Abstract Body (Do not enter title and authors here): Introduction: Although left bundle branch area pacing (LBBAP) has become a popular approach to physiologic pacing in heart failure (HF), there is still substantial uncertainty regarding ECG predictors of favorable outcomes with this approach. We have shown that machine learning (ML) with functional principal component decomposition (FPCD) generates accurate predictors of outcomes for cardiac resynchronization therapy (CRT), but its effectiveness for LBBAP in HF is unknown.
Research Question: How does ML with FPCD applied to 12-lead ECG QRS waveforms predict improvement in the LVEF with LBBAP in HF patients compared with standard ECG criteria for selective LBBAP?
Aims: The aims were to determine which FPCD lead parameters from 12-lead ECGs before and after LBBAP in HF were associated with improvement in LV function post-pacing and how these predictors compared with known ECG and clinical predictors.
Methods: 12-lead ECGs in paper format were digitized, and FPCD was applied to the QRS waveforms before and after LBBAP in patients with LVEF ≤ 40% and ≥ 80% LBBAP. Logistic regression with the least absolute shrinkage and selection operator (LASSO) (Figure 1) was used to identify the ECG leads with the best FPCD predictors, and these predictors were compared with the commonly used R-wave peak time (RWPT) in lead V6 and other clinical parameters.
Results: In 25 patients (40% female), the median baseline QRS duration was 156 ms (IQR 116 to 170 ms) and the median baseline LVEF was 33% (IQR 28 to 35%). In this cohort, 40% of patients had an absolute LVEF improvement of at least 15% with LBBAP. The ECG leads with the best FPCD parameters to identify patients with a 15% increase in the LVEF after LBBAP were lead III post-pacing and lead aVF pre-pacing (area under the receiver operating characteristic curve [AUROC] for model = 0.88, p=0.003) (Figure 2). The RWPT in V6 (3 V pacing output) had an AUROC of 0.60 for this outcome (p=0.63), and the difference in QRS duration before and after LBBAP had an AUROC of 0.58 (p=0.34). The model with the first FPCD weight for lead III post-pacing and ischemic cardiomyopathy (ICM) had an AUROC of 0.86 (0.0025) (Figure 3).
Conclusion: Machine learning applied to ECG waveforms using FPCD was more effective in identifying LV functional improvement with LBBAP in HF compared with known ECG and clinical parameters. This approach could have an important impact for prognostication, therapy selection, and optimization during the LBBAP procedure.