Missing but Not Lost: Reconstructing Arbitrarily Missing EEG Channels with Pre-trained Diffusion

Yan-Kai Liu, Hao-Long Yin, Bao-liang Lu, Wei-Long Zheng

Keywords: EEG，diffusion model

Abstract: Electroencephalography (EEG) plays a pivotal role in brain–computer interface (BCI) research owing to its non-invasive nature and high temporal resolution. However, the presence of missing channels during acquisition often compromises signal quality and limits practical applications. Existing EEG channel reconstruction methods remain constrained by limited accuracy and generalization. In this work, we present Uni-SVDiffusion, a scalable pre-training–based diffusion framework for cross-channel EEG reconstruction that generates high-fidelity signals for arbitrarily missing channels. Our approach leverages singular value decomposition (SVD) to disentangle EEG signals into spatial and temporal components, and employs a diffusion model trained on spatial representations to achieve precise reconstruction conditioned on observed channels. To enable a unified model across diverse EEG configurations, we further propose a forward/backward channel mapping strategy that preserves spatial structure and facilitates cross-dataset generalization via convolutional pre-training. Evaluations on three datasets demonstrate that Uni-SVDiffusion achieves state-of-the-art performance, even under severe channel-missing scenarios. This work provides a generalizable, plug-and-play solution for EEG reconstruction. Code and pre-trained models will be released upon acceptance.

Primary Area: applications to neuroscience & cognitive science

Submission Number: 9004