Classification of Normal and Abnormal S3 Heart Sounds Using Spectral Features and LSTM Neural Networks

This study presents an exploratory approach to classifying heart sounds, with a particular focus on the S3 sound, using a digital stethoscope combined with an LSTM neural network. The methodology involves the acquisition of heart signals, their processing using a fourth-order Butterworth filter, and the extraction of Mel-Frequency Cepstral Coefficients (MFCC), which capture the relevant spectral characteristics of the sound. These coefficients are used as input to the LSTM model to classify the sounds into normal and abnormal categories. The experiments were conducted with a limited sample of 10 participants, evenly divided between healthy individuals and those with previously diagnosed cardiac pathologies. Preliminary results demonstrated a classification accuracy of 95% during training and 90% in validation, suggesting the potential of the proposed approach for applications in telemedicine and non-invasive diagnostics. However, further studies with larger and more diverse datasets are needed to confirm the clinical applicability of this methodology.