Deep Learning for Heart Sound Abnormality of Infants: Proof of Conceptual Study of 1D and 2D Representations

Timely diagnosis and treatment of Congenital Heart Defects (CHDs) in pediatric patients are critical, as approximately 1% of neonates worldwide are affected by these anomalies. Traditional stethoscope auscultation relies on the clinician’s skills, which can lead to missed subtle symptoms. This study introduces a deep-learning framework for the early diagnosis of congenital heart disease, utilizing time-series data from cardiac auditory signals captured via stethoscopes. The audio signals were transformed using Mel Frequency Cepstral Coefficients (MFCC) to create time-frequency representations. Our novel architecture combines Convolutional Neural Networks (CNN) for feature extraction with Long Short-Term Memory (LSTM) networks to capture temporal dependencies. This model achieved an impressive accuracy of 98.91% in early disease detection. While methods such as Electrocardiograms (ECG) and Phonocardiograms (PCG) are necessary for confirming diagnoses, previous AI-driven studies have largely focused on ECG and PCG datasets. Our approach emphasizes the potential of using cardiac acoustics for the early diagnosis of CHDs, enhancing clinical outcomes for infants.