Po-04-174 an electrographic deep learning model for long-term mortality of the patients surviving after acute myocardial infarction

The diagnosis of acute myocardial infarction (AMI) initially depends on a patient’s history and the 12-lead ECG. Significant changes of the ST segment on ECG are one of the diagnostic markers of AMI and are associated with higher short-term or long-term mortality of patients. However, scoring systems or predictors for long-term risk stratification are still lacking for patients with AMI. Recently deep learning (DL) model built based on 12-lead ECG showed its great potential for early diagnosis of heart failure and atrial fibrillation. We aimed to develop a prognosis model for the long-term mortality of AMI patients based on 12-lead ECG using DL. Twelve-lead ECGs of patients during AMI were retrieved retrospectively between 2011 to 2017, and a convolutional neural network (CNN) using 12-lead ECG and patient’s age has been built to predict all-cause death within one year. The patients’ ECG recordings were split into training, validation, and testing set at a ratio of 5:2:3 and the patients were not overlapped among different datasets. Survival analysis was conducted using the Kaplan-Meier method with available follow-up data in a testing set stratified by the built model. Furthermore, Cox proportional hazard model was performed to get crude hazard ratio (HR) and covariate-adjusted HR for TIMI and GRACE score. 16827 ECGs from 2238 patients (77% male, age 64.8±13.8 years) with AMI were included. The AUC of the DL model is 0.80, with a sensitivity of 82%, and specificity of 66% using an optimal threshold on the testing set. In addition, we also found that the patients labeled as dead by DL model have a higher probability of death over a 6-year follow-up with the HR of 6.00 (95% confidence interval (CI) 3.92-9.20) while the HR of TIMI score is 1.38 (95% CI 0.97-1.97) and HR of GRACE score is 2.72 (95% CI 1.92-3.86). By applying multivariate adjustment, the DL model can be an independent predictor with the HR of 5.68 (95% CI 3.69-8.74) for the sex-adjusted model, 6.11 (95% CI 3.95-9.45) for the TIMI-socre-adjusted model and 5.22 (95% CI 3.38-8.06) for GRACE-score-adjusted model. (Figure 1) Our result showed that the DL model built based on 12-lead ECG can predict long-term mortality after AMI for up to 6 years. The DL model can provide an instant prognostic indicator using the 12-lead ECG complemented with TIMI or GRACE scores.