Interpretable machine learning and incremental prognostic model for all-cause and cardiovascular mortality in obesity with MASLD population: Comorbidity insights from triglyceride-glucose and its obesity-related derivatives.

OBJECTIVE

To investigate comorbid correlation and prognostic value of triglyceride-glucose (TyG) and its obesity-related derivatives for all-cause and cardiovascular mortality in the obesity with metabolic dysfunction-associated fatty liver disease (MASLD), and develop incremental predictive model using machine learning and SHapley Additive exPlanations (SHAP) analysis.

METHODS

4807 adults with obesity and MASLD enrolled from National Health and Nutrition Examination Survey 1999-2018. Kaplan-Meier, Cox regression, and restricted cubic splines (RCS) were used to examine associations. Machine learning and incremental prognostic model were developed to predict mortality, with SHAP interpreting feature importance.

RESULTS

Both TyG and TyG-WHtR were associated with all-cause mortality after full adjustment, whereas TyG-WHtR showed a more robust association with cardiovascular mortality. TyG-WHtR showed the strongest association: each 1-unit increase linked to 31.7% and 46.7% higher risk of all-cause and cardiovascular mortality respectively in the fully adjusted model, with RCS showing an S-shaped relationship. Time-varying hazard ratio patterns suggested that the associations of TyG and TyG-WHtR with mortality were stronger in the early follow-up period and attenuated over time. XGBoost performed best with SHAP analysis identified age, fasting plasma glucose (FPG), male as primary factors. The incorporation of TyG and TyG-WHtR into predictive models modestly improved the prediction performance for mortality outcomes.

CONCLUSIONS

TyG-WHtR showed a stronger time-dependent association with mortality in obesity-MASLD comorbidity, while TyG mainly associated with all-cause mortality. S-shaped curve for TyG-WHtR indicates the critical need for early intervention. Integrating TyG-WHtR into XGBoost prognostic models enhances risk identification, while SHAP values clarify age, FPG, male as key drivers.