Aim
Transcatheter aortic valve implantation (TAVI) is widely used for the treatment of severe aortic stenosis, particularly in older and high-risk patients. However, post-procedural conduction disturbances remain a relevant complication and may require permanent pacemaker implantation (PPI). This study evaluates whether deep-learning-assisted ECG analysis can predict PPI after TAVI using ECG signals alone, reflecting the routine availability of ECGs in clinical care.
Methods
This retrospective multicentre study used 1042 pre- and 1243 post-procedural ECG signals of patients undergoing TAVI to train a deep-learning neural network, consisting of a multi-layer convolutional block and recurrent LSTM network. ECGs were obtained from 1,070 patients, with a mean age of 81 years and a standard deviation of 6.7.
Separate models were trained using either pre-procedural or post-procedural ECGs. For each training run, the ECG data were divided into an 80/20 split, with subsets grouped by patient and stratified by label. To account for potential split-wise bias, five-fold cross-validation was employed. Final performance metrics were calculated as the average performance across models trained on the different folds.
Given the substantial class imbalance in the dataset, with positive labels representing approximately 14% of the samples, the positive class in the training subset was oversampled to match the number of negative samples. Models were trained locally using a TensorFlow backend, with experiment tracking performed via MLflow.
Results
Models trained on pre-procedural ECGs achieved an average validation AUC of 0.6218, a validation AUPRC of 0.2584, and an accuracy of 0.6876, indicating that predictive patterns above random chance could be learned from the ECG signal. In comparison, models trained on post-procedural ECGs demonstrated substantially higher generalisability, achieving a validation AUC of 0.7806, an AUPRC of 0.3846, and an accuracy of 0.7734.
Conclusions
ECG-based deep-learning models showed potential for identifying patients at risk of permanent pacemaker implantation after TAVI. Pre-procedural ECGs provided limited but above-chance predictive information, whereas post-procedural ECGs yielded substantially better discrimination, suggesting that procedure-related conduction changes are more informative for prediction. These findings support further investigation of ECG-only models for post-TAVI risk stratification, but external validation and assessment of clinical utility are required before clinical use.
