across studies[31]. Recent reviews have confirmed that AI applications in ultrasound image analysis improve overall performance, reduce interobserver variability, and enhance the early detection of steatosis[47].

AI integration into other imaging modalities is also under active development and has shown promising results[53]. In the case of CT, some DL models focused on fibrosis detection have reported AUROCs above 0.95 for staging ≥ F2 fibrosis, with strong correlation to histopathological findings[54]. In MRI, studies combining AI with radiomic analysis have demonstrated good performance in detecting various stages of fibrosis and identifying signs of portal hypertension non-invasively[55]. The application of DL to elastography has also improved diagnostic accuracy; one study reported AUROCs of 0.85 for ≥ F2 fibrosis and 0.97 for F4, although these data were obtained in a population with conditions other than MASLD[56].

AI support in histopathological analysis

Histopathological analysis remains the gold standard for diagnosing MASLD and differentiating simple MASLD from steatohepatitis (MASH). It allows direct evaluation of several key parameters: (1) Steatosis, which must be present in more than 5% of hepatocytes to confirm MASLD; (2) Lobular inflammation, characterized by infiltration of lymphocytes, neutrophils, eosinophils, and Kupffer cells; (3) Ballooning, which reflects hepatocellular injury and presents as hepatocyte edema - both inflammation and ballooning are essential for MASH diagnosis; and (4) Fibrosis, which, although not a diagnostic criterion, carries significant prognostic value[57,58].

The most commonly used assessment tool is the NASH Clinical Research Network (NASH CRN) system, which grades the severity of steatosis, lobular inflammation, and ballooning, as well as the degree of fibrosis. The NAFLD Activity Score is also widely applied, particularly in clinical trials, where a score ≥ 5 typically correlates with MASH diagnosis[59]. However, these systems are limited by considerable inter- and intra-observer variability, even among expert pathologists, as well as by the inherent heterogeneity of liver biopsies, which can complicate interpretation[60,61].

In this context, the integration of AI with digital pathology has emerged as a promising approach to improve diagnostic accuracy, reproducibility, and efficiency[62]. The most widely used AI models are based on DL and utilize data from whole-slide imaging (WSI), which enables the scanning and analysis of entire histological slides[58]. These models can be employed to detect, localize, quantify, and assign scores to various histopathological parameters[60].

Initial studies using murine liver samples demonstrated that DL algorithms could accurately detect both microvesicular and macrovesicular steatosis, with acceptable performance and good correlation with expert pathologist evaluations[63,64]. Follow-up research on human liver tissue further validated the utility of AI for assessing steatosis and fibrosis. For instance, Munsterman et al[65] developed a model that identified the steatotic proportional area in WSI with an AUROC of 0.970 (95%CI: 0.968-0.973). Similarly, Gawrieh et al[66] evaluated fibrosis grades in trichrome-stained digital images using automated collagen area quantification, achieving AUROCs ranging from 0.78 to 0.90 for identifying different fibrosis stages.

Recently, AI advancements have led to the development of comprehensive models capable of simultaneously evaluating all components of the NASH CRN system. One such model, AIM-MASH, was trained using expert pathologist annotations for all four scoring components. Analytical validation showed high repeatability and reproducibility, with values ranging from 0.93 to 0.96, exceeding those typically achieved by human observers[67].

Forlano et al[68], employed ML techniques combined with image processing to quantify steatosis, inflammation, ballooning, and fibrosis in a cohort of 246 patients, divided into training and validation sets. The model was compared with the assessments of two hepatobiliary pathologists, yielding intraclass correlations ranging from 0.92 to 0.97 across all parameters. Additionally, it demonstrated intra- and inter-observer agreement of 0.95 and 0.99, respectively, compared to traditional semi-quantitative methods, which achieved values ranging from 0.58 to 0.88. These findings highlight AI’s considerable potential to support pathologists by standardizing criteria and improving the reproducibility of diagnostic assessments.

AI IN THE PREDICTION AND DETECTION OF MASLD COMPLICATIONS

AI has also shown promising utility in predicting long-term complications associated with MASLD. Notably, it has proven effective in anticipating the development of HCC, even in patients without established cirrhosis. For example, Sarkar et al[69] compared five different ML models using clinical data from the electronic medical records of 2247 patients. The best-performing models - gradient boosting and RF - achieved 92.7% accuracy and an AUROC of 0.97 for HCC prediction. The most influential variables included FIB-4, total cholesterol, bilirubin, and arterial hypertension.

Another example is the HCC-Scope model, a neural network-based tool developed to detect early-stage HCC in patients with MASLD using demographic and basic laboratory data. The training cohort included 175 patients, with external validation in an independent sample of 55. This system demonstrated excellent performance, achieving AUROCs greater than 0.98 in both cohorts, outperforming widely used biomarkers such as alpha-fetoprotein and the GALAD score[70]. Additionally, DL-assisted ultrasound models have been developed for HCC detection and classification, demonstrating sensitivities above 95% in most studies[71].

