Utilizing Pretrained Convolutional Neural Networks for Classification of Lorenz Plots of RR Intervals

Title

Utilizing Pretrained Convolutional Neural Networks for Classification of Lorenz Plots of RR Intervals

Authors

Qingqing Song1,*; Shaoliang Xia1 ; Zhen Wu1
1 Faculty of Applied Mathematics and Computer Science, Belarusian State University, 220030, Belarus

Published In

1. Journal of Investigative Medicine
2. 3rd International Conference on Life, Health and Modern Medicine

Published Date

February 1, 2025

Abstract

Background: Heart rate variability (HRV) and the Lorenz plots of RR intervals are the key indicators for evaluating heart health. The Lorenz plots can effectively describe the dynamic changes of heart rate, but their classification is challenging. At present, mathematical calculation is the main classification method, but there are some problems, such as the difficulty of regional division, complex modes and so on. Convolutional neural network model has been proved to be effective in many problems, but its application in the classification of Lorenz plots is relatively limited.
Subjects and Methods: The purpose of this study is to propose an advanced method for classifying the Lorenz Plots of RR Intervals using the pretrained convolutional neural network model, and to classify them according to the morphological characteristics of Lorenz plots. This method aims to improve the diagnostic accuracy and efficiency of cardiac diseases such as arrhythmia and heart failure. It is applied to the analysis and interpretation of ECGs in various clinical diagnostic environments.
In this study, the convolutional neural network model based on the AlexNet structure and the hyperparameters optimization algorithm are used to process the pre-processing and enhanced Lorenz plots dataset for 7-class classification and 19-class classification tasks. Specifically, in the preprocessing stage, this study preprocesses and enhances the original image, including removing the text, reference lines and borders in the image, and modifying the image to black-and-white mode. In the enhancement phase, this study only uses small amplitude rotation, Gaussian blur, smoothing and other methods to enhance the dataset at a 1:11 ratio. In the pre-training stage, we adjust the input of the model to 1281281. In the hyperparameters optimization stage, the grid search enhanced with coordinate ascent method is used to adjust the three hyperparameters of batch size, learning rate and epochs to obtain the optimal hyperparameters combination for the model.
Results: In this study, for the 7-class classification task, when the values of learning rate, batch size and epoch times are 1e-4, 5 and 10 respectively, the accuracy of the model on the validation set reaches 99.45%. For the 19-class classification task, when the values of the three hyperparameters are 1e-5, 5 and 13 respectively, the accuracy of the model on the validation set reaches 97.77%. The method used in this study is superior to the current five classification methods, and shows significant advantages in complex multi-classification tasks, which proves the effectiveness of the method.
Conclusions: The proposed method proves the effectiveness of the pretrained convolutional neural network, especially the AlexNet structure convolutional neural network, in the Lorenz plots classification task. Expanding the sample size of dataset and exploring more complex CNN model structure are helpful for further research. This study is helpful to develop accurate and effective diagnostic tools to meet the urgent challenges of heart health.

Identifiers

• DOI: 10.1177/10815589241311635
• WoS: 001460875600013

Links

Sega Journals
https://journals.sagepub.com/doi/10.1177/10815589241311635