LRR-Unet provides an effective and interpretable deep learning solution for EEG signal denoising, improving both artifact removal and downstream classification performance.
BACKGROUND: Electroencephalogram (EEG) signals are crucial for brain-computer interface research but are highly susceptible to noise contamination, necessitating effective denoising. While deep learning has been widely applied, its "black-box" nature limits interpretability. In contrast, traditional model-based methods like Low-Rank Recovery (LRR) offer strong interpretability by decomposing signals into low-rank and sparse components. OBJECTIVE: This paper aims to develop an interpretable deep-learning model for EEG denoising that combines the performance of deep learning with the interpretability of traditional LRR methods. METHODS: We propose LRR-Unet, a deep unfolding network that transforms the traditional iterative LRR algorithm into a neural network architecture. Specifically, the time-consuming Singular Value Decomposition (SVD) and sparse optimization processes in LRR are replaced with learnable neural network modules. RESULTS: Extensive experiments demonstrate that LRR-Unet outperforms other state-of-the-art models in removing ocular and electromyographic artifacts, achieving superior performance on both quantitative and qualitative metrics. Furthermore, in downstream classification tasks, EEG signals preprocessed with LRR-Unet yield better results across various evaluation indicators. CONCLUSION: The proposed LRR-Unet provides an effective and interpretable solution for EEG denoising. Its superiority in denoising performance and practical utility in enhancing downstream application performance is validated through comprehensive experiments.
Yue et al. (Wed,) studied this question.