Detecting 'Strict' Left Bundle Branch Block from 12-lead Electrocardiogram using Support Vector Machine Classification and Derivative Analysis

Abstract

Cardiac Resynchronization Therapy (CRT) is generally indicated for heart failure patients with a left bundle branch block (LBBB). 'Strict' LBBB criteria have been proposed as a better predictor of benefit from CRT. Automatic detection of 'strict' LBBB criteria may improve outcomes for heart failure patients by reducing high false positive rates in LBBB detection. This study proposes an algorithm to automatically detect 'strict' LBBB, developed and tested using ECGs made available via the International Society of Computerized Electrocardiology (ISCE) LBBB initiative. The dataset consists of 12-lead Holter ECGs recorded before the therapy from the MADIT-CRT clinical trial. The algorithm consists of multi-lead QRS complex detection using length transform, a support vector machine (SVM) classifier to identify QS- or rS- configurations and identification of mid-QRS notching and slurring by analyzing the variation of first and second derivatives of the signals respectively. The algorithm achieved an accuracy of 80%, sensitivity of 86%, specificity of 73%, positive predictive value (PPV) of 81% and negative predictive value of 79% on the training set. It achieved accuracy, sensitivity, specificity, PPV and NPV of 81%, 88%, 75%, 79% and 85% on the test set. High sensitivity to minor slurring and errors in QRS detection result in low specificity for LBBB detection.