DOI: 10.1161/circ.148.suppl_1.19210 ISSN: 0009-7322

Abstract 19210: Interpretation of Deep Learning Model That Identifies Regional Myocardial Dysfunction From the Surface Ecg

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

Introduction: Cardiac wall motion abnormalities (WMA) are key prognostic indicators for major cardiovascular events and mortality, yet they are poorly detected by current ECG screening tools such as Q waves. Deep learning (DL) models have improved analysis of complex data and novel tools may provide interpretation of these models.

Hypothesis: ECG features beyond the Q wave may be identified by analyzing an ECG-trained DL model to improve WMA detection over standard ECG measurements.

Methods: We developed a DL neural network using 36,964 unique patients (62.9±16.9 years, 47% female, 58% white) labeled by natural language processing of echocardiography reports within 60 days. To understand how the CNN model identifies WMAs from the ECG input, we implemented Shapley Additive exPlanations (SHAP) and DeepExplainer on R-wave aligned beats over 300ms surrounding the QRS complexes to identify hot-spots across the ECG. We then retrained classifiers on 60mg windows of the ECG to identify which timepoints produced the greatest performance (ablation method).

Results: Probing the CNN revealed that ECG regions throughout the QRS and T-waves identified WMA. Analysis output shows that such ECG regions arose throughout the QRS and T-waves, and not just at early regions corresponding to Q waves (Panel A) . An aggregate view of the ECG reports summarizes the most discriminatory ECG regions for all patients (Panel B) . Stepwise analysis of limited windows within aligned ECG beats also confirmed this analysis (Panel C) . We found that the highest performance window was 60-120ms (AUC 0.814), containing most of the QRS, with the lowest performance in the window of 0-60 ms (AUC 0.699) and 0.20-0.80ms (AUC 0.707).

Conclusions: The DL model show promise in accurately detecting WMAs and provides key insights into critical ECG regions for WMA detection beyond the Q wave. Further analysis of WMA detection with these expanded regions is needed.

More from our Archive