Artificial intelligence in metabolic dysfunction-associated steatotic liver disease: A short narrative review of diagnostic and therapeutic implications

Table 1 Artificial intelligence-assisted diagnosis in metabolic dysfunction-associated steatotic liver disease and related liver-disease settings

Ref.	Data/model	Reference standard and validation	Main result	Focused limitation
Corey et al[15]	Derivation cohort of 620 patients; adaptive LASSO logistic-regression algorithm using EHR narrative mentions, ICD-9 codes, and triglyceride levels; subsequently applied to 8458 individuals	Manual chart review with biopsy or clinical diagnosis as the reference standard	The combined algorithm achieved an AUC of 0.85, compared with 0.75 for ICD-9 codes alone, and identified an independent cohort of 8458 individuals with a high probability of historically defined NAFLD	Retrospective design; historical terminology; effects on clinical outcomes were not evaluated
Stuart et al[16]	NLP-supported EHR identification of 957 patients with MASLD over six months	Manual review of monthly cohorts; NLP identification of steatosis from abdominal imaging reports and structured assessment of MASLD criteria	Overall PPV exceeded 93%; the imaging-report NLP component achieved a PPV of up to 99.4%; only 140 of 957 patients (14.6%) had a MASLD-related diagnosis code	Single-health-system study; dependent on documentation and coding practices; no external validation
Lee et al[19]	Deep convolutional neural network trained on 13608 ultrasonography images from 3446 patients; internal test set of 266 patients and external test set of 572 patients	Pathological specimens or METAVIR scores estimated by transient elastography; comparison with five radiologists	Four-class METAVIR accuracy was 83.5% internally and 76.4% externally. Cirrhosis AUC was 0.901 internally and 0.857 externally; external performance exceeded that of the five radiologists	Not MASLD-specific; reference standards were heterogeneous; tertiary-centre transportability requires evaluation
Wang et al[23]	Prospective multicentre study of 398 patients and 1990 shear-wave elastography images from 12 hospitals; deep-learning radiomics model	Histological fibrosis staging in patients with chronic hepatitis B	AUCs were 097 for F4, 0.98 for ≥ F3, and 0.85 for ≥ F2; overall performance exceeded conventional two-dimensional shear-wave elastography and biomarkers for most fibrosis thresholds	Indirect evidence for MASLD because the study population had chronic hepatitis B
Forlano et al[24]	Machine-learning analysis of 246 liver biopsies: 100 in the derivation cohort and 146 in the validation cohort	Independent pathologist assessment using NASH Clinical Research Network criteria and manual histological annotations	In the derivation cohort, ICCs were 097 for steatosis, 0.96 for inflammation, 0.94 for ballooning, and 0.92 for fibrosis. Quantitative findings were subsequently confirmed in the validation cohort	Single-centre biopsy cohort; sampling variability and transportability across staining and scanner platforms remain concerns
You et al[18]	Retrospective cohort of 152 patients; 17 routine clinical and laboratory variables; comparison of ten machine-learning models using five-fold cross-validation	MRI-PDFF-defined steatosis as the quantitative reference standard	Logistic regression achieved the highest AUC of 0.873. Random forest achieved the highest accuracy of 0.816 and F1 score of 0.807	Small, single-centre Taiwanese cohort; internal validation only; prospective external validation and assessment of clinical decision thresholds are required

AUC: Area under the curve; EHR: Electronic health record; ICC: Intraclass correlation coefficient; ICD-9: International Classification of Diseases, Ninth Revision; LASSO: Least absolute shrinkage and selection operator; MASLD: Metabolic dysfunction-associated steatotic liver disease; MRI-PDFF: Magnetic resonance imaging proton-density fat fraction; NAFLD: Non-alcoholic fatty liver disease; NASH: Non-alcoholic steatohepatitis; NLP: Natural language processing; PPV: Positive predictive value.

Table 2 Artificial intelligence-assisted prognostication and fibrosis-risk assessment in Metabolic dysfunction-associated steatotic liver disease

Ref.	Population/data	Model/output	Main quantitative finding	Focused limitation
Van Vleck et al[14]	EHR data from 38575 participants; 2281 patients were identified with historically defined NAFLD, including 619 documented only in radiology reports	NLP combining structured and unstructured EHR data to identify disease and examine progression to NASH or cirrhosis	NLP achieved a sensitivity of 0.93 and an F2 score of 0.92. Among the 619 radiology-only cases, 170 patients (28.4%) were later considered likely to have developed NASH or cirrhosis after a median of 1057.3 days	Retrospective and documentation-dependent; historical terminology; progression was inferred from EHR documentation rather than prospectively adjudicated outcomes
Sherman et al[17]	More than 1200 pathology reports used for manual validation; prognostic cohort of 3134 patients with biopsy-confirmed MASLD followed for 20604 person-years	NLP extraction of steatosis, lobular inflammation, ballooning, and fibrosis stage from free-text pathology reports	Positive and negative predictive values ranged from 93.5% to 100%. Compared with simple steatosis, adjusted hazard ratios for cirrhosis were 1.04 for MASH/F0, 1.19 for MASH/F1, 1.89 for MASH/F2, and 4.21 for MASH/F3	Retrospective study within one multihospital health system; transportability across institutions and pathology-reporting practices requires evaluation
Njei et al[37]	Development cohort of 1487 NHANES 2017-2020 participants; external biopsy cohort of 337 patients; mortality cohort of 4276 NHANES III participants	Explainable XGBoost FibroX model for advanced fibrosis and long-term mortality prediction, interpreted using SHAP	FibroX achieved an AUROC of 0.97 vs 0.62 for FIB-4 in development and 0.84 vs 0.82 in the biopsy cohort. The mortality C-statistic was 0.80; the adjusted hazard ratio for cardiovascular mortality was 1.22	The development reference standard combined non-invasive assessment with transient elastography rather than biopsy; prospective clinical-impact and decision-threshold evaluation remain necessary
Njei et al[38]	Systematic review of 26 studies published during 2014-2025 across Europe, the United States, Asia, and South America	AI-based prediction of MASH and fibrosis stages ≥ F2-F4 using clinical, laboratory, imaging, and elastography variables	Validation/test AUROCs were approximately 0.76-0.95 for MASH, with external-validation performance as low as 0.67, and approximately 0.72-0.94 for fibrosis ≥ F2-F4	Considerable heterogeneity in populations, predictors, outcomes, and study designs; data diversity and model interpretability remain important barriers to clinical translation

