You Only Acquire Sparse-channel (YOAS): A Unified Framework for Dense-channel EEG Generation

TL;DR

This work tackles the problem of generating dense-channel EEG signals from sparse channels by introducing YOAS, a four-stage framework that leverages regional electrode organization, bias signal modeling, Transformer- and diffusion-based biased EEG generation, and a deduction-based synthetic paradigm. By formulating cross-channel generation with regional references and aligning generated signals to electrode positions, YOAS achieves efficient, high-fidelity dense-channel EEG synthesis validated on the FACED dataset and across multiple emotion-recognition tasks. The study contributes a rigorous problem formulation, a unified generation framework, and experimental evidence that generated data can improve cross-subject and intra-subject emotion classification while reducing hardware requirements. This approach has practical implications for making high-density EEG more accessible in clinical, consumer, and BCI contexts through reduced sensor arrays and enhanced data quality.

Abstract

High-precision acquisition of dense-channel electroencephalogram (EEG) signals is often impeded by the costliness and lack of portability of equipment. In contrast, generating dense-channel EEG signals effectively from sparse channels shows promise and economic viability. However, sparse-channel EEG poses challenges such as reduced spatial resolution, information loss, signal mixing, and heightened susceptibility to noise and interference. To address these challenges, we first theoretically formulate the dense-channel EEG generation problem as by optimizing a set of cross-channel EEG signal generation problems. Then, we propose the YOAS framework for generating dense-channel data from sparse-channel EEG signals. The YOAS totally consists of four sequential stages: Data Preparation, Data Preprocessing, Biased-EEG Generation, and Synthetic EEG Generation. Data Preparation and Preprocessing carefully consider the distribution of EEG electrodes and low signal-to-noise ratio problem of EEG signals. Biased-EEG Generation includes sub-modules of BiasEEGGanFormer and BiasEEGDiffFormer, which facilitate long-term feature extraction with attention and generate signals by combining electrode position alignment with diffusion model, respectively. Synthetic EEG Generation synthesizes the final signals, employing a deduction paradigm for multi-channel EEG generation. Extensive experiments confirmed YOAS's feasibility, efficiency, and theoretical validity, even remarkably enhancing data discernibility. This breakthrough in dense-channel EEG signal generation from sparse-channel data opens new avenues for exploration in EEG signal processing and application.

Figures (12)

• Figure 1: The YOAS framework. The Data Preparation module extracts Bias Signal from Raw Data through regional division, followed by Data Preprocessing to eliminate outliers. The Biased-EEG Generation module captures long-term features and aligns spatial positions to generate Two-stage Bias. Finally, the Synthetic EEG Generation module synthesizes the final signals.
• Figure 2: Hypotheses illustration for dense-channel EEG generation pathways: (A) Direct one-way generation involves generating signals directly with reference to the channel; (B) Indirect one-way generation requires an intermediate channel for the reference channel to generate signals indirectly; (C) Mutual generation involves two channels serving as reference channel mutually.
• Figure 3: BiasEEGGanFormer employs an adversarial generation architecture, where the Biased-EEGGenerator generates data fed into the Biased-EEGDiscriminator for loss feedback, thereby improving the generation quality. The resulting output comprises the One-stage Generated Bias, enhancing EEG signal refinement.
• Figure 4: BiasEEGDiffFormer aligns One-stage Generated Bias with the correspondingly physical locations, first applying 1D convolution in ConvBlock for feature extraction, and subsequently involving DiffusionBlock to generate Two-stage Generated Bias based on Regional Reference Signal and diffusion model.
• Figure 5: The 32-channel EEG 2D distribution map and the initial EEG channel regional division, with Fp1, Fz, Cz, A1, CP1, CP2, Pz, T5, Oz as reference channels for their respectively regional division.