Artificial intelligence (AI) enhanced ECG predicts obstructive cardiovascular disease and death using secondary data; a benchmarked comparison of different methods

Abstract

Abstract Introduction Assessment of pretest-probability of chronic coronary artery disease (CAD) is based on the modified Diamond and Forrester criteria, including age, gender, and symptoms. However these criteria suffer from low accuracy, requiring further diagnostics prior to coronary angiography (CA), e.g. via coronary computed tomography angiography. Artificial intelligence (AI)-enhanced evaluations of broadly available resting ECGs have shown excellent results in the detection of multiple cardiovascular diseases. In this analysis, we aimed to evaluate and compare the ability of neural networks, random forest and support vector machines as different AI-based approaches regarding obstructive CAD and long-term mortality using structured ECG data. Methods The cohort of the ECAD registry includes patients undergoing CA at our heart and vascular center between 2004 and 2019. Patients with a digitally available resting ECG within 90 days prior to CA, containing structured data on 648 characteristics for each ECG, were included. The overall cohort was divided in a learning (60%), validation (20%) and a test cohort (20%). Obstructive CAD was defined by percutaneous intervention as by discretion of the interventional cardiologist. The incidence of death due to any cause was evaluated during follow-up. Logistic regression (LR) and linear discriminate analysis (LDA) were used as benchmarks. Calculations were run 100 times and compared by their mean area under the receiver operating curve (AUC) in the validation cohort. An initial neural net (NN1) was improved by sophisticated hyperparameter tuning, dropout layers and weight regularization (NN2). Feature reduction lead to NN3. The initial random forests (RF1) was adapted for the risk factors (RF2). Results Data from 7076 CAs were included. An obstructive CAD requiring revascularization was present in 2075 cases (29.3 %). Neural networks with feature selection based on ECG characteristics outperformed traditional risk factors and provided highest area under the ROC-curve (table 1). During a median follow-up of 2.4 years (Q1: 0.8; Q3: 6.3), 1137 patients (16.1 %) died. For prediction of all-cause mortality, neural networks based on structured ECG data provided better prediction as compared to risk factors. All NNTs and the RF models exceeded the LR and LDA as benchmark, while SVM did not hold up to conventional multivariate analyses (table 1). Conclusion Based on structured data from resting ECGs, NNTs and RF analyses improve the prediction of obstructive CAD and long-term survival in patients undergoing CA.