Background <p>Lean metabolic dysfunction-associated fatty liver disease (MAFLD) is increasingly recognized but often goes unnoticed during health checkups and primary care due to low perceived risk and limitations of imaging. Cost-effective, automated screening methods using routine data could enhance triage and referrals for confirmatory imaging.</p> Methods <p>We used routinely collected data from a hospital health checkup cohort (<i>N</i> = 3346; Huai’an First People’s Hospital, China) to develop a machine learning model for lean MAFLD incorporating prespecified predictors––triglyceride-glucose (TyG) and uric acid–to–high-density lipoprotein cholesterol ratio (UHR). The cohort was randomly split 70:30 into training and test sets. In the training set, additional clinical predictors were selected via elastic-net regression, followed by the tuning of eight algorithms with strict leakage control, utilizing average precision as the primary optimization metric. Extreme gradient boosting (XGBoost) was chosen based on test set discrimination, calibration, and clinical utility. External validation utilized data from the National Health and Nutrition Examination Survey 2017–2020 (<i>N</i> = 2216), split 50:50 into calibration and hold-out validation subsets. Post hoc recalibration was performed on the calibration subset. Performance in the hold-out validation subset was summarized using the receiver operating characteristic curve, calibration plots, decision curve analysis, and Shapley Additive Explanations (SHAP) for model interpretability.</p> Results <p>Restricted cubic spline analysis revealed a nonlinear association for TyG, whereas UHR showed an almost linear association. Elastic-net regularization identified additional clinical predictors, including systolic blood pressure (SBP), diastolic blood pressure (DBP), age, the alanine aminotransferase/aspartate aminotransferase (ALT/AST) ratio, and serum creatinine (Scr). In the test set, XGBoost achieved an area under the curve (AUC) of 0.995, a Brier score of 0.008, a log loss of 0.035, and an expected calibration error of approximately 0.007, offering the highest net benefit across threshold probabilities of 0.05 to 0.60. In the hold-out validation subset, discrimination remained robust both before and after recalibration (AUC 0.826 vs. 0.810), while probability accuracy improved significantly (Brier score decreased from 0.286 to 0.109; log loss from 1.253 to 0.388). Recalibrated probabilities demonstrated acceptable calibration (intercept 0.103; slope 0.684), with consistent positive net benefit observed across thresholds of approximately 0.02 to 0.30. Regarding interpretability, SHAP values ranked TyG (2.538) as the most influential factor, followed by SBP (1.727) and UHR (0.999), with modest but consistent contributions from age (0.810), Scr (0.713), and the ALT/AST ratio (0.657), whereas DBP (0.266) had the smallest effect.</p> Conclusions <p>An explainable machine learning model centered on TyG and UHR can cheaply screen for lean MAFLD using routine health checkup data, support threshold-based triage in primary care, and be adapted to external populations with post hoc recalibration. Adoption should include prediction paired with lightweight local recalibration to maintain accuracy and make implementation easier across different sites.</p>