The added value of deep learning to submaximal exercise electrocardiogram for the 10-year prediction of major adverse cardiovascular and cerebrovascular events

Abstract Background The added value of exercise ECG data to traditional risk factors in predicting cardiovascular events remains unclear. Deep learning of ECG signals has demonstrated state-of-the-art performance in detecting subtle abnormalities indicative of cardiovascular disease. We investigated whether deep learning analysis of exercise ECGs could improve the prediction of major cardiovascular and cerebrovascular events (MACCE) with respect to established risk models in a large population-based cohort. Purpose To assess the added value of submaximal exercise ECG measurements to established cardiovascular risk prediction models. Methods We obtained ECG recordings from 41,076 UK Biobank participants without cardiovascular disease (CVD) who underwent a submaximal exercise ECG test. We obtained ECG recordings from 41,076 UK Biobank participants without cardiovascular disease (CVD) who underwent a submaximal exercise ECG test. We created two deep neural network ECG risk scores: one derived from conventional ECG parameters, measured at rest, peak exercise and late-stage recovery, and another based on the complete ECG (Q-ECG). We assessed the added value of conventional ECG parameters and Q-ECG risk scores to the UK's current prediction algorithm (QRISK3). We estimate the association between ECG parameters and MACCE, using Cox Proportional hazard models, following adjustment for traditional risk factors. All models were internally validated via 5-fold cross-validation and 1000 bootstrap iterations. Predictive performance was evaluated using Harrel's C-index, Net Reclassification Index (NRI) and net benefit. Findings: Incident MACCE was reported in 4,082 (9.9%) individuals in the study population and 3,463(9.7%) individuals with valid ECG parameters over a median follow-up period of 12.5 years. We found combined conventional ECG and Q-ECG scores were independently associated with MACCE, following adjustment for multiple testing: adjusted hazard ratio [HZ] = 1.76 (95% CI:1.63-1.91); HZ = 1.14 (95% CI:1.10-1.18), respectively, per standard deviation increase. Both conventional ECG parameters and Q-ECG were predictive of MACCE, independent of clinical risk factors, C-index = 0.64 (95%CI 0.63-0.65); net benefit = 0.09(95% CI 0.07 - 0.11) and C-index = 0.56 (95% CI 0.55-0.57); net benefit = 0.07(95% CI 0.05 - 0.09). ECG measurement's modestly improved model discrimination over the bassline QRISK3 risk score when combined with QRISK3 risk factors for conventional markers and Q-ECG score, respectively; ΔC-index 0.03 (95% CI: 0.02 – 0.04) and ΔC-index 0.03 (95% CI: 0.02 – 0.03); However, we observed no significant improvements in classification at the current recommended threshold of 10%. Conclusion In individuals without a history of prior cardiovascular disease, ECG measures independently predict the risk of MACCE. When combined with QRISK3, neural-network-derived ECG risk scores marginally improve cardiovascular risk prediction over QRISK3.