Multi-Feature Fusion for Robust Heart Sound Classification in Cardiovascular Disease Diagnosis

A novel five-layer LSTM-based architecture integrated with multi head attention mechanism is proposed to classify heart sound data into five categories: normal, murmur, extrastole, artifact, and extrahls. Mel Frequency Cepstral Coefficients (MFCCs), spectrogram and wavelet features were extracted from each audio file. To address class imbalance, data augmentation techniques, like time stretching, pitch shifting, and noise addition, were applied. The attention mechanism effectively captures critical time steps, while multi-head attention further enhances long-term dependencies. The LSTM model captures the sequential patterns from the data. And finally both the outputs are concatenated to provide final output. Experimental results demonstrated a significant improvement in classification accuracy from 70–85% after augmentation, the accuracy we got as 0.85, 0.865 with different features dimension like 40 and 20. The artifact and extrahls classes achieved near-perfect F1-scores, while challenging classes like extrastole and normal showed notable improvement. The model is performed better when compared to other prescribed models.