A multi-scale convolutional LSTM-dense network for robust cardiac arrhythmia classification from ECG signals

Abstract

Cardiac arrhythmias are irregular heart rhythms that, if undetected, can lead to severe cardiovascular conditions. Detecting these anomalies early through electrocardiogram (ECG) signal analysis is critical for preventive healthcare and effective treatment. However, the automatic classification of arrhythmias poses significant challenges, including class imbalance and noise interference in ECG signals. This paper introduces the Multi-Scale Convolutional LSTM Dense Network (MS-CLDNet) model, an advanced deep-learning model specifically designed to address these issues and improve arrhythmia classification accuracy and other relevant metrics. This paper aims to develop an efficient deep-learning model, MS-CLDNet, for accurately classifying cardiac arrhythmias from electrocardiogram (ECG) signals. Addressing challenges like class imbalance and noise interference, the model integrates bidirectional long short-term memory (LSTM) networks for temporal pattern recognition, Dense Blocks for feature refinement, and Multi-Scale Convolutional Neural Networks (CNNs) for robust feature extraction. To achieve accurate classification of different types of arrhythmias, the Classification Head refines these extracted features even further. Utilizing the MIT-BIH arrhythmia dataset, key pre-processing techniques such as wavelet-based denoising were employed to enhance signal clarity. Results indicate that the MS-CLDNet model achieves a classification accuracy of 98.22 %, outperforming baseline models with low average loss values (0.084). This research highlights how crucial it is to combine sophisticated neural network architectures with efficient pre-processing techniques to improve the precision and accuracy of automated cardiovascular diagnostic systems, which could have important healthcare applications for early and accurate arrhythmia detection.

Keywords: Arrhythmia classification; Class imbalance; Deep learning; ECG signal processing; Wavelet denoising.