Elastic Net Regularization and Gabor Dictionary for Classification of Heart Sound Signals using Deep Learning
Summary: arXiv:2604.12483v1 Announce Type: cross
Abstract
In this article, we propose the optimization of the resolution of time-frequency atoms and the regularization of fitting models to obtain better representations of heart sound signals. This is done by evaluating the classification performance of deep learning (DL) networks in discriminating five heart valvular conditions based on a new class of time-frequency feature matrices derived from the fitting models.
Introduction
Heart diseases have become a significant public health concern worldwide. Early detection and accurate diagnosis are crucial for effective treatment. Recent advancements in deep learning (DL) technologies offer new avenues for analyzing heart sound signals, providing a non-invasive method for diagnosing heart conditions.
Methodology
This study focuses on the following key components:
- Time-Frequency Feature Matrices: We derive a new class of time-frequency feature matrices based on fitting models optimized for heart sound signal representations.
- Elastic Net Regularization: We utilize elastic net regularization to enhance model fitting, balancing between L1 and L2 penalties to improve sparsity and prediction accuracy.
- Gabor Dictionary: An overcomplete dictionary of Gabor atoms is employed to extract features from the heart sound signals, enabling better discrimination of valvular conditions.
Deep Learning Architectures
We explore two distinct deep learning architectures:
- The first architecture consists of a 1D convolutional neural network (CNN) layer followed by a long short-term memory (LSTM) layer, designed to capture temporal dependencies in the heart sound signals.
- The second architecture incorporates both 1D and 2D CNN layers followed by an LSTM layer, allowing for more complex feature extraction and improved performance.
Training Algorithms
Two training algorithms are utilized in this study:
- Stochastic Gradient Descent with Momentum (SGDM): A widely used optimization method that accelerates SGD by adding a fraction of the previous update.
- Adaptive Moment Estimation (ADAM): Combines the advantages of two other extensions of SGD, maintaining an exponentially decaying average of past gradients and squared gradients.
Results
Extensive experimentation was conducted using a database containing heart sound signals of five heart valvular conditions. The second architecture achieved the best classification accuracy of 98.95% when trained with ADAM. This performance was attained using feature matrices derived from optimal models obtained with a Gabor dictionary emphasizing high-time low-frequency resolution while imposing sparsity on the models.
Conclusion
The findings of this study highlight the effectiveness of using elastic net regularization combined with a Gabor dictionary for enhancing the classification of heart sound signals. The proposed deep learning architectures demonstrate significant potential in improving diagnostic capabilities for heart valvular conditions, paving the way for further research in the field of medical signal analysis.