Task-Oriented Learning for Automatic EEG Denoising

Electroencephalography (EEG) denoising methods typically depend on manual intervention or clean reference signals. This work introduces a task-oriented learning framework for automatic EEG denoising with only task labels without clean reference signals. EEG recordings are first decomposed into components based on blind source separation (BSS) techniques. A learning-based selector then assigns a retention probability to each component, and the denoised signal is reconstructed as a probability-weighted combination. A downstream proxy-task model evaluates the reconstructed signal, with its task loss supervising the selector in a collaborative optimization scheme that relies solely on task labels, eliminating the need for clean EEG references. Experiments on three datasets, covering two paradigms and multiple noise conditions, demonstrate average gains of 2.59% in task accuracy and 0.80 dB in signal-to-noise ratio under standard signal-quality metrics. Further analyses demonstrate that the task-oriented learning framework is algorithm-agnostic, as it accommodates diverse decomposition techniques and network backbones for both the selector and the proxy model. These results indicate the proposed framework can denoise EEG using only task labels, supporting its potential utility in neuroscience research and EEG-based interaction systems.