Abstract 16427: Despite Machine Learning Models, QRS Duration Remains the Superior ECG Criterion for Predicting Left Ventricular Dilation in Patients With Left Bundle Branch Block

Abstract

Introduction: The utility of ECG to diagnose left ventricular (LV) dilation in patients with left bundle branch block (LBBB) is not known. We sought to compare the diagnostic yield of ECG using (i) QRS duration, (ii) published LVH criteria, and (iii) machine learning (ML) models to detect increased left ventricular end diastolic volume indexed (LVEDVi) in the setting of LBBB. Hypothesis: ML is superior to QRS duration and LVH criteria to detect increased LVEDVi among LBBB. Methods: 12-lead ECGs were processed to reconstruct orthogonal X, Y, Z leads using Kors’s matrix and obtain root-mean-squared (3D) ECG. R wave, S wave and overall amplitudes, voltage-time-integrals (VTIs), and other ECG features were extracted from all ECG leads. ML algorithms [logistic regression (LR), support vector classifier (SVC), decision trees (DT), random forest (RF), gradient boosted machine (GBM) and boosted trees (BT)] were trained to predict increased LVEDVi (women >61 mL/m 2 , men >74 mL/m 2 ) from ECG features on a training set of 2668 ECGs with typical LBBB and echocardiogram within 45 days before or after ECG. LVEDV was measured using ASE biplane method of discs. We obtained ROC AUCs for prediction of increased LVEDVi by (i) QRS duration, (ii) published LVH criteria, and (iii) ML models in a separate validation set of adults with typical LBBB. Results: Among the validation set of 413 adults (53% women, age 73±12 yr) with LBBB, QRS duration alone had a higher AUC (women 0.668, men 0.699) for diagnosing increased LVMi compared to standard LVH criteria (Table). The best ML model (RF with overall AUC 0.694) did not substantially outperform QRS duration alone. Conclusions: In patients with LBBB, QRS duration ≥150 in women and ≥160 in men is a superior predictor of LV dilation than traditional ECG-based LVH criteria, with no additional value added by ML methods.