Abstract 16866: Artificial Neural Networks as Prediction Tools: Predicting Permanent Pacemaker Implantation (PPMI) in Patients Undergoing Transcatheter Aortic Valve Replacement (TAVR) Utilizing a Shallow Feed-Forward Neural Network

Abstract

Introduction: Despite increasing TAVR volumes after a series of recent favorable clinical trials, adverse outcomes remain frequent, including new or worsened conduction disease requiring PPMI, life-threatening bleeding, paravalvular leak, and stroke. PPMI carries a reported incidence ranging from 8.8-14.6% and is of particular concern given the increased risk of mortality and rehospitalization. New techniques in signal processing may inform novel statistical approaches to better predict PPMI from a set of clinical variables. Artificial neural networks (ANN) comprise a family of algorithms that utilize non-linear activation functions to enable improved prediction of PPMI in TAVR patients. Objective: To examine the predictive utility of a feed-forward neural network in classifying PPM implantation in patients undergoing TAVR. Methods: Pre and post-operative data from a single institution were collected for all patients undergoing TAVR without prior pacemaker implantation from January 2016 to December 2019. Data was imported into Matlab, partitioned into training, validation, and test sets, and processed in a two-layer feed-forward neural network with sigmoid hidden and softmax output neurons. Performance data included confusion matrices and receiver operating characteristic (ROC) curves. Results: The total sample size for the cohort was 513 patients with a PPMI incidence of 8.6%. The training set contained 40 variables and 359 patients, while the validation and test sets contained 77 patients each. The final optimized model showed cross-entropy of 0.25 with 6 iterations and an area under ROC curve of 0.73. Overall model accuracy was 92.7% in the validation set and 88.3% in the test set. Conclusions: In summary, we show that feed-forward neural networks can be useful in processing multiple interdependent variables to aid clinical prediction. Our network demonstrated modest discriminatory ability in predicting the need for PPMI after TAVR.