Comprehensive and low-cost machine learning models for predicting MASLD prevalence: cross-sectional evidence from the...

Metabolic dysfunction-associated steatotic liver disease (MASLD) represents a growing global health burden, yet early detection remains difficult, especially in low-resource settings. This study aimed to develop an interpretable machine learning model using both conventional and non-invasive, low-cost predictors to assess the probability of MASLD in the general population. Data from 3120 adults were split into training (n = 2,184) and testing (n = 936) sets. Six ML algorithms, including Random Forest, Gradient Boosting Machine, kernel-based Support Vector Machine, k-Nearest Neighbors, Neural Network, and Recursive Partitioning, were evaluated using standard performance metrics. Model interpretability was examined through SHapley Additive exPlanations. Analyses were also stratified by sex to explore heterogeneity in predictor influence. In the whole-variable model, Random Forest and Gradient Boosting Machine achieved the strongest performance (AUC 0.858 and 0.855, respectively). SHAP consistently identified composite hepatic indices, particularly the Fatty Liver Index and Hepatic Steatosis Index, along with waist circumference and adiposity-related measures, as leading contributors to MASLD probability. Sex-specific patterns were evident, with metabolic traits such as fasting glucose and diastolic blood pressure contributing more strongly in women, while hepatic markers and lipid-related variables were more influential in men. In a reduced model restricted to non-invasive, low-cost predictors, waist circumference remained the dominant determinant across all subgroups, followed by diastolic blood pressure, age, and a healthy lifestyle score. These findings highlight the association of visceral adiposity and metabolic dysfunction with MASLD and demonstrate that models based on widely accessible indicators can distinguish individuals with and without MASLD within this cohort, suggesting the potential feasibility of scalable screening approaches in resource-limited settings.