Given the strong association between MASLD and CVDs, several studies have applied ML techniques to identify predictors of major adverse cardiovascular events and heart failure. For instance, Shibata et al[72] implemented a classification and regression tree model in a cohort of 2962 patients, identifying age over 60 years and the presence of four or more cardiovascular risk factors as primary predictors. Similarly, Nabrdalik et al[73] used ML models and found that atrial fibrillation, hyperuricemia, and glomerular filtration rate (GFR) were significant predictors of heart failure in patients with MASLD.

Some studies have also employed AI for automated analysis of fibrosis progression in histopathological samples and used unsupervised learning techniques to identify patient subgroups. These approaches have revealed that both the hepatic fibrosis stage and FIB-4 score are significantly associated with renal function decline, as measured by decreased GFR, and with increased all-cause mortality[74,75].

Additionally, an unsupervised ML analysis of the NHANES III cohort identified distinct clusters of patients with differentiated metabolic profiles. The cluster with the highest mortality risk exhibited elevated fasting glucose, systolic blood pressure, and body mass index, surpassing the risk levels observed in groups with average values or predominantly lipid-related abnormalities[76]. Integrating such variables into predictive models has further reinforced their utility as key prognostic markers for risk stratification in patients with MASLD.

ROLE OF AI IN TREATMENT

Current MASLD treatment primarily relies on lifestyle modifications, including dietary interventions, regular physical activity, and weight loss. In patients with metabolic comorbidities, pharmacologic agents such as glucagon-like peptide-1 receptor agonists and sodium-glucose cotransporter-2 inhibitors are used, which have demonstrated benefits in reducing hepatic steatosis, inflammation, and fibrosis, as well as improving cardiovascular outcomes. However, no pharmacological treatment has yet been approved explicitly for MASLD, although several therapies are currently in advanced stages of clinical development[1].

AI is emerging as a promising tool to optimize the effectiveness and sustainability of lifestyle interventions through its personalization capabilities. By generating targeted recommendations, guiding nutritional and behavioral adjustments, monitoring adherence to goals, and predicting individual responses, AI allows interventions to be tailored to each patient’s unique characteristics[77]. In this context, accessible tools such as ChatGPT - an AI-powered chatbot - have been evaluated for their ability to generate personalized diet and exercise plans tailored to the needs, preferences, and specific objectives of patients with MASLD[78].

For example, Ozlu Karahan et al[79] simulated 48 patients with MASLD and evaluated the diets generated by ChatGPT (GPT-4o version). While the recommendations met caloric and fiber requirements, they exhibited deficiencies, including a low carbohydrate content and excessive protein, fat, and saturated fatty acids, compared to clinical guidelines, highlighting the need for further refinement before widespread clinical application. Similarly, Dergaa et al[80] assessed ChatGPT’s physical activity recommendations, which were generally safe and showed progressive intensity but were limited in their specificity and personalization. Moreover, ChatGPT’s ability to respond to frequently asked questions from patients with MASLD regarding lifestyle, weight control, and referral to specialists has been evaluated, yielding favorable results in terms of accuracy, completeness, and comprehensibility, according to expert review[81].

The field of therapeutic research in MASLD has also undergone significant transformation with the integration of AI at multiple stages. In omics medicine, AI facilitates the analysis of large datasets from genomics, transcriptomics, proteomics, and metabolomics, enabling a deeper understanding of disease pathophysiology and progression, its relationship with the gut microbiome, and the identification of new therapeutic targets and response biomarkers[82].

In clinical research, AI-powered models such as AIM-MASH and qFibrosis enable continuous and objective quantification of key histopathological parameters, overcoming the limitations of traditional semi-quantitative scales. This enhances patient selection for clinical trials, allows for the exploration of new clinical endpoints, and provides a more accurate assessment of responses to both behavioral and pharmacological interventions[83,84]. For instance, Ratziu et al[85] conducted a post hoc analysis of a randomized controlled trial evaluating semaglutide in patients with MASLD. By analyzing baseline and 72-week digital biopsies using an ML-based continuous quantification model, they identified an antifibrotic effect of semaglutide that was not apparent with conventional histopathological assessments.

AI has also accelerated the development of new drugs through automated discovery platforms that model drug-receptor interactions, identify candidate compounds, and reduce the time and cost of preclinical research. A notable example is the search for farnesoid X receptor (FXR) agonists, a promising therapeutic approach in MASLD due to their capacity to reduce hepatic steatosis. Qin et al[86]. employed computational approaches and ML models to identify a new non-steroidal FXR agonist with a superior safety and efficacy profile, particularly in modulating lipid metabolism, compared to previous molecules.

