A machine learning-based risk prediction model for major adverse cardiovascular events within one year after discharge in patients with acute coronous syndrome
T Li, X X Wei, M X Zhang, K L Lu, C ZhangAbstract
Background
Risk stratification after ACS is critical for long-term prognosis, but current tools like GRACE rely on traditional variables, excluding novel biomarkers and environmental exposures (e.g., PM2.5, temperature) known to threaten cardiovascular health. Moreover, assessment of composite outcomes beyond all-cause death is clinically needed yet limited.
Purpose
To integrate clinical variables, biomarkers, and environmental exposures (long-term PM2.5, seasonal temperature) using machine learning to develop prediction models for 1-year all-cause death, all-cause death/myocardial infarction (MI), and composite MACEs (cardiovascular death, non-fatal MI, ischemic stroke), and to construct a composite voting decision model for refined risk stratification.
Methods
From the National Chest Pain Center registry, 148,493 ACS patients (866 districts/counties) formed the derivation cohort, with 9,343 patients (407 districts/counties) for external validation. Boruta algorithm selected predictors from demographics, medical history, biomarkers, treatments, and environmental exposures. Six machine learning algorithms developed models for three outcomes. After obtaining optimal single models (all XGBoost), a composite voting decision model integrated their risk levels via majority voting.
Results
One-year incidence rates were 1.3% (all-cause death), 1.8% (all-cause death/MI), and 1.5% (MACE). Boruta identified age, NT-proBNP, and PM2.5 as key predictors across all outcomes. The three single models achieved AUCs of 0.857 (all-cause death), 0.872 (all-cause death/MI), and 0.838 (MACEs) in internal validation, and 0.826, 0.803, and 0.754 in external validation, respectively. In the external validation cohort where GRACE 2.0 scores were available, both the all-cause death model and the all-cause death/myocardial infarction model demonstrated superior predictive performance to the GRACE 2.0 score (AUC: 0.831 vs. 0.823 for all-cause death; 0.818 vs. 0.791 for all-cause death/MI), with better calibration. The construction of composite outcome models (all-cause death/MI and MACEs) extended risk assessment to a broader spectrum of event types, providing more comprehensive prognostic information for clinical practice than the single all-cause death model alone. The composite voting model further optimized stratification: low-risk groups had event rates ≤0.6%, intermediate-risk 1.8%-2.8%, and high-risk groups showed markedly elevated rates (12.4%, 13.2%, and 8.2%). For MACEs, the composite model demonstrated significant net reclassification improvement (NRI=0.089, P=0.008).
Conclusion
This study integrated environmental exposures into machine learning models for ACS prognosis. Environmental factors independently enhanced prediction, composite outcome models provided comprehensive risk assessment, and the composite voting decision model enabled precise risk stratification, offering a robust tool for individualized secondary prevention.Performance and interpretation of the modelsEvent rates in low- and high-risk groups