An Ensemble of Machine Learning Models for Multilabel Classification of Cardiovascular Diseases by ECGs

The study focuses on automating ECG analysis, a crucial tool for cardiac evaluation and treatment decisions. Two classifiers were developed to detect cardiac conduction disorders and myocardial infarctions from 12-lead ECGs. These classifiers were trained on three open datasets after ECG preprocessing to remove noise and select relevant features. The first classifier (10 classes) identifies conditions of the ventricular conduction system: normal ECG, LBBB, ILBBB, premature ventricular complex, LAFB, LPFB, IRBBB, RBBB, nonspecific intraventricular conduction disturbance, and ventricular preexcitation. The second classifier (5 classes) distinguishes between non-infarcted ECGs and various types of myocardial infarction locations. For each QRS complex, we calculated statistical, time-domain, frequency features, and scalograms. Then we trained classifiers combining XGBoost for statistical features, an autoencoder+neural network for deep-learning features, and ResNet for scalograms. An ensemble approach was employed to get the final class prediction. We show that the developed ensemble model achieve the mean F1-score of 0.70 and 0.76 for the first and second classifier respectively. The study's key strength lies in its diverse feature extraction methods, enhancing the predictive power of machine learning models.