Classification of heart sounds based on audio signals using deep learning methods
DOI:
https://doi.org/10.15330/itee.2026.4.07Keywords:
audio signal processing, heart sound classification, random forest, gradient boosting, convolutional neural networks, deep learning, machine learningAbstract
Cardiovascular diseases remain the leading cause of death in developed countries worldwide, making the development of effective diagnostic methods a critically important task for public health systems. The development of diagnostic methods and risk assessment based on heart rhythm data analysis occupies an important place in interdisciplinary research created at the intersection of mathematics, data analysis, and medicine. The results of such research find wide application in developing applications for mobile smartphones and various portable devices, where increasing attention is being paid to the problems of self-diagnosis and early detection of heart diseases.
This work investigates the classification of heart sound audio signals collected using a stethoscope for the automatic diagnosis of diseases caused by improper heart function. The study utilized the Classifying Heart Sounds dataset containing 813 audio recordings ranging from 1 to 30 seconds in duration, distributed across four categories: normal heart sounds, murmurs, extrasystoles, and artifacts. The methods considered in this work enable their use for recording heartbeats using mobile devices and subsequent analysis through specialized applications.
Data preprocessing included normalization of recording duration and dataset augmentation through the generation of synthetic examples using time-stretching coefficients of 1.2 and 0.8. Mel-frequency cepstral coefficients (MFCC) with 20 components were employed for feature extraction. Various classification approaches were compared: Random Forest and XGBoost gradient boosting algorithms achieved accuracies of 89% and 85% respectively, revealing difficulties in distinguishing between extrasystole and artifact classes.
The main focus was on developing a hybrid deep neural network architecture CNN-LSTM that combines convolutional layers for spatial feature extraction with recurrent LSTM layers for temporal dependency analysis. The model consists of three blocks of Conv1D convolutional layers (1024, 512, 256 filters), MaxPooling and BatchNormalization layers, followed by an LSTM layer with 128 units and Dense fully connected layers with Dropout regularization for preventing overfitting. The work was implemented in Python using TensorFlow, Librosa, Pandas, and SciPy libraries. Results demonstrate that the proposed architecture achieves a classification accuracy of 95% on the test dataset, significantly outperforming traditional machine learning methods, particularly in detecting challenging extrasystole (F1-score 0.86) and artifact (F1-score 1.0) classes.
References
W. Chen, Q. Sun, X. Chen, G. Xie, and H. Wu, “Deep learning methods for heart sounds classification: A systematic review,” Entropy, vol. 23, no. 6, Art. no. 667, May 2021. doi: https://doi.org/10.3390/e23060667 .
W. Zhang, J. Han, and S. Deng, “Heart sound classification based on scaled spectrogram and tensor de-composition,” Expert Syst. Appl., vol. 84, pp. 220–231, Oct. 2017. doi: https://doi.org/10.1016/j.eswa.2017.05.014 .
J. Chen, X. Dang, and M. Li, “Heart sound classification method based on ensemble learning,” in Proc. 2022 16th IEEE Int. Conf. Signal Process. (ICSP), Beijing, China, 2022, pp. 8–13. doi: https://doi.org/10.1109/ICSP54964.2022.9778383 .
T. Tuncer, S. Dogan, R.-S. Tan, and U. R. Acharya, “Application of Petersen graph pattern technique for automated detection of heart valve diseases with PCG signals,” Inf. Sci., vol. 565, pp. 91–104, Jul. 2021. doi: https://doi.org/10.1016/j.ins.2021.01.088 .
P. Aparna, G. Jayalaxmi, and P. Vishwanath, “Heart sound classification system using deep-learning neural networks,” in Proc. 4th Int. Conf. Frontiers Comput. Syst. (COMSYS 2023), Singapore: Springer, 2024, pp. 647–658. doi: https://doi.org/10.1007/978-981-97-2614-1_46 .
Yaseen, G.-Y. Son, and S. Kwon, “Classification of heart sound signal using multiple features,” Appl. Sci., vol. 8, no. 12, Art. no. 2344, Nov. 2018. doi: https://doi.org/10.3390/app8122344 .
J. Huang, B. Chen, B. Yao, and W. He, “ECG arrhythmia classification using STFT-based spectrogram and convolutional neural network,” IEEE Access, vol. 7, pp. 92871–92880, 2019. doi: https://doi.org/10.1109/ACCESS.2019.2928017 .
D. Kozłowski, “Method in the chaos – a step-by-step approach to ECG interpretation,” Eur. J. Transl. Clin. Med., vol. 1, no. 1, pp. 74–90, 2018. doi: https://doi.org/10.31373/ejtcm/92255 .
T. S. M. Tsang, W. P. Abhayaratna, M. E. Barnes, et al., “Prediction of cardiovascular outcomes with left atrial size: Is volume superior to area or diameter?,” J. Am. Coll. Cardiol., vol. 47, no. 5, pp. 1018–1023, Mar. 2006. doi: https://doi.org/10.1016/j.jacc.2005.08.077 .
A. Y. Hannun, P. Rajpurkar, M. Haghpanahi, et al., “Cardiologist-level arrhythmia detection and classi-fication in ambulatory electrocardiograms using a deep neural network,” Nat. Med., vol. 25, no. 1, pp. 65–69, Jan. 2019. doi: https://doi.org/10.1038/s41591-018-0268-3
Downloads
Published
How to Cite
Issue
Section
License
Copyright (c) 2026 Юрій Петриняк, Ігор Пірко

This work is licensed under a Creative Commons Attribution 4.0 International License.