CNN-Driven Detection of Abnormalities in PCG Signals Using Gammatonegram Analysis

Phonocardiogram (PCG) signals carry essential data about heart health, offering significant potential for early disease detection. This study presents a method to identify heart abnormalities using Gammatonegram images derived from PCG signals, which are analyzed with a Convolutional Neural Network (CNN). The Gammatonegram, a time-frequency representation mimicking human auditory processing, captures both temporal and spectral features of PCG signals. The images are pre-processed through resizing and data augmentation, the images are input into the CNN model for classification. The model distinguishes between healthy and unhealthy heart sounds, and its performance is evaluated using several optimizers—SGD, Adagrad, Adadelta, RMSprop, and Adam. Adam yielded the best performance, achieving a test accuracy of 100%. This method provides a robust, non-invasive solution for heart disease diagnosis, leveraging deep learning to improve accuracy in detecting heart abnormalities from PCG signals.