Hybrid CNN-LSTM Framework for Automated Detection of Pediatric Heart Murmurs

This study presents a comprehensive deep learning framework for automated heart murmur detection from phonocardiogram (PCG) signals to address the subjective limitations of cardiac auscultation. Using the CirCor DigiScope dataset of 5,272 pediatric heart sound recordings from 1,568 subjects, we systematically evaluated multiple approaches including traditional machine learning models (Logistic Regression, SVM, KNN, Decision Tree, AdaBoost), basic deep learning architectures (standalone CNN and LSTM), and advanced transfer learning models (DenseNet121, EfficientNet B0/B1/B2, ResNet 50/101, MobileNet V2/V3, GhostNet, AttentionCNN). Our hybrid CNN-LSTM architecture combines mel-spectrogram spatial feature extraction with temporal sequence modeling using 128 traditional audio features and comprehensive preprocessing with data augmentation. The CNN-LSTM model achieved superior performance with 84.6% accuracy, 0.84 F1-score, 0.748 precision, and 0.800 recall, significantly outperforming traditional methods (best SVM: 0.716 F1-score) and advanced architectures (best DenseNet121: 0.72 F1-score), demonstrating robust classification with 0.879 AUC. Results indicate that hybrid deep learning architectures effectively automate cardiac auscultation, providing objective diagnostic tools with potential for improved healthcare accessibility and clinical decision-making.