1Segeberger Kliniken GmbH Herz- und Gefäßzentrum Bad Segeberg, Deutschland; 2Technische Hochschule Mittelhessen Fachbereich Gesundheit Gießen, Deutschland; 3Universitätsklinikum Gießen und Marburg GmbH Medizinische Klinik I - Kardiologie und Angiologie Gießen, Deutschland; 4Klinikum Oldenburg AöR Klinik für Kardiologie Oldenburg, Deutschland
Background:
The most common complication during implantation of transcatheter aortic valve replacement (TAVR) when treating aortic valve stenosis is a compromise of the conduction system, with the need for implantation of a permanent pacemaker (PPM) system (approx. 3 - 19% of cases). Some preprocedural electrocardiogram (ECG) pathologies are suspected to be associated with a higher risk for PPM implantation, providing a basis for calculating a “PPM score”. However, it is sufficiently likely that PPM risk depends on more variables. Due to advancing capabilities of automatic, image-based registration and segmentation in the context of diagnostic parameters, the investigators sought to develop an artificial intelligence (AI)-based algorithm to optimize current risk stratification tools.
Purpose:
The aim of this study is to develop and validate a novel deep learning (DL)-based framework for automatic pre-procedural risk stratification for PPM implantation by using AI.
Methods:
Within our previously established all-comers TAVR database of more than 2000 datasets, we identified 50 consecutive patients in the period of May 2022 to February 2023 who had well-characterized data with standardized pre- and post-procedural ECG recordings (<48 h). We gathered in-office or telephone follow-up data at 30 days, 6 months, and 12 months to evaluate whether PPM implantation had occurred. Within this patient collective 18% needed permanent pacing and an additional 3 patient already had a PPM.
To develop an AI-based stratification system, we will train artificial neural networks integrating pre- and post-procedural ECGs and all available patient characteristics, diagnostic parameters, and procedural aspects. Different neural networks will be used by either varying the training set or adjusting components of the architecture of the networks.
Transfer learning from related classification problems will be considered to reduce the number of cases and generalize the model. For assessment of the performance, the PPM score will be compared with clinical features alone, ECG implementation, and a combination of the two using ROC analysis and AUC calculation.
Conclusions:
Due to a relatively high rate of PPM requirement after TAVR with a vulnerability in the first few days after implantation, accurate prediction of patients at risk will lead to patient-centered and thus optimized therapy. In theory, such a tool will offer an individualized course of treatment for each patient, leading to faster mobilization and hospital discharge as well as a reduction in serious adverse events due to a compromise of the conduction system. In the context of precise medicine, such tools will be valuable and mandatory to support the heart team’s decision and improve therapy and patient outcome in the near future.