BlendNet: a blending-based convolutional neural network for effective deep learning of electrocardiogram signals.

Introduction: In recent years, Deep Learning (DL) architectures such as Convolutional Neural Network (CNN) and its variants have been shown to be effective in the diagnosis of cardiovascular disease from ElectroCardioGram (ECG) signals. In the case of ECG as a one-dimensional signal, 1-D CNNs are deployed, whereas in the case of a 2D-represented ECG signal, i.e., two-dimensional signal, 2-D CNNs or other relevant architectures are deployed. Since 2D-represented ECG signals facilitate better feature extraction, it is a common practice to convert an ECG signal into a scalogram image using a continuous wavelet transform (CWT) approach and then subject it to a DL architecture such as 2-D CNN. However, this traditional approach captures only a limited set of features of ECG and thereby limits the effectiveness of DL architectures in disease detection. Methods: This work proposes "BlendNet," a DL architecture that effectively extracts the features of an ECG signal using a blending approach termed "alpha blending." First, the 1-D ECG signal is converted into a scalogram image using CWT, and a binary version of the scalogram image is also obtained. Then, both the scalogram and binary images are subjected to a sequence of convolution and pooling layers, and the resulting feature images are blended. This blended feature image is subjected to a dense layer that classifies the image. The blending is flexible, and it is controlled by a parameter α, hence the process is termed as alpha blending. The utilization of alpha blending facilitates the generation of a composite feature set that incorporates different characteristics from both the scalogram and binary versions. Results: For experiments, a total of 162 ECG recordings from the PhysioNet database were used. Experimental results and analysis show that, in the case of α = 0.7, BlendNet's performance surpasses the performance of (i) traditional approaches (that do not involve blending) and (ii) state-of-the-art approaches for ECG classification. Discussion: Experimental outcomes show that the proposed BlendNet is flexible regarding dense layer settings and can accommodate faster alternatives [i.e., machine learning (ML) algorithms] for faster convergence. The superior performance at α = 0.7 indicates that alpha blending allows for richer composite feature sets, leading to improved classification accuracy over conventional feature extraction and classification methods.