A TyG-UHR-based machine learning model for screening lean MAFLD: development and external validation.

Area of Science:

• Hepatology
• Machine Learning
• Public Health

Background:

• Lean metabolic dysfunction-associated fatty liver disease (MAFLD) is underdiagnosed in primary care due to low risk perception and imaging limitations.
• Cost-effective screening using routine data is needed to improve early detection and referral for MAFLD.

Purpose of the Study:

• To develop and validate an explainable machine learning model for screening lean MAFLD using routinely collected health checkup data.
• To assess the model's performance, calibration, and clinical utility in both internal and external validation cohorts.

Main Methods:

• Developed an Extreme Gradient Boosting (XGBoost) model using triglyceride-glucose (TyG) and uric acid-to-high-density lipoprotein cholesterol ratio (UHR), plus other clinical predictors.
• Utilized elastic-net regression for predictor selection and rigorous leakage control during model tuning.
• Externally validated the model on a separate cohort, employing post hoc recalibration to improve accuracy.

Main Results:

• The XGBoost model demonstrated high discrimination (AUC 0.995 in test set) and improved probability accuracy after recalibration in external validation (Brier score 0.109).
• Key predictors included TyG, systolic blood pressure, and UHR, with the model showing significant net benefit across various thresholds.
• The model proved interpretable via SHAP values, highlighting TyG as the most influential factor.

Conclusions:

• An explainable machine learning model using TyG and UHR can effectively and affordably screen for lean MAFLD from routine data.
• The model supports primary care triage and can be adapted to external populations via recalibration for easier implementation.
• Paired prediction with local recalibration is recommended to maintain accuracy across diverse healthcare settings.