An enhanced convolutional neural network for the classification of heart diseases using Softmax dropout and Butterworth moving average filter

Background and aim: Limitations in existing deep learning algorithms have prevented us to successfully implement deep learning in the classification of heart diseases. This is mainly due to poor signal pre-processing. This research aims to improve the classification accuracy of heart diseases and keep the processing time minimal. Methodology: The proposed system consists of ECG (Electrocardiogram) signal regularization using a Butterworth filter and moving average filter. The modified ECG is then prepared to be fed as the input in a 1-Dimensional Convolutional Neural Network (CNN). The system also includes Softmax with a dropout function to ensure the prevention of overfitting while enhancing generalization. The proposed solution is implemented using three datasets: MIT-BIH, PTB Diagnostic, and St.Petersburg. Results: With a total of 800 samples from these three datasets, around 720 samples were separated for training, and the remaining is used for the validation purpose. The solution was able to enhance the classification accuracy by 98% in terms of cross-entropy loss, compared to state of the art which is 96% to 97%. Conclusion: The proposed system achieved highly improved classification accuracy by deploying enhanced Butterworth and moving average filter for ECG signal pre-processing by reducing the noise in the signal. Furthermore, the proposed system enhances classification accuracy also by applying several filters during the signal pre-processing due to high noise in raw ECG signals. It also deploys dropout function which is added to Softmax function in the CNN to remove unnecessary nodes in the neural network, which prevents from overfitting. Removal of these unnecessary nodes provides improved performance in the classification accuracy.