CHALLENGES AND LIMITATIONS

Despite recent advances and the transformative potential of AI, its implementation in the diagnosis and treatment of MASLD still faces significant challenges. It is essential to improve the representativeness of the datasets utilized to ensure that AI algorithms are applicable beyond the initial training and validation phases. This requires the development of epidemiologically robust and diverse datasets that capture the full clinical and imaging spectrum of the disease. However, marked heterogeneity remains in the cohorts used to train these models, limiting their generalizability[87].

The successful adoption and systematic integration of AI into clinical practice depends on multiple critical factors. A major limitation is the “black box” problem - a lack of explainability inherent to many AI systems, particularly DL models. These models generate predictions through multiple hidden computational layers that are not easily interpretable, making it difficult to understand their internal decision-making processes. This opacity can reduce transparency, undermine clinical validation and human trust, and hinder overall comprehension of AI systems[88]. Another essential challenge is the standardization and quality of input data. Reliable AI performance depends on rigorous procedures for data collection, curation, and validation, processes that remain highly inconsistent across studies[87,89].

Beyond technical limitations, regulatory and ethical concerns must be addressed for the successful integration of AI in medicine. Unlike traditional drugs or devices, which follow established approval pathways, AI-based systems must demonstrate not only safety and efficacy but also transparency, accuracy, reproducibility, robustness, and validation across diverse populations. Currently, AI is primarily regulated under the framework of Software as a Medical Device. Regulatory agencies such as the United States Food and Drug Administration and the European Medicines Agency increasingly emphasize evidence of generalizability and require ongoing monitoring of real-world performance using a risk-based approach. Moreover, given the dynamic nature of AI technologies, continuous updates to regulatory agencies regarding changes in algorithm performance and input are necessary. From an ethical perspective, the use of AI requires strict safeguards for patient privacy, informed consent, and the protection of sensitive medical data. Advanced security measures are imperative to guarantee patient rights and confidentiality. These regulatory and ethical requirements underscore the need for tailored regulatory frameworks to ensure the safe and reliable integration of AI into MASLD management[87,90].

The effective integration of AI into clinical care requires substantial upgrades in healthcare infrastructure, including comprehensive digitalization of clinical units and advanced technological systems for data processing, storage, and connectivity. These upgrades demand significant financial investment, which may be particularly challenging in resource-limited settings. Although preliminary analyses suggest that AI could prove cost-effective by reducing unnecessary procedures and optimizing resource utilization, these benefits are highly context-dependent. Their interpretation must account for the specific healthcare setting, target disease, and outcomes measured[91].

Furthermore, it is essential to ensure that healthcare professionals are adequately trained - not only in the basic use of these tools but also in the critical interpretation of their outputs and understanding their benefits, limitations, and potential biases[87].

CONCLUSION

The integration of AI in the management of MASLD represents a disruptive advancement with the potential to radically transform diagnosis, risk stratification, monitoring, and treatment. ML and DL-based tools have demonstrated clear superiority over traditional methods in predicting steatosis and fibrosis, as well as in the early detection of complications. Furthermore, in therapeutic research, these tools have enabled more objective, reproducible, and continuous quantification of key histopathological parameters, thereby optimizing the design and evaluation of new interventions. However, widespread integration of AI into clinical practice requires overcoming significant technical, regulatory, and ethical challenges. Short-term priorities include incorporating explainable AI techniques to enhance transparency and build trust in clinical applications. Methods such as Local Interpretable Model-agnostic Explanations, SHapley Additive exPlanations, and attention-based mechanisms help clinicians identify which variables most influenced a given prediction. This makes AI outputs more interpretable and clinically actionable while empowering healthcare professionals to critically evaluate AI-supported decisions. Another key priority is standardizing data collection and validation processes, along with determining the most appropriate models for each phase of the clinical-therapeutic pathway. The development of multicenter, diverse, and representative datasets that encompass all disease stages and major comorbidities is essential. Such datasets will enable prospective studies to evaluate the real-world impact and cost-effectiveness of AI models on hard clinical outcomes, including mortality, fibrosis progression, and the onset of complications. This process must be accompanied by continuous evolution of regulatory frameworks that govern ethical data use and ensure robust security. In this context, federated learning emerges as a promising strategy for collaborative model development while preserving data privacy and protecting sensitive information. In the medium to long term, key milestones include achieving seamless integration of AI into healthcare systems with real-time functionality through robust technological infrastructure, ensuring equitable access in resource-limited settings, and strengthening clinical staff training in the use and critical interpretation of AI tools. Equally important is advancing basic research to develop adaptive, generalizable algorithms capable of delivering personalized predictions. Together, these challenges and opportunities position AI as a transformative paradigm in precision medicine, with immense potential to revolutionize clinical management of MASLD from risk stratification to treatment personalization.

ACKNOWLEDGEMENTS

The authors thank Dr. Diego Niño for his critical review and valuable input during the final revision of the manuscript.