Enhancing cardiovascular diagnostics using fused capsule network for heartbeat sound classification

Heartbeat sound classification plays a crucial role in the early detection of cardiovascular abnormalities. In this study, a novel framework, CRCapsNet, that integrates Convolutional Neural Networks (CNNs), Recurrent Neural Networks (RNNs), and Capsule Networks to enhance classification accuracy and robustness is proposed. Analysis of the proposed model is performed on two datasets: dataset 1, with 832 audio samples in WAV format, and dataset 2, with 3240 heart sound recordings. The pre-processing techniques, including noise addition, time shifts, time stretching, and pitch shifts, are applied to the datasets, and Mel-Frequency Cepstral Coefficients (MFCC) are employed for feature extraction. The spectrograms are passed through a CNN with four convolutional blocks for spatial feature extraction, followed by an RNN module to capture temporal patterns in the heartbeat sequences. A Capsule Network is further incorporated to retain hierarchical relationships that are typically lost in traditional max-pooling operations. The achieved classification accuracies are 88.5% for the CNN, 98.67% for RNN, and an impressive 99.32% and 99.64% for the proposed CRCapsNet model on dataset 1 and dataset 2, respectively, demonstrating its superior performance in heartbeat sound classification. This research underscores the significance of heartbeat sound classification in augmenting traditional diagnostic practices and highlights the role of advanced computational techniques in healthcare innovation. Future directions include exploring multimodal integration and real-time clinical deployment.