Development and validation of an explainable machine learning model using routine laboratory biomarkers for identifying prevalent MASLD: Evidence from two observational studies

Although many predictive models for metabolic dysfunction-associated steatotic liver disease (MASLD) have been developed, their performance remains suboptimal. We aimed to develop an interpretable machine learning (ML)-based plasma biomarker model for identifying prevalent MASLD. Data from the National Health and Nutrition Examination Survey (NHANES 2017–2020) were randomly divided into a training cohort (N = 2760) and an internal cohort (N = 1184). Eleven ML algorithms were employed to construct classification models. Model interpretability was visualized via the SHapley Additive exPlanations (SHAP) method. External validation of these models was further conducted using data from the Korea NHANES (KNHANES) 2019–2021. The association between the selected features and prevalent MASLD was evaluated using restricted cubic spline regression analysis. Feature selection was performed using LASSO regression and the Boruta algorithm. Key predictors included diabetes mellitus (DM), waist circumference (WC), age, hypertension, and atherogenic index of plasma (AIP). All evaluated ML algorithms demonstrated robust predictive capabilities, with areas under the curve (AUC) exceeding 0.70. Among these, the Extra Trees (ET) performed the best, achieving an AUC of 0.879 (95% CI 0.856–0.897) in the internal testing group and maintaining good performance in the external KNHANES cohort with an AUC of 0.822 (95% CI 0.815–0.829). The DeLong test revealed significant differences in AUC between ET and other algorithms. These findings suggest that age, WC, DM, hypertension, and AIP are informative features associated with prevalent MASLD. The ET model showed strong discriminative performance and may serve as a practical tool for MASLD screening.