Document Type : Original/Review Paper

Authors

• AGHIL Kashir Taghartapeh ¹
• Nader Javadifar ¹
• Ali Harimi ²
• Seyed Mehdi BagheriMofidi ³
• Aziz Kalteh ¹

¹ Department of Electrical Engineering, AK.C., Islamic Azad University, Aliabad Katoul, Iran.

² Department of Electrical Engineering, Shahrood Branch, Islamic Azad University, Shahrood, Iran.

³ Department of Electrical Engineering, AK.C., Islamic Azad University, Aliabad Katoul, Iran

Abstract

This study investigates the effectiveness of integrating nonlinear dynamical representations derived from reconstructed phase space (RPS) analysis with deep convolutional neural networks for phonocardiogram classification. It evaluates how the nonlinear dynamic information present in cardiac signals can be captured through the proposed unified framework that embeds RPS within a deep learning architecture. Utilizing the ResNet-34 model, the method efficiently extracts critical spatial and temporal patterns from these enhanced tensors, offering a novel integration of nonlinear system dynamics with deep feature learning for improved diagnostic performance. Performance evaluation conducted on the PhysioNet Challenge dataset demonstrates the method's capability to achieve high accuracy, precision, recall, and F1 scores consistently across different tensor dimensions and time lag parameters. Best results were obtained using tensors with dimension 4 and time lags of 3 and 5, underlining the model's robustness and flexibility in accommodating variability inherent to PCG signals. With an F1 score of 93.38%, the proposed method performs competitively within the upper range of current state-of-the-art techniques.

Keywords

• Phonocardiogram (PCG) classification
• Reconstructed Phase Space (RPS)
• Deep Convolutional Neural Networks (DCNNs)

Main Subjects

• H.6.3.3. Pattern analysis