Predicting 3-year mortality after myocardial infarction in a structured care program: A comparative analysis of Cox regression and machine learning models

Le résumé fourni par la source

BACKGROUND: Accurate risk stratification after myocardial infarction (MI) remains essential for optimizing long-term management. The advantage of machine learning (ML) over conventional statistical models in real-world clinical settings remains uncertain. AIMS: Our aim was to compare the performance of Cox proportional hazards regression and ML models in predicting 3-year all-cause mortality after MI. METHODS: This retrospective cohort study included 965 patients enrolled in a structured post-MI care program between 2018 and 2022, all with complete 3-year follow-up. A development cohort (n = 740) and a temporal validation cohort (n = 225) were defined. A Cox model and three ML models (Random Forest, XGBoost, and Random Survival Forest) were developed using baseline clinical variables. Performance was assessed at a fixed 3-year horizon using area under the receiver operating characteristic curve, calibration, and the Brier score. RESULTS: During the 3-year follow-up, 126 patients (13%) died. In temporal validation, discrimination was similar across models, with area under the receiver operating characteristic curves of 0.742 (Cox), 0.738 (Random Survival Forest), 0.727 (Random Forest), and 0.712 (XGBoost) (P = 0.59). Random Survival Forest achieved the lowest Brier score (0.105). In multivariable Cox analysis, age, comorbidity index, log-transformed creatinine, and log-transformed C-reactive protein were independently associated with mortality. CONCLUSIONS: In a moderate-sized real-world cohort with limited predictors, ML models did not outperform Cox regression in predicting 3-year mortality after MI. Age, comorbidity burden, renal dysfunction, and inflammatory status emerged as the most important predictors of mortality. These findings require external validation in independent cohorts before clinical implementation.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.