JMIR Publications

ABSTRACT

Background:

Left bundle branch block (LBBB) remains a common complication after transcatheter aortic valve replacement (TAVR). Traditional methods to predict the occurrence of LBBB have limitations, especially as TAVR expands to low-risk patients, necessitating new approaches.

Objective:

This study aims to develop a machine learning (ML)-based model to predict LBBB after TAVR, using diverse clinical and imaging data.

Methods:

Data from 242 TAVR patients excluding pre-existing LBBB or prior pacemaker implantation across three centers were retrospectively analyzed. A transformer-based ML model, integrating tokenizing and classification layers, was developed to identify significant predictive features. The OURS method, which utilizes feature sets, were evaluated against traditional and domain knowledge-driven methods using key metrics: accuracy, precision, recall, and F1-score.

Results:

The proposed OURS approach demonstrated balanced performance, with a gradient boosting algorithm demonstrating an accuracy of 78.05%, along with the most balanced precision, recall, and F1 scores at 47.78 ± 0.08, 53.47 ± 0.09, and 50.46 ± 0.08, respectively, outperforming conventional models. This method also showed strong capability in identifying significant features suggested by medical knowledge, achieving an accuracy of 68.11%, precision of 60.00%, recall of 46.15%, and an F1-score of 52.17%. Significant features identified exclusively included height, peripheral vascular disease, Society of Thoracic Surgeons Predicted Risk of Mortality score, sinotubular junction area, diameter of the right coronary cusp, and coronary heights as measured by computed tomography.

Conclusions:

This study demonstrates that ML algorithms, particularly the proposed OURS method, could effectively predict LBBB risk post-TAVR. Incorporating diverse clinical data and advanced feature selection enhances predictive accuracy, offering potential for tailored clinical strategies. Clinical Trial: Not applicable.