Kolmogorov-Arnold-Based Network With Lightweight Feature Fusion Schema for Single-Lead Electrocardiogram Atrial Fibrillation Detection

Atrial fibrillation (AF) is a prevalent cardiac arrhythmia that poses a serious threat to patients' cardiovascular health. Deep learning-based single-lead wearable ECG devices have been widely studied and shown satisfactory performance in early AF detection. However, most existing models increase complexity by employing deep networks and feature fusion techniques to enhance robustness and generalization. Despite several model compression techniques also have been applied to reduce model parameters by compromising certain aspects of performance. In this study, to strike a balance between model performance and complexity, we proposed a Piecewise Aggregate Approximation(PAA)-based lightweight feature fusion schema in the proposed KAN-based Network, with an improved connected layer from Kolmogorov-Arnold Network(KAN) to replace the fully connected layer. Our innovation leveraged the learnable nonlinear activation functions of the KAN Layer to enhance the complex high-dimensional representations at our network's output. Additionally, to reduce the number of parameters generated by feature fusion and the KAN Layer, we applied large-scale dimensionality reduction to the ECG signals using PAA as the secondary feature, allowing the lowdimensional features to represent part of the high-dimensional features through sequence-level fusion. Testing on the PhysioNet2017 dataset demonstrates that our method outperformed the benchmark, achieving a <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{2. 2 \%}$</tex> improvement in the F1 score for AF detection while reducing parameters by <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$13 \%$</tex> compared to traditional feature fusion.