Abstract 4141993: Artificial Intelligence ECG Analysis More Accurately Identifies Epicardial versus Endocardial Ventricular Tachycardia and Pacing Compared with Visual Criteria

Background: Accurate prediction of whether a ventricular tachycardia (VT) source or critical isthmus is epicardial may facilitate efficient and effective catheter ablation. Prior visual criteria have been described but performance may vary depending upon region. Hypothesis: We hypothesized that an artificial intelligence (AI) algorithm would provide more accurate identification of epicardial versus endocardial sources distributed throughout the left ventricle (LV) compared with visual analysis. Aims: We endeavored to assess the performance of a novel AI algorithm in separating epicardial and endocardial VT and pacing sources compared with 7 previously published ECG criteria. Methods: Patients with LV VT or pacing were enrolled. Episodes were grouped as either LV-endocardial or non-LV-endocardial (i.e., epicardial) according to site of successful ablation or location of the pacing electrode. 12-lead ECG data was analyzed using the AI algorithm to predict source location and accuracy was computed. For comparison, accuracy of 7 previously described visual ECG criteria was assessed. Results: A total of 136 arrhythmia or pacing episodes from 72 patients (mean age 66±14 y, 32% female, EF 39±16%) were analyzed, evenly distributed between epicardial (n=68, 50%) and endocardial (n=68) sources. Episodes were localized to all 17 AHA LV segments. Algorithm accuracy was 91.1% (124 of 136 episodes), exceeded the best performing visual ECG criteria (shortest precordial rS >121 msec; 62.5%; 85 of 136 episodes; P<0.001). Algorithm sensitivity was 0.971 [95% CI: 0.898 - 0.996] and specificity was 0.853 [95% CI: 0.746 - 0.927]. Figure A shows an example ECG from a 73-year-old patient undergoing epicardial pacing at the inferior left ventricular apex. Figures B, C, and D show orthogonal views of the local normal vector (blue) and the electrical activation vector (red), consistent with an epicardial origin. Conclusions: The AI algorithm provided improved accuracy versus visual ECG criteria. Future studies are required to assess whether such information improves procedural efficiency.