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

Area of Science:

• Biomedical Engineering
• Artificial Intelligence in Medicine
• Pediatric Cardiology

Background:

• Congenital Heart Defects (CHDs) affect approximately 1% of neonates globally.
• Traditional CHD diagnosis relies on expert stethoscope auscultation, risking oversight of subtle acoustic signs.
• Advanced, non-invasive diagnostic tools are crucial for timely intervention in pediatric populations.

Purpose of the Study:

• To develop and evaluate a deep-learning framework for the early diagnosis of Congenital Heart Defects.
• To leverage cardiac acoustic signals captured via stethoscopes for CHD detection.
• To improve upon traditional diagnostic methods by utilizing advanced AI techniques.

Main Methods:

• Cardiac auditory signals were processed into time-frequency representations using Mel-Frequency Cepstral Coefficients (MFCCs).
• A hybrid deep learning model combining Convolutional Neural Networks (CNNs) and Long Short-Term Memory (LSTM) networks was employed.
• The model architecture was designed to effectively extract features and model temporal dependencies in audio data.

Main Results:

• The deep learning model achieved a high accuracy of 98.91% in the early detection of CHDs.
• The study highlights the potential of cardiac acoustics, analyzed via AI, for early CHD diagnosis.
• The research utilized a publicly available dataset, promoting reproducibility and further development.

Conclusions:

• This AI-driven approach using stethoscope audio shows significant promise for early CHD detection in neonates.
• The framework offers a potential complementary tool to existing diagnostic methods like ECG and PCG.
• Further research and clinical validation can enhance the impact of acoustic AI on pediatric cardiac care.