Development and validation of an integrated residual-recurrent neural network model for automated heart murmur detection in pediatric populations

Fig. 1. Heart Sound Segmentation. During the analysis, each 10-second sound recording was divided into seven overlapping temporal segments. The segmentation used 2.5-second windows with a 1.25-second overlap between consecutive segments. The neural network model then classified each of these seven windows to determine whether the overall recording represented a normal sound or a murmur.

Fig. 2. Visual analysis of the CQT representations. The Constant-Q Transform revealed a distinctive contrast between (a) normal and (b) abnormal heart signals. Normal heart signals are characterized by more compact energy spectrums across individual heartbeats, while abnormal heart sounds display notably more extended and prominent energy spectrums. This advanced transform technique effectively captures nuanced frequency characteristics, enabling clear differentiation between healthy and pathological cardiac acoustic patterns.

Fig. 3. The integrated Residual-Recurrent Neural Networks (RRNN) model architecture. (a) The RRNN architecture employs a sophisticated multi-stage approach to heart sound analysis. The processing begins with a convolutional layer that extracts fundamental features from preprocessed heart sound data. To improve computational efficiency, max-pooling is applied, reducing the dimensionality of the feature space. The model’s architectural core consists of four residual block layers, which enable robust feature learning and help mitigate potential gradient vanishing problems. Following these residual blocks, a Long Short-Term Memory (LSTM) network is integrated to capture and analyze long-term temporal dependencies within the heart sound signals with enhanced precision. (b)The final classification stage leverages a fully connected layer that processes the LSTM output, strategically incorporating eXtreme Gradient Boosting to make the final diagnostic determination. This hybrid approach combines deep learning’s feature extraction capabilities with the predictive power of advanced machine learning techniques.

Fig. 4. Optimized Window-based Classification and Model Performance Enhancement. (a) Window-based Classification Analysis: The 10-second heart sound recordings were segmented into seven overlapping windows to optimize murmur detection. The optimal classification strategy emerged with a four-window threshold, achieving an accuracy of 84.5%. While increasing the window threshold to five improved specificity, it simultaneously compromised overall sensitivity, resulting in reduced diagnostic performance. (b) Model Performance Enhancement: By incorporating eXtreme Gradient Boosting into the RRNN architecture, the study significantly enhanced the model’s diagnostic capabilities. The integrated approach achieved a robust performance of 88.5% accuracy, with a sensitivity of 85.5% and specificity of 90.7%. This advanced methodology demonstrated the potential of combining deep learning techniques with conventional machine learning algorithms to improve cardiac sound classification.