Intelligent Arrhythmia Classification Using Deep Learning on Multichannel ICU Physiological Signals

Abstract

Arrhythmias is an abnormality that is found in heart rhythm which poses risks to cardiovascular health this abnormality is very critical for ill patients those are in intensive care units (ICUs). Electrocardiogram signals can be used to serve as valuable alternatives or complements to ECG data for arrhythmia detection which ensures a continuous monitoring of ECG signals when they are unavailable. In this study, different deep learning-based approach to classify arrhythmias using a combination of electrocardiogram (ECG), arterial blood pressure (ABP), central venous pressure(CVP) signals, including long short-term memory (LSTM) networks and convolutional neural networks (CNN), with various residual CNNs like ResNet architectures for arrhythmia classification. Among all these models evaluated ResNet50 has achieved the highest training accuracy of 98.77% and validation accuracy of 98.88% from all five of the arrhythmia classes when utilizing all three signal types (ECG, ABP, and CVP). ResNet50 has also demonstrated strong performance results when being trained solely on ABP and CVP signals which have achieved accuracies of 98.79% and 96.67%. Furthermore, when it was applied to the MIT-BIH arrhythmia database on the ResNet50 model, it had an accuracy of 98.88%. These results have highlighted both the scalability and robustness of the different deep learning models it also has shown the potential of ABP and CVP signals that they are reliable inputs for arrhythmia detection.