Risk Factors for Coronary Heart Disease and Stroke Among Older Adults With Metabolic Dysfunction-Associated Steatotic...

Older adults with metabolic dysfunction-associated steatotic liver disease have elevated cardiovascular disease (CVD) risk, yet conventional models, such as the Framingham Risk Score, rely on linear assumptions and were developed for younger populations. Nonlinear machine learning approaches may improve risk prediction in this high-risk group.

Using Korean National Health Insurance Service data, this population-based cohort study followed up 67 575 adults, aged ≥65 years, with metabolic dysfunction-associated steatotic liver disease (assessed 2009-2010) for incident CVD, coronary heart disease, and stroke from 2011 to 2019. Cox proportional hazards (CoxPH) models identified time-to-event predictors, whereas an extreme gradient boosting (XGBoost)-based survival model explored nonlinear associations and predicted 9-year CVD risk.

Both CoxPH and XGBoost models outperformed the Framingham Risk Score, with CoxPH showing slightly better discrimination than XGBoost in the validation sets. The 2 models identified broadly similar leading predictors. For coronary heart disease, both identified Charlson comorbidity index, antiplatelet agent use, and serum creatinine among the most important predictors. In the CoxPH model, the Charlson comorbidity index had the highest explained relative risk (0.031), and the XGBoost model also ranked Charlson comorbidity index first by Shapley Additive Explanations value (0.159). For stroke, both models identified age, Charlson comorbidity index, and antiplatelet agent use as the top 3 predictors, with age ranked first (explained relative risk, 0.065; Shapley Additive Explanations, 0.239). Similar findings were observed for total CVD.

In older adults with metabolic dysfunction-associated steatotic liver disease, CoxPH and survival XGBoost showed similar predictive performance and identified key predictors not emphasized in traditional models, like the Framingham Risk Score. These findings underscore limitations of applying younger-population-based tools to older adults with metabolic dysfunction-associated steatotic liver disease and support the development of tailored CVD risk models.