Abstract 17676: An Electrocardiographic Deep Learning Model is Informative for the Prediction of Atrioventricular Block After Transcatheter Aortic Valve Replacement

Abstract

Introduction: Atrioventricular block requiring a permanent pacemaker (PPM) is an important complication after transcatheter aortic valve repair (TAVR). While several risk factors have been identified, a clinically useful risk prediction algorithm has been challenging to develop. Furthermore, while artificial intelligence methods utilizing structured clinical data have been used, a deep learning approach using raw electrocardiogram (ECG) data is a novel method in this field. Research Question: The objective of our study was to develop an electrocardiographic artificial intelligence (ECG-AI) model to assess the risk of PPM placement after TAVR using only the pre-procedure ECG. Methods: We analyzed data from 651 patients that underwent TAVR at Atrium Health Wake Forest Baptist in the last 10 years. We obtained raw ECG files in XML format from General Electric (Boston, MA, USA). The ECG data were used as inputs within a convolutional neural network (CNN) model. The CNN was an adjusted residual network with a batch size of 64, had 100 epochs, and used Adam optimization with default parameters. The model was trained using a 5-fold cross-validation strategy on 80% of the data and validated on a 20% hold-out set. The CNN was supervised on need for new PPM within 30 days after TAVR. All model evaluations were performed on the hold-out set. Results: Of the 651 patients in our sample, 61 patients required PPM placement within 30 days after TAVR. The average age was 77.2±9 years and there were 260 females (40%). In our sample, 34/268 (12.7%) patients with self-expanding valves and 26/377 (6.9%) patients with balloon-expanding valves received a PPM. Results of our final model on the hold-out set yielded an area under the receiver-operator characteristic curve (AUC) of 0.74 (95% CI 0.56-0.92). The accuracy of our preliminary model was 77% with a sensitivity of 60% and specificity of 79%. Conclusion: Our ECG-AI model achieved a moderate AUC for the prediction of PPM after TAVR. This serves as a proof of concept that pre-TAVR ECGs alone are informative for prediction of PPM. Future work includes model optimization and adding demographic and clinical variables to refine and develop a robust predictive model with continual improvement of the accuracy for prediction of TAVR outcomes.