AI-Based Predictive and Diagnostic Electrocardiography

International Society of Electrocardiology, Tempus-authored
Authors Christopher M. Haggerty, Sushravya Raghunath, Alvaro E. Ulloa Cerna, Linyuan Jing, Jeffrey A. Ruhl, Dustin N. Hartzel, Christopher R. Kelsey, Daniel B. Rocha, Noah Zimmerman, Steven Steinhubl, Thomas B. Morland, Ruijun Chen, John M. Pfeifer, Brandon K. Fornwalt

Background – The combination of artificial intelligence (AI) and large historical electrocardiography (ECG) datasets has produced models able to both detect previously unrecognized clinical disease and predict risk of future events with remarkable performance. This presentation will highlight several recent examples of model development—focusing on prediction of atrial fibrillation (AF) and structural heart disease—as well as demonstrate the potential for such models to improve upon current clinical paradigms.    

Methods – Digital 12-lead ECG data were extracted from the clinical archives of a large, regional US healthcare system (Geisinger) and linked with clinical patient data available in the electronic health records. Prevalent disease at the ECG encounter and incident events following the ECG, such as AF, stroke, or aortic stenosis, were algorithmically defined and validated by physician-led chart review. Deep neural networks (DNNs) were trained—using digital ECG voltage data as inputs—with either a 5-fold cross-validation scheme or time-based censoring scheme. 

Results – As previously reported, the AF DNN model exhibited strong performance in predicting new diagnosis of AF within 1-year of the ECG (cross-validation AUROC=0.85). In a simulated clinical deployment scenario evaluating risk using a single ECG per patient, this model was more efficient in finding AF than contemporary, primarily age-based patient screening criteria, such as those used in recent clinical trials (STROKESTOP, mSToPS, SCREEN-AF). Specifically, the DNN model achieved a comparable, if not lower, number-needed-to-screen for AF (10) and stroke (160) compared to the clinical trial approaches (ranges: 11–18 and 155–317, respectively), but with far superior AF sensitivity (65% vs. 4–7%). 

The composite structural heart disease model also exhibited excellent performance, as previously reported, (cross-validation AUROC=0.91) in detecting significant underlying disease from digital ECG data. In a simulated clinical deployment scenario, for every 100 patients over age 40 without previously known structural disease (including valvular disease, inter-ventricular septal thickening, and left ventricular systolic dysfunction), the model classified 12% as high-risk for underlying disease, 77% of whom were echocardiographically confirmed within the subsequent year. 

Conclusions – Enhancing clinical workflows through integration of robust DNN models may help improve diagnosis and risk-based stratification of previously unrecognized, yet actionable disease. Future work is needed to prospectively validate these findings and optimize clinical integration.