Automatic Left Bundle Branch Block Diagnose Using a 2-D Convolutional Network

Abstract

Left bundle branch block (LBBB) patients are the population that benefits most from cardiac resynchronization therapy (CRT), a therapy applied in heart failure. However, CRT presents about 40% non-responders rates. A plausible explanation to this fact, is a precarious LBBB diagnosis. QRS duration is currently one of three pillars in LBBB diagnosis. However, ECG morphology is severely altered in the presence of LBBB, affecting seriously the process of ECG delineation. Thus, QRS duration becomes a highly unreliable measure in LBBB diagnosis. Herein, we propose a LBBB classification framework complettely independent of temporal measures. In this line, a 2-D convolutional network (CNN) was utilized to separate strict LBBB patients from (not strict/not) LBBB patients, obtained from a subset of the Multi-center Autonomic Defibrillator Implantation (MADIT) trial. In order to fit the 2-D architecture, we fed the CNN with 10 s- spectrograms, constructing and validating 6 separated unilead models, one per precordial lead. From all analyzed models, the one using lead $V_1$ turned out to be the most informative. The latter, produced an 89% accuracy and 90% positive predictive value. These results encourage the use of such statistical models to provide a more reliable and automated LBBB diagnosis.