关键词: 脂肪肝, 机器学习预测模型, 甘油三酯-葡萄糖-体质指数, 体检人群

Abstract: Objective The aim of this study was to set up and validate a precise yet low-cost early prediction model for fatty liver disease based on routine indicators available in health-checkup centers. Methods A retrospective cohort of 1212 individuals for physical examination was analyzed, and the fatty liver was diagnosed based on ultrasonography. Various indexes were calculated based on clinical materials. Nested cross-validation (10-fold outer loop for validation and 5-fold inner loop for tuning) was combined random-forest and XGBoost with LASSO Logistic regression was conducted for feature selection. Variable importance was interpreted with SHAP values. Internal validation was used 1000-bootstrap optimism-corrected AUC and external validation was employed by a 30 % random split. Results Of the 1212 individuals, fatty liver was found in 542 cases(44.7%);the final model retained four variables, e.g.,triglyceride-glucose-body mass index (TyG-BMI), body-fat percentage, diastolic blood pressure and monocyte to high-density lipoprotein cholesterol (MHR); AUC of nested-cross-validation was 0.874 (95 % CI: 0.855-0.893), AUC of final-model was 0.880 (95 % CI: 0.861-0.898), AUC of optimism-corrected was 0.878 (95 % CI: 0.860-0.897) and AUC of external was 0.866 (95 % CI: 0.830-0.902); calibration was excellent (slope ≈ 1; Hosmer-Lemeshow P=0.433) and robust under 30 % Gaussian noise (AUC=0.878); SHAP analysis identified TyG-BMI as the dominant contributor. Conclusion The four-variable model demonstrates high discrimination, excellent calibration, easy acquisition and strong generalizability, which might offer health-checkup centers a“precise, efficient and low-cost” screening tool for fatty liver disease.

Key words: Fatty liver, Machine learning prediction mode, Triglyceride-glucose-body mass index, Health check-up population