Classification of Short-Segment Pediatric Heart Sounds Based on a Transformer-Based Convolutional Neural Network

• Published in: IEEE access Vol. 13; pp. 93852 - 93868
• Main Authors: Hassanuzzaman, Md, Ghosh, Samit Kumar, Hasan, Mohammad Nurul Akhtar, Mamun, Mohammad Abdullah Al, Ahmed, Khawza I., Mostafa, Raqibul, Khandoker, Ahsan H.
• Format: Journal Article
• Language: English
• Published: Piscataway IEEE 2025; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Abnormalities; Artificial neural networks; attention transformer; Blood vessels; Classification; congenital heart disease; Ethics; Feature extraction; Heart; Heart diseases; Heart valves; mel-frequency cepstral coefficients; Neural networks; Pediatrics; Phonocardiogram; Phonocardiography; Public health; Recording; Segments; signal duration; Signal quality; Stethoscope; Transformers; Valves
• ISSN: 2169-3536; 2169-3536
• DOI: 10.1109/ACCESS.2025.3573870

Congenital heart diseases (CHDs), caused by structural abnormalities in the heart and blood vessels, pose a significant public health concern and contribute significantly to the socioeconomic burden, particularly in pediatric populations. Phonocardiograms (PCGs), as a non-invasive and cost-effective diagnostic modality, capture vital acoustic signals that reflect the mechanical activity of the heart and can reveal pathological patterns associated with various CHD types. This study investigates the minimum signal duration required for accurate automatic classification of heart sounds and evaluates signal quality using the root mean square of successive differences (RMSSD) and the zero-crossing rate (ZCR). Mel-frequency cepstral coefficients (MFCCs) are extracted as features and fed into a transformer-based residual one-dimensional convolutional neural network (1D-CNN) for classification. Experimental results show that a threshold of 0.4 for RMSSD and ZCR yields optimal classification performance, with a minimum signal length of 5 seconds required for reliable results. Shorter segments (3 seconds) lack sufficient diagnostic information, while longer segments (15 seconds) may introduce additional noise. The proposed model achieves a maximum classification accuracy of 93.69% with 5-second signals.