Classification of Bundle Branch Blocks with QRS Templates Extracted from 12-lead ECGs

Abstract

Being non-invasive, cheap and widely available, the 12-lead electrocardiogram (ECG) is a standard method to assess cardiac function. Still, its reliable interpretation requires specialized knowledge and experience, rendering a second opinion valuable. We evaluated the performance of machine learning based classification of 11,705 healthy and bundle branch block 12-lead ECGs from 3 open databases. For each lead of the ECG signal, a representative QRS-complex template was extracted automatically. Principal component analysis (PCA) was applied to the concatenated, normalized and rescaled QRS signals to reduce their dimensionality. Multilayer perceptron and support-vector machine classifiers were trained using the principal components of weighted and non-weighted QRS template signals as input data. Classifiers achieved F1 scores between 0.92 and 0.96 on the test set for different input configurations. Anomaly based weighting slightly improved the performance of the classifiers. Neither class-wise PCA for feature extraction nor adding information on sex, gender and electrical heart axis to the input data yielded considerable improvement of the F1 scores. The achieved classification accuracy is similar to deep learning classifier performances and should generalize robustly to other ECG datasets. Our results suggest that this simple and well interpretable approach based on morphological signal characteristics is suitable for automatically and non-invasively identifying bundle branch block pathologies in clinical or smart electronics contexts.

Keywords

• 12 lead ecg
• qrs
• classification
• bundle branch block
• preprocessing
• machine learning