Atrial Fibrillation Identification via Joint SCSB and CCB Neural Network With the 2DSSA and CNN

To identify atrial fibrillation (AF) using deep learning, electrocardiogram (ECG) signals are often represented as spectrograms to capture their time frequency characteristics; however, these representations often introduce irrelevant or redundant information that adversely affects classification performance. This paper proposes a joint two dimensional singular spectrum analysis (2DSSA) component selection block (SCSB) and convolutional neural network based classification block (CCB) to achieve effective noise suppression, redundancy reduction, and feature extraction. Conventional 2DSSA component selection methods rely on empirical and binary selection schemes, which limit optimality and flexibility. The proposed SCSB overcomes these drawbacks by employing a weighted sum approach to fuse the 2DSSA components, in which the component weights are adaptively learned via stochastic gradient descent. Integrated with the CCB for feature extraction and AF identification, the proposed framework demonstrates significant improvements in accuracy and F1 score through numerical simulations, validating its effectiveness.