OR03-1
Neural Networks Yield Good Prediction Models for Patient Pacemaker Implantation Post TAVR
Presenter
Chris S Abrahim, MD, Hackensack University Medical Center, Wood Ridge, NJ
Maor Shir1, Marissa Heyer, MD1, Ryan Kaple, MD1, Craig Basman, MD, FACC, FSCAI1 and Chris S Abrahim, MD2, (1)Hackensack University Medical Center, Hackensack, NJ, (2)Hackensack University Medical Center, Wood Ridge, NJ
Keywords: TAVI/TAVR/Aortic Valve
Background
Electrical disturbances are commonly seen during transcatheter aortic valve replacement (TAVR) despite efforts to reduce permanent pacemaker (PPM) placement. This study set out to explore the effectiveness of Neural Networks' predictive capabilities regarding PPM implantation 30 days post-TAVR.
Methods
430 patients, without prior PPM, who underwent TAVR at a large tertiary center from 01/2017 to 06/2021 were included in the study. Features used in the Neural Network included prior conduction abnormalities from EKGs, prior medical history, and CT scan measurements. The sample was randomly split into a training data set, validation data set, and testing data set, ensuring the percentage of PPMs implanted was consistent in each data set through stratification. The three Neural Networks were independently trained using the training data set. Early stopping was used to prevent overfitting using the validation data set, in order to determine the best model to test using the test data set. The three Neural Networks varied in their activation functions and number of nodes per layer (Figure).
Results
Of the 430 patients, 14.4% had a PPM implanted 30 days post TAVR. The area under the curve (AUC) the three Neural Networks achieved on the test data set were (AUC: 0.798), (AUC: 0.805), and (AUC: 0.786) respectively.
Conclusions
All three Neural Network models yielded accurate predictions for PPM implantation post-TAVR. Further studies are required to implement such artificial intelligence models to help predict the need for PPM after TAVR for clinicians and patients.
