One-year survival prediction models following ST-elevation myocardial infarction: A comparative analysis of the Cox Frailty Model and machine learning

Background: The aim of this study was developing and comparative analyzing prediction models using a Cox proportional hazards model with and without frailty, random survival forests (RSF) and survival support vector regression (SVR). Methods: In this study, 2800 patients with STEMI have been used and two machine learning methods for survival analysis have been applied: RSF and SVR, then the Cox model with and without frailty has been employed. The main outcome was 1-year mortality after STEMI. In this study, 16 variables have missing data. After applying four multiple imputation via chained equations methods, the “Sample” algorithm was selected as the appropriate model with complete data and the modeling process was continued with this data and Hazard Ratio (HR) were calculated. Results: Overall, 1628 (58.1%) patients received primary percutaneous coronary intervention and 737 (26.3%) received thrombolytic therapy. Based on the experimental results, between all the models, the Cox with frailty model performed the best, with the highest overall C-index (0.891) and time-dependent area under the curve (0.9134) and the least Brier score (0.0458). Ever smoking (HR= 1.46), systolic blood pressure (HR= 0.98), left ventricular ejection fraction (HR= 0.96), glomerular filtration rate (HR= 0.96), and reperfusion therapy (No reperfusion HR= 2.71) independently associated with 1-year mortality of STEMI patients. Conclusion: The findings suggest that there are advantages in developing frailty models further than the fundamental Cox proportional hazards regression for estimating the likelihood of survival for STEMI patients to account for the unobserved heterogeneity in grouped observations.