Figure 3
An in-depth analysis is performed on a CNN that is sequentially trained on each lead of a 12-lead clinical ECG dataset, one lead at a time, to predict single-lead ECG on real data. The proposed memory-optimized CNN architecture consists of 32, 64, 128, and 128 filters, various feature map size dimensions of each layer i.e., 48\(\times\)7, and ‘ReLU’ activation function. The model is trained in conjunction with preprocessing techniques (given in the algorithms) to ensure compatibility and effectiveness with the 12-lead format.
