BCIwiki (bciwiki.com) — Researchers at Tianjin University and Tiangong University proposed MTGNet, a task-oriented and spectrally guided EEG denoising framework that cut spectral relative root-mean-square error by 18.9%, 31.5% and 14.0% for EMG, EOG and hybrid artifacts on the public EEGDenoiseNet dataset, and raised classification accuracy by 6.20 to 6.69 percentage points over unprocessed input on a real-world fatigue EEG dataset. The results were reported in Journal of Neural Engineering on August 26, 2026.
EEG sits at the microvolt level, so electromyography, electrooculography and mixed physiological artifacts readily contaminate it, while aggressive filtering can strip information that downstream tasks need. MTGNet binds denoising to a specific task: a pretrained 11.97 million-parameter backbone learns to preserve intrinsic EEG characteristics from paired noisy-clean data, while 0.33 million LoRA parameters handle task-specific adaptation, removing the need for paired clean EEG references.
The team reported the artifact suppression results on EEGDenoiseNet (p<0.001) and the accuracy gain on a real-world fatigue EEG dataset (p<0.05). Ablation, cross-classifier and cross-dataset analyses were used to validate the components and support the transferability of MTGNet across the evaluated settings. The authors note that its applicability across diverse tasks, artifact types and acquisition conditions still requires further validation.