Predicting of His-Ventricle (HV) interval in patients with syncope and bifascicular block on the 12 lead ECG using artificial intelligence

Patients that present with syncope of unknown etiology and bifascicular block are at risk of developing complete heart block (CHB). Electrophysiological study (EPS) is often used to measure their HV interval, that correlates with the risk of subsequent CHB and guide prophylactic pacemaker (PM) implantation.

Currently, HV interval that warrants PM implantation cannot be predicted from the surface ECG and is measured by an invasive EPS. Artificial Intelligence (AI), using deep learning may predict the HV interval in this scenario non-invasively from the 12 lead ECG.

To predict a pathological HV interval using AI algorithms and deep machine learning from ECG in patients with syncope and bifascicular block on their 12 lead ECG

Study patients were ≥18 years of age with syncope of unexplained etiology and bifascicular block on their ECG who admitted between January 1st, 2012, to December 31st, 2022. All the patients underwent EPS to measure HV interval. A pathological HV interval defined above 70ms and followed by PM implantation.

We applied preprocessing to clean the ECG data, extracting essential properties and removing damaged records to retain only high-quality 12-lead ECG’s. Additionally, we captured R-peaks from the signals, which are critical for analyzing cardiac activity. For analysis, we used a Residual Network (ResNet), a Convolutional Neural Network (CNN) designed to extract meaningful patterns from these features and provide accurate predictions.

A total number of 109 patients and a total of 1,472 ECG records were available. The mean age was 74 years and 75% were males. The most common comorbidities were diabetes mellitus, hypertension and ischemic heart disease.

Our results demonstrated strong predictive capabilities of the HV interval. The mean AUC was 0.77 (Figure 1a), showing good overall performance, and the mean PR-AUC (Figure 1b) was 0.67, reflecting reliable precision-recall balance between sensitivity and PPV. PR-AUC measures the model’s ability to correctly identify patients who require a pacemaker (sensitivity) while minimizing false positives (PPV), making it particularly useful when there is an imbalance between patients who need and do not need a pacemaker. At the optimal AUC threshold of 0.85, the model had a sensitivity (recall) of approximately 0.8 and specificity of approximately 0.5. The PPV (precision) was 0.7 and F2-score was 0.75 (Figure 2), which prioritizes sensitivity over positive predictive value (PPV). The F2 score is declining beyond this point, while specificity is increasing. Metrics across thresholds highlighted this threshold as optimal for identifying patients who may require a pacemaker.

AI algorithms and deep machine learning were helpful in predicting HV interval from surface ECG in patients with syncope and bifascicular block. Nevertheless, our results call for a larger study to further confirm our findings.