Abstract 9374: Machine Learning on Automated Measurement of 12 Leads Electrocardiography Can Predict Appropriate Shock in Patients With Implantable Cardioverter Defibrillator

Introduction: Prediction for appropriate shock in patients with implantable cardioverter defibrillator (ICD) is still challenging. We tried to predict the shock by machine learning (ML) model on automated microvolt-level measurement of 12-leads electrocardiography (ECG). Hypothesis: ML on 12-leads ECG can predict appropriate ICD shock. Methods: Consecutive 177 patients (61.5±14.4years, 141males, organic heart disease: 121cases) with ICD were enrolled. ECG was measured by ECAPs12c system (Nihon-Koden) at ICD implantation. Shock was defined as appropriate ICD shock and anti-tachycardia overdrive pacing. Statistic significant predictors were extracted by univariate Cox regression analysis. Because of many correlation/confounding/multicollinearity among the predictors, multivariate Cox was not performed, and machine learning (ML) predictive model was utilized to compare the importance of the predictors. Results: Fifty-six patients were treated by appropriate shock during the observation period (median during implantation to shock: 1.85years). Thirteen significant predictors were extracted, and T-axis showed the smallest P value of univariate Cox (P=0.0007). Random Forest Classifier model demonstrated high accuracy (0.740) and T-axis showed the most important role to build the model. Receiver operating characteristics (ROC) curve analysis indicated the cut-off value as 105 degree, and Kapan-Meier curve analysis demonstrated T-axis ≥105 degree group showed worse prognosis than T-axis >105 degree group. Conclusions: ML of microvolt-level measurement of 12-leads ECG potentially had high predictive value for appropriate shock of ICD, and T-axis played an essential role for the prediction.