Invasive and non-invasive variables prediction models for cardiovascular disease-specific mortality between machine learning vs. traditional statistics

Abstract

This study examined the predictive performance of cardiovascular disease (CVD)-specific mortality using traditional statistical and machine learning models with non-invasive indicators, and assessed whether adding blood lipid profiles improves prediction. Data were from 1,749,444 Korean adults (44.7% female) from the Korea Medical Institute. Non-invasive predictors included sex, age, waist-to-height ratio, diabetes, hypertension, and physical activity; invasive variables included triglycerides, fasting glucose, and cholesterol. CVD-specific mortality was tracked over a 10-year follow-up. We applied Cox proportional hazards models (with and without elastic net penalty), Random Survival Forest, Gradient Boosting Survival, and Survival Tree models. Predictive performance was compared using area under the curve (AUC), c-index, and Brier score. All models using only non-invasive predictors achieved AUCs > 0.800 and were not inferior to models including blood profiles. Machine learning models showed slightly higher predictive performance over time than traditional models, but differences were not substantial. Both approaches appear valid for predicting CVD-specific mortality using non-invasive data. Machine learning models may offer marginally improved prediction, and the addition of invasive variables may not substantially enhance model performance.