AI: Artificial intelligence; AUROC: Area under the receiver operating characteristic curve; EHR: Electronic health record; FIB-4: Fibrosis-4 Index; MASH: Metabolic dysfunction-associated steatohepatitis; MASLD: Metabolic dysfunction-associated steatotic liver disease; NAFLD: Non-alcoholic fatty liver disease; NASH: Non-alcoholic steatohepatitis; NHANES: National Health and Nutrition Examination Survey; NLP: Natural language processing; SHAP: SHapley Additive exPlanations; XGBoost: Extreme gradient boosting.

Table 3 Evidence relevant to artificial intelligence-supported therapy and therapeutic development in metabolic dysfunction-associated steatotic liver disease

Ref.	Evidence base	Potential therapeutic contribution	Evidence maturity	Focused limitation
Naoumov et al[53]	Exploratory post hoc analysis of the FLIGHT-FXR trial involving 99 patients with F2-F3 NASH and 198 paired biopsies obtained before and after 48 weeks of placebo or tropifexor treatment	AI-assisted SHG/TPEF digital pathology quantified fibrosis, steatosis, ballooning, collagen architecture, and treatment-related histological changes. It identified reductions in overall and perisinusoidal fibrosis that were not detected by conventional ordinal scoring	Clinically proximate but investigational therapeutic-endpoint evidence	Post hoc analysis of a relatively small trial; evaluated treatment-response measurement rather than AI-guided treatment selection or improved patient outcomes
Harrison et al[42]	Phase 3 randomized trial involving 966 patients with biopsy-confirmed non-cirrhotic NASH and fibrosis who received resmetirom 80 mg, resmetirom 100 mg, or placebo	Demonstrates the expanding therapeutic context in which future decision-support systems may need to consider fibrosis stage, metabolic risk, treatment eligibility, safety, and monitoring requirements	Direct therapeutic evidence, but not an AI study	Does not validate an AI model or demonstrate that AI improves treatment selection, monitoring, or clinical outcomes
Muralidharan et al[30]	Serum lipidomic study of 151 patients with biopsy-paired MASLD data and independent validation in 105 patients; 481 serum lipids were evaluated	Identified lipidomic signatures that may inform future risk stratification, trial enrichment, or biomarker-based therapeutic research	Investigational biomarker evidence; not an AI treatment study	The study assessed disease classification rather than treatment response. The at-risk MASH model achieved an AUROC of 0.912 in development but declined to 0.76 during independent validation
Guo et al[49]	Review of inflammation-mediated crosstalk between diabetes and liver disease, including discussion of AI and systems biology	Provides mechanistic context for using computational approaches to characterize patient-specific inflammatory profiles and identify potential precision-therapy targets	Conceptual and hypothesis-generating evidence	No AI model was developed or validated, and no prospective AI-guided MASLD treatment comparison was conducted
Xia et al[55]	Preclinical drug-discovery study combining pose-filter ensemble learning with structure-based virtual screening of FXR agonists	Supported screening and lead identification, resulting in the discovery of XJ034, which showed FXR activity and metabolic effects in vitro and in high-fat-diet-induced obese mice	Preclinical proof-of-concept for AI-assisted drug discovery	No human MASLD evaluation, clinical treatment comparison, or patient-outcome evidence; the study supports lead discovery rather than established therapy selection
De Chiara et al[56]	Narrative review of the potential integration of AI with liver organ-on-chip platforms for historically defined NAFLD	Describes how AI could analyze complex organ-on-chip data to support target identification, drug screening, and preclinical efficacy or toxicity assessment	Conceptual and preclinical	The publication is a review and does not report direct experimental validation of an AI-integrated organ-on-chip platform in MASLD

AI: Artificial intelligence; AUROC: Area under the receiver operating characteristic curve; MASLD: Metabolic dysfunction-associated steatotic liver disease; MASH: Metabolic dysfunction-associated steatohepatitis; NAFLD: Non-alcoholic fatty liver disease; NASH: Non-alcoholic steatohepatitis; SHG/TPEF: Second-harmonic generation/two-photon excitation fluorescence.