AI-Based Models for Risk Prediction in MASLD: A Systematic Review

Metabolic dysfunction-associated steatotic liver disease (MASLD) is a rapidly growing global health concern. This systematic review evaluates the efficacy of AI-based models in predicting risk and stratifying patients with MASLD, with a focus on identifying individuals at risk for clinically significant disease, such as ≥ F2 fibrosis, ≥ F3 advanced fibrosis, or MASH, addressing a critical gap in the literature to enhance clinical management and patient outcomes.

Following PRISMA guidelines, we conducted a comprehensive literature search across PubMed, MEDLINE, Web of Science, EMBASE, and CINAHL on November 19, 2024 and June 9, 2025. Predefined keywords were used to identify relevant studies focused on AI-based risk prediction for MASLD. Inclusion and exclusion criteria were established, and data were extracted using a structured form, emphasizing AI model efficacy and quality assessment via the QUADAS-2 tool. This study is registered in the PROSPERO database (CRD42024606950) and conducted without external funding.

We included 26 studies (2014-2025) from Europe, the U.S.A., Asia, and South America spanning retrospective cohorts and cross-sectional designs. Risk of bias (QUADAS-2) was predominantly low overall: patient selection and index test were low risk in all studies; flow/timing was uniformly low risk; reference standard was low risk in most studies with a minority rated unclear. Performance by endpoint (validation/test cohorts): AI models for MASH reported AUROCs of ~ 0.76-0.95 (with external validation AUROCs as low as 0.67), while models for fibrosis ≥ F2-F4 reported AUROCs of ~ 0.72-0.94. Models leveraged machine learning and deep learning approaches; key predictors frequently included age, BMI, liver enzymes (ALT/AST ± GGT), platelets, and other routine clinical features, with multimodal (clinical ± elastography/imaging) inputs often yielding the strongest discrimination within study contexts.

This review highlights the promising role of AI-based models in MASLD risk prediction, showcasing their strong predictive capabilities. Addressing limitations in data diversity and model interpretability will be crucial for integrating AI into clinical practice.