Dynamic deep residual shrinkage networks for adaptive removal of ocular artifacts in EEG signals

Electroencephalogram (EEG) signals, which are crucial in clinical applications such as diagnostics, drug development, and neuroscience research, often suffer from noise, particularly ocular artifacts (OAs). Traditional denoising methods like filtering and regression analysis have limitations such as loss of signal information and high computational costs. Recent deep learning models offer improvements but face challenges in handling different signal-to-noise ratios (SNR). EEG SNR often fluctuates significantly over time in practice, causing models to perform well under specific conditions but poorly under others. We propose a novel multi-network dynamic denoising framework for removing OAs that includes a baseline denoising model for low SNR signals and a fine-tuning model for high SNR conditions. This framework dynamically adjusts based on the SNR, ensuring robust denoising for low SNR signals while preserving details in high SNR scenarios. An SNR classification network guides the process with prior information. Our method significantly outperforms traditional approaches, presenting an effective solution for enhancing the quality of EEG signals in clinical applications.