Abstract 9562: Prediction of Post-TAVR Pacemaker Implantation Using Machine Learning on Pre-TAVR ECG

Abstract

Introduction: Transcatheter Aortic Valve Replacement (TAVR) is increasingly performed with a decreasing overall complication rate. The risk of complete heart block requiring pacemaker implant (PPI) remains an important complication. Simple and accurate pre-TAVR risk stratification is essential to identify the patients who are at high risk for PPI. Unfortunately, an accurate risk prediction tool is not established yet. Hypothesis: We hypothesized that a pre-existing “electrical fingerprint” on electrocardiography (ECG) is associated with the need for post-TAVR PPI and that using machine learning (ML) approaches on pre-TAVR ECG can predict the need for post-TAVR PPI. Methods: This is a single-center retrospective study of 1032 patients who underwent TAVR from March 2011 to November 2019 at University Hospitals Cleveland Medical Center. Patients with preexisting cardiac implantable electronic devices were excluded. Pre-TAVR ECGs were obtained within 90 days prior to TAVR and a total of 637 features were extracted using ECG MUSE system (GE Healthcare; Milwaukee, WI). We applied hierarchical clustering on 637 ECG features and explored the incidence of post-TAVR PPI in each cluster. Using supervised machine learning, we trained and tested 25 ML algorithms using 445 ECG features (66:34 train:test split). We explored the accuracy of the best 5 ML algorithms on clinical data alone, 26 pre-TAVR clinical variables+ECG, and ECG alone to predict PPI after TAVR. Analysis was done using R Studio (Vienna, Austria) and Waikato Environment for Knowledge Analysis (WEKA, NZ). Results: Hierarchical clustering identified two distinct clusters (cluster 1 has 69% of the patient and cluster has 31% of the patient). The incidence of PPI was 13% in cluster 1 and 33% in cluster 2 (P<0.0001). The best performing supervised ML had an accuracy of 83% and AUC-ROC of 0.76 with no significant improvement with including clinical variables. Conclusions: A pre-existing “electrical fingerprint” accounts for significant variation in post-TAVR PPM implant. Unsupervised and supervised ML approaches can be helpful in predicting PPI without the need for inclusion of clinical data. ML-based models can be broadly applied to automatically-extracted ECG features with minimal data preparation