Deep-Learning based approach to ED chest pain risk stratification by ECG

Background/Introduction Convolutional neural networks (CNNs), have successfully predicted ASCVD, arrhythmias, and structural heart disease but have not been applied in acute ED settings for chest pain risk stratification. A CNN-driven, ECG-derived risk score for cardiovascular (CV) risk could enhance clinical decision-making and improve early intervention strategies. Purpose To develop and validate an ECG-Derived Risk Estimator (ECG-Risk) using a CNN model capable of predicting 30-day CV mortality in patients presenting with chest pain. Methods Using a large dataset of resting 12-lead ECGs collected at an academic institution, we trained and validated CNNs to create the continuous "ECG-Risk" variable, an estimator of adverse CV risk. The primary outcome was 30-day mortality. Mortality was confirmed using the Medicare Social Security Death Index (SSDI). Multivariable logistic regression models were adjusted for age, troponin, race, blood pressure, hyperlipidemia, type 2 diabetes mellitus statis, previous myocardial infarction, and smoking status. Model discrimination was assessed using area under the receiver characteristic curve (AUC-ROC). De Long’s test was used to compare the model AUCs. Results Among 1,532,156 total patients with 12-lead ECGs performed at the institution from 2005 to 2019, 22,590 (1.5%) presented to the emergency department or ambulatory clinic with chest pain. ECG-Risk was not significantly correlated with troponin values (p > .05). There were no 30-day mortality events in those with an ECG-Risk score < -1.25 (Figure 1). In fully adjusted models, ECG-Risk testing was associated with increased rates of 30-day mortality (OR [95%CI]: 3.47 [2.66 – 2.54]). The addition of ECG-Risk to fully adjusted models that included troponin to predict 30-day mortality significantly improved model performance (AUC 0.80 vs 0.75, p<.05) (Figure 2). Conclusions The CNN-based ECG-Risk score demonstrated superior performance in predicting short-term cardiovascular mortality and has the potential to serve as a rule-out exam compared to traditional clinical models including troponin. These findings highlight the potential of deep learning-enhanced ECG interpretation for objective, real-time risk stratification in acute cardiovascular settings.