Leveraging machine learning approaches to identify behavioural and lifestyle risk factors and predict mortality among people with cardiovascular disease

Background People living with cardiovascular disease (CVD) remain at high risk of mortality. Machine Learning (ML) offers opportunities to improve risk prediction and identify under-recognised predictors that may be overlooked by traditional statistical models. Purpose To assess the performance of ML models in predicting mortality risk in people with documented CVD and identify associated risk factors. Methods Baseline data from the Sax Institute’s 45 and Up Study (2005-2008) were linked with New South Wales hospitalisation and mortality records through 30 September 2018. The performance of seven ML survival and classification models (Fast Survival SVM, XGBoost, Gradient Boosting, Gaussian naïve Bayes, multilayer perceptron, random forest, and LightGBM) in CVD mortality risk was compared with three traditional approaches (Cox proportional hazards, accelerated failure time models and logistic regression). Models were trained using an 80/20 train-test split. Discrimination was assessed by C-index and Area Under the Receiver Operating Characteristic Curve (AUROC). SHapley Additive exPlanations (SHAP) were used to interpret the best-performing model overall and by sex, compared with Cox hazard ratio-based rankings. External validation used the University of California Irvine Heart Failure Clinical Records dataset. Results Among 18,603 participants diagnosed with CVD at baseline, 7,308 deaths occurred during a 10-year follow-up. Across 101 predictors, classification models showed higher discrimination for mortality, while survival models provided time-to-event risk estimation. LightGBM achieved the highest discrimination overall and in sex-stratified analyses (AUROC≈0.85). SHAP analyses identified age and diagnosed diabetes as dominant risk factors, while self-reported health status and quality of life were the strongest protective predictors. Limitations in physical functioning and unhealthy lifestyle were consistently associated with higher mortality risk. Sex-stratified analyses revealed distinct risk profiles: preventive and health-seeking behaviours (lifetime mammography screening, aspirin use) were prominent protective factors in females, whereas mortality risk for males was more strongly driven by behavioural and lifestyle factors, including sedentary time, sleep duration, alcohol intake, fruit juice intake, and severe mobility limitations. Compared with traditional statistical models, LightGBM identified a broader range of prognostic factors, including physical functioning (bathing, walking, bending, climbing), dietary, preventive screening behaviours (sugar, processed meat intake) and hearing loss, which were weakly represented or overlooked. Conclusion LightGBM demonstrated the highest discrimination for mortality over classical Cox-based modelling, highlighting under-recognised multidomain risk and protective factors that may inform clinical practice and targeted risk stratification and prevention strategies.