Hybrid Deep Learning Model for Paediatric Cardiac Anomaly Detection and Classification

Congenital heart disease (CHD) is a serious health concern that affects a sizeable proportion of newborns worldwide. Detecting CHD during the initial auscultation can expedite the treatment process, enhancing the overall quality of life. The study proposes a hybrid deep learning model that employs the fusion of the Mel Frequency Cepstral coefficient (MFCC) and label encoded auscultation points to classify the phonocardiogram signals into five distinct classes. The use of location corresponding to the loudest murmur as a feature and 5-class classification of paediatric heart sounds using a deep learning algorithm makes this study ground-breaking in this area of research. The proposed methodology was trained and tested with a real-time dataset collected by paediatric cardiologists in the clinical environment. The model achieved 93.44% accuracy in 10 cross-validations, proving its robustness. The model's performance was evaluated using nine metrics, and it gained more than 90% for all the metrics, including Cohen's Kappa score and Matthews's Correlation Coefficient. The results demonstrate the model's efficiency and suitability for the classification of CHD from phonocardiogram signals.