Enhanced electrocardiogram machine learning-based classification with emphasis on fusion and unknown heartbeat classes

Abstract

Building an electrocardiogram (ECG) heartbeat classification model is essential for early arrhythmia detection. This research aims to build a reliable model that can classify heartbeats into five heartbeat types: normal beat (N), supraventricular ectopic beat (SVEB), ventricular ectopic beat (VEB), fusion beat (F), and unknown beat (Q), with a focus on enhancing the predictions of the uncommon Q and F heartbeats. The base dataset used is the MIT-BIH SupraVentricular Database, which was used to train and compare the performance of five machine learning models: logistic regression, Random Forest (RF), K-nearest neighbor, linear support vector machine, and linear discriminant analysis. In addition to using the synthetic minority oversampling technique, data extracted from multiple databases for the F and Q classes were combined with the original base dataset. These methods resulted in significant improvement in the recall for the rare F and Q classes when compared to the literature. The RF algorithm produced the best performance with an accuracy of 97% and recall values equal to 97%, 93%, 95%, 95%, and 30% for N, SVEB, VEB, F, and Q, respectively.

Enhancing ECG Heartbeat classification with feature fusion neural networks and dynamic minority-biased batch weighting loss function

“…Although many methods attempt to solve the problem of class imbalance by improving the respective classification model (Niu et al 2020, their eventual performance shows that simply increasing the complexity of the model often results in better recognition of majority classes, while the recognition of minority classes is not significantly improved. Alternative approaches, such as synthesis or sampling methods employed by Acharya et al (2017), Al-Mousa et al (2023), and Pramukantoro and Gofuku (2022), often lead to overfitting or underfitting, limiting their practical utility. Researchers have explored other strategies to address the imbalanced classes challenges, such as integrating local features into network structures and modifying loss functions.…”