An ensemble-based transfer learning model for predicting the imbalance heart sound signal using spectrogram images

Heart sound signal analysis is an important area in healthcare, and the detection of imbalanced heart sounds can provide valuable diagnostic information. However, due to heart sound variability, accurate prediction of imbalanced signals remains challenging. The issue of class inequality has received much attention from numerous scientific domains. The correct classification becomes increasingly challenging as data scale and data imbalance increase. Traditional classifiers tend to favor the dominant class and overlook the minority class, which is frequently considerably more significant when dealing with imbalanced learning problems. We propose an ensemble learning algorithm based on a transfer learning convolutional neural network (CNN) model to solve these challenges to predict imbalanced heart sound signals. We employ spectrogram images and STFT to extract the relevant features from Phonocardiogram (PCG) data. Our model leverages the pre-trained CNN architecture and fine-tunes it on the spectrogram images to improve the prediction performance. Moreover, we incorporate an ensemble approach to improve the model’s robustness and accuracy. Our experimental results on a publicly available PhysioNet PCG dataset demonstrate that the proposed algorithm outperforms existing state-of-the-art methods in terms of accuracy, sensitivity, and specificity. The ensemble methodology comprising AlexNet, SqueezeNet, and VGG19 models was proposed and achieved the highest level of performance, resulting in an accuracy of 99.20% and a sensitivity rate of 99.47%. Our study showcases the potential of leveraging technological advancements to predict unbalanced Phonocardiogram (PCG) signals using spectrogram images. This research opens up promising avenues for future exploration in cardiac diagnostics. Specifically, the ensemble-based transfer learning model proposed in this study holds great promise.