Brundavanam Satyasai

The absence of medical resources at remote places prevents many patients from receiving a prompt and accurate diagnosis of cardiovascular disorders. To address this, we proposed a novel deep learning model based on partially fine-tuned VGG16 and ConvMixer for the automatic identification of various heart valve diseases from phonocardiogram (PCG) signals. The method involves preprocessing PCG signals and converting them into Gammatone filterbank (GFB) based 2D time-frequency images. In order to generate time-frequency images, we utilized Gammatonegram, Gammtone Cepstral Coefficients (GTCC), and Gammatone Discrete Wavelet Coefficients (GDWC) techniques. These time-frequency images are augmented to reduce overfitting and then fed into a VGG16 model. Partial fine-tuning of the VGG16 model accelerates convergence and further improves performance. By extending the VGG16 model with ConvMixer, Global Average-Pooling, Dense, and Softmax layers, we enhance its capacity to capture intricate patterns. The ConvMixer enriches spatial and channel-wise features using Depthwise and Pointwise convolutions. We also performed an ablation analysis to highlight the effect of ConvMixer with VGG16. In addition, performance evaluation based on precision, recall, f1 score, test accuracy, and validation accuracy reveal the efficacy of the proposed method. Comparisons between Gammatonegram, GTCC and GDWC show superior performance of Gammatonegram, achieving a test accuracy of 99.60% and validation accuracy of 99.75%. Our approach demonstrates significant advances over existing methods, offering a promising solution for remote diagnosis of heart valve diseases.