EEG-SDL-Net: biomedical EEG signal classification via multi-level supervised Kolmogorov-Arnold networks

This paper presents a novel supervised deep learning framework for EEG signal classification based on hierarchical one-dimensional convolutional neural networks with Sobel-based temporal enhancement, channel attention, Kolmogorov-Arnold Networks (KAN), and multi-level deep supervision. The proposed framework is designed to improve feature representation, robustness to EEG variability, and computational efficiency for Edge AI deployment. The proposed architecture addresses the challenges of EEG signal variability, nonlinear temporal dynamics, and robust feature extraction. The framework integrates a 1D Sobel projection module for local temporal feature enhancement, a channel-wise attention mechanism for adaptive electrode selection, and KAN to capture complex non-linear relationships. Furthermore, a deep supervision strategy is employed to stabilize training and ensure robust gradient flow. Experimental evaluations on the SEED-VIG dataset demonstrate that the proposed model achieves superior classification accuracy compared to state-of-the-art methods. Performance benchmarks on edge devices, such as the Raspberry Pi 4, show a real-time inference latency of 15 ms, confirming the framework's suitability for Internet of Medical Things (IoMT) and portable monitoring systems. The complete source code has been made openly available and permanently archived through Zenodo: https://doi.org/10.5281/zenodo.21441822.