Detection of Bundle Branch Block From 12-Lead ECG Using Fifth-Order Tensor-Domain Machine Learning

Bundle branch block (BBB) is a cardiac disease that occurs due to the delay in the heart’s electrical activity during a heartbeat. The early detection of BBB using 12-lead electrocardiogram (ECG) is crucial in clinical studies for monitoring the progress of this disease and initiation of treatment. This article proposes a fifth-order tensor-domain machine learning (FOTDML) approach for the automated detection of BBB using 12-lead ECG recordings. The entire duration of the 12-lead ECG recording of each subject is initially segmented into 12-lead beats using a multilead fusion-based QRS complex detection method. Multivariate fast iterative filtering (MVFIF) decomposes each 12-lead beat into intrinsic mode functions or local components. Then, the continuous wavelet transform is utilized to evaluate the time-frequency representation of the MVFIF mode of the 12-lead beat. A fifth-order tensor containing the information on 12-lead ECG beats, leads, MVFIF-domain local components, frequencies (or scales), and samples is evaluated from the entire duration of the 12-lead ECG recording of each subject. Multilinear singular value decomposition is employed to extract features from the fifth-order tensor. Different machine learning (ML)-based methods are utilized to detect BBB from the fifth-order tensor-domain features of each subject’s entire duration 12-lead ECG recording. The suggested approach is evaluated using 12-lead ECG recordings of subjects from two public databases. The results show that the proposed FOTDML approach has yielded the classification accuracy values of 99.88% and 100%, respectively, for healthy control (HC) versus left BBB versus right BBB and HC versus BBB schemes with the subject-independent hold-out validation strategy. The suggested FOTDML approach has demonstrated higher classification performance than the existing methods to detect BBB using 12-lead ECG.