Fig. 5

Performance evaluation of deep learning models. A. Classification accuracies achieved by individual models in evaluating mitral regurgitation severity. This graph shows the accuracy of three audio processing models—CNN6, PaSST, and ResNet38—using two types of features to analyze audio data: Fbank and mel spectrograms. Each set of bars illustrates the accuracy results for each model, with lighter shades representing Fbank and darker shades representing mel spectrograms. Standard deviations are indicated by error bars. B. Confusion matrix for CNN-based mitral regurgitation classification. The confusion matrix depicts the performance of our CNN in classifying the severity of mitral regurgitation. The true severity levels are plotted on the y-axis, and the predicted severity levels are plotted on the x-axis. Each cell contains the percentage of instances for each predicted true-label pair. CNN, convolutional neural network; Fbank, filter bank; PaSST, patch-mix audio spectrogram transformer; ResNet, residual neural network