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

Background: Heart rate variability (HRV) and RR-interval Lorenz plots are key indicators for assessing cardiac health status. Lorenz plots can effectively characterize the dynamic changes of heart rate, yet their classification presents challenges. Currently, mathematical computation serves as the primary classification approach, but issues such as difficulty in region delineation and pattern complexity persist. Convolutional neural network models have proven effective across numerous problems; however, their application to Lorenz plot classification remains relatively limited.

Approach and Methods: This study aims to propose an advanced Lorenz plot classification method based on pretrained convolutional neural network models, classifying Lorenz plots according to their morphological characteristics. The method is intended to improve the diagnostic accuracy and efficiency for cardiac conditions such as arrhythmias and heart failure, and is applicable to ECG analysis and interpretation across various clinical diagnostic settings. This study employs a convolutional neural network model based on the AlexNet architecture and a hyperparameter optimization algorithm, conducting 7-class and 19-class classification tasks on a preprocessed and augmented Lorenz plot dataset. Specifically, during the preprocessing stage, original images are preprocessed and augmented, including the removal of text, reference lines, and borders, and conversion to black-and-white mode. During augmentation, only mild rotation, Gaussian blur, and smoothing are applied, augmenting the dataset at a 1:11 ratio. During pretraining, model input is adjusted to 128×128×1. During hyperparameter optimization, the grid search method enhanced with coordinate ascent is employed to tune three hyperparameters — batch size, learning rate, and epochs — to obtain the optimal hyperparameter combination for the model.

Results: In this study, under the 7-class classification task, with a learning rate, batch size, and epochs of 1e-4, 5, and 10 respectively, the model achieves 99.45% accuracy on the validation set; under the 19-class classification task, with the three hyperparameter values set at 1e-5, 5, and 13, the model achieves 97.77% accuracy on the validation set. The proposed method outperforms five existing classification methods, demonstrating marked advantages in complex multi-class classification tasks, thereby validating its effectiveness.

Conclusion: The proposed method confirms the effectiveness of pretrained convolutional neural networks — particularly those based on the AlexNet architecture — for Lorenz plot classification tasks. Expanding the dataset sample size and exploring more complex CNN model architectures will facilitate deeper subsequent research. This study contributes to the development of accurate and effective diagnostic tools to address pressing challenges in cardiac health.