An artificial intelligence based model for predicting long-term all-cause mortality after acute myocardial infarction(the AIMI model)

Background and aims Prediction of long-term mortality after acute myocardial infarction (AMI) remains challenging. Artificial intelligence (AI) may overcome the limitations of traditional statistical methods1.We aimed to establish an AI-based model for predicting long-term all-cause mortality after acute Myocardial Infarction (the AIMI model). Methods We consecutively enrolled AMI patients undergoing emergent coronary angiography and PCI within 24 hours of symptom onset. Five machine learning (ML) algorithms (RF, SSVM, CoxPH, GBSA and XGB) were employed2. We used recursive feature elimination (RFE) to select predictive variables. Individual predictions were visualized by SHAP plots. AIMI model was compared against GRACE and TIMI scores using time-dependent receiver operating characteristic (ROC) curves, Brier scores, and Kaplan-Meier (K-M) analyses. External validation was also performed at the same way. (Figure 1) Results A total of 4825 patients were used to train and test the AIMI model and 723 patients for external validation. Model incorporated 15 variables achieved robust performance (C-index=0.81). As indicated by AUCs in test set, the model outperformed GRACE and TIMI risk scores across short-, mid- and long-term periods, especially for long-term prediction (1, 3 and 5 years). K-M curves confirmed precise discrimination between low-, median-, and high-risk groups (all p<0.05). External validation confirmed AIMI model's good generalization and robustness (AUCs: 0.88, 0.91, 0.83, 0.75, 0.78 and 0.77 during hospitalization, at 30 days, half-year, 1-year, 2-years and 3-year follow-up; comparisons of K-M curves across three risk groups, all p<0.05). Brier scores showed its good performance in individual level (mean value of internal validation cohort: 0.022; external validation: 0.021). (Figure 2) Conclusions AIMI model surpassed traditional methods in predicting all-cause death, especially for long-term prediction. AI/ML-based model demonstrated potential to enhance risk stratification, guiding clinical decision-making and personalized treatment3.Figure 1Figure 2