Diffusion-based reconstruction of high-density EEG signals in an EEG with a limited number of electrodes for portable brain-computer interfaces

Abstract

Wearable BCI systems need compact EEG systems for applications in daily life settings. However, reducing the number of electrodes largely sacrifices spatial resolution. We introduced a new method using denoising diffusion probabilistic models (DDPM) to synthesize high-density EEG recordings from sparse electrode setups. We assumed that DDPM might capture the spatial sparsity of scalp EEG and precisely recover direction information of missing channels by conditional generation. Five healthy participants participated in visual event-related potential (ERP) tasks and were recorded with 64-channel EEG. We have developed the inpainting task on top of a DDPM with conditionally sampled inputs. For the first experiments, either 30%, 50% or 70% of channels were masked randomly and models were trained on recovering the full scalp topography. For second spatial reconstruction experiments, we also select a reduced number of electrodes(6 channels)in either the temporal or occipital regions. The DDPM was trained with 1000 steps. Reconstruction performance was quantified by MSE between original and reconstructed signals in every 500ms window of the length 500ms. Analysis of the subject-level MSE results demonstrated that the MSE performed consistently with very low between-subject variability. Random channel selection at 50% also revealed very strong ERP component reconstruction, if you consider the respective characteristic components, such as P100, N170 and P300. It was shown that 6-channel configurations of either occipital or temporal regions could decode whole-brain activity patterns with a fair level of accuracy while preserving the spatiotemporal dynamics of visual processing. These results suggest that spatial correlations embedded in EEG signals can be exploited efficiently by the DDPM to recover high-resolution estimations with a limited number of recording electrodes. This methodology may open the doors for the lightweight and portable BCI systems while reducing the hardware complexity.