DOI: 10.1161/circ.148.suppl_1.17420 ISSN: 0009-7322

Abstract 17420: Enhanced Identification of Cardiac Wall Motion Abnormalities: An Externally Validated Deep Neural Network Approach Outperforms Expert and Quantitative Analysis of Electrocardiograms

Albert J Rogers, neal bhatia, James Tooley, Vyom Thakkar, JUSTIN XU, Rayan Ansari, Jessica Torres, Jagteshwar Tung, Mahmood Alhusseini, Paul L Clopton, Reza Sameni, Gari Clifford, Euan A Ashley, Marco V Perez, Matei Zaharia, Sanjiv M Narayan
  • Physiology (medical)
  • Cardiology and Cardiovascular Medicine

Background: Cardiac wall motion abnormalities (WMA) independently predict mortality and other adverse events beyond ejection fraction. Clinical screening relies on detection of ECG Q waves yet has poor predictive accuracy.

Hypothesis: We hypothesized that ECG features beyond the Q wave could be identified by deep learning to improve WMA detection over standard ECG measurements.

Methods: We collected ECGs and echocardiogram pairs (Panel A) in 35,210 unique patients who underwent both studies within 60 days at Stanford University. WMAs were determined by natural language processing (NLP) of clinical reports (Panel B). Five second ECG signals were extracted digitally and used to develop and test a DNN. For development, 80% of the pairs were used and 20% (N = 7042) were reserved for internal validation. We then tested the model performance on 2,338 echo/ECG pairs from Emory University. For comparison, qualitative physician statements from the ECG interpretation suggesting myocardial injury, ischemia, or infarct were labeled using NLP, and a logistic regression model was developed to include quantitative Q-wave measurements.

Results: In the internal validation cohort (62.9±16.9 years, 47% female, 58% white), DNN of the ECG detected echocardiographic WMA with a c-statistic of 0.781 (CI: 0.764-0.798) which was superior to quantitative ECG Q wave measurement (0.632, CI: 0.605-0.661) or qualitative interpretation (0.595) (Panel C). Performance was not improved by incorporating demographic data. In the external cohort (61.9±16.0 years, 48% female, 48% white), the DNN AUC was 0.723 (CI:0.689-0.758).

Conclusion: Deep learning improves detection of cardiac WMAs from the electrocardiogram in an externally validated model compared with standard ECG analysis. Application of this model may provide improved screening for regional myocardial injury. Further work will determine which features beyond the Q wave provide improved performance.

More from our Archive