Early diagnosis of acyanotic congenital heart disease using fused feature based hybrid model with Bahdanau attention mechanism

Critical information about congenital heart diseases can be obtained from heart sounds. Heart sounds act as a diagnostic indicator for heart health. Heart disease detection from heart sounds is a low-cost, non-invasive diagnostic method. However, this method is challenging and requires highly experienced specialist doctors. The presence of an innocent murmur, which is non-pathological, makes the diagnostic process difficult in children, and all subjects with a murmur are referred for an echocardiogram to confirm the diagnosis. The study proposes an attention-based deep learning method for the early detection of acyanotic congenital heart disease during auscultation. When developed into a diagnostic support tool, the proposed model has the potential to help doctors distinguish between pathological and innocent murmurs, which could support their decision on whether echocardiographic evaluation is necessary and potentially reduce unnecessary echocardiogram referrals. The focus of the study is on the detection of heart diseases and the classification of heart sounds into five distinct classes: Normal, Innocent Murmur, Atrial Septal Defect, Ventricular Septal Defect, and all other pathologies grouped into the Abnormal class. The research endeavours to formulate a paediatric heart sound dataset comprising these five well-defined categories, as the publicly available multi-class paediatric heart sound datasets are limited. The study introduces a novel method that combines clinical and frequency-domain features of heart sounds. The fused features are processed using a hybrid model consisting of a 2D Convolutional Neural Network and a Bidirectional Long Short-Term Memory followed by the Bahdanau attention mechanism. The work also studied the scaled dot attention mechanism, but the Bahdanau model outperformed with a 10-fold cross-validation accuracy of 95.6%, supported by a 94% Kappa Coefficient and Matthews Correlation Coefficient.