Reliable Multimodal Heartbeat Classification using Deep Neural Networks

Abstract

Copyright © 2023 Authors. Arrhythmias are deviations from the normal heart rhythm with impact on the cardiovascular health. Their prompt detection plays an important role in mitigating potential negative outcomes, particularly in patients in the intensive care units (ICU). Heartbeat detection has mainly been focused on electrocardiogram (ECG) signals. However, ICU patient mobility frequently leads to disconnection of certain ECG leads, potentially compromising the accurate heartbeat classification. Arterial line blood pressure (ABP) and central venous pressure (CVP) signals are routinely monitored in ICU patients. Changes in the ABP and CVP suggest alterations in the haemodynamic status and cardiac function of the patients. Thus, leveraging these signals for heartbeat classification, either independently or in conjunction with ECG data, present a viable approach to ensure that even in scenarios where ECG signals are unavailable, alarm systems alerting healthcare providers of arrhythmias remain functional. Moreover, while many researchers have successfully created methodologies to accurately classify heartbeats including paced beats, none were able to distinguish various sub-classes of paced heartbeats. A more comprehensive distinction is crucial as it not only aids in the identification of pacing settings but also facilitates the detection of inadequate pacing settings, a critical aspect in patient care. In this paper, we employ a hybrid model using long-short term memory networks (LSTM) and convolutional neural network (CNN), along with different residual CNN (ResNet) models for multimodal arrhythmia classification and for comprehensive paced heartbeats classification. When using all three channels, ResNet50 achieved the best accuracy of 99.58% on 5 different arrhythmia classes, whereas ResNet34 achieved an accuracy of 93.82% on 12 paced classes. The significant efficiency of utilizing ABP and CVP signals independently for classification, was also highlighted. ResNet50 was trained with ABP and CVP signals independently and correctly identified arrhythmias with an accuracy of 98.79% and 96.67%, respectively. For classifying 12 different paced heartbeats, ResNet34 achieved 74.04% accuracy with ABP signals and 74.38% with CVP signals. Moreover, the same ResNet50 model was trained on the MIT-BIH arrhythmia database, achieving an accuracy, sensitivity, and precision of 98.78%, 98.77% and 98.80%, which demonstrates the scalability of the proposed model.