Title: Enhancing Wearable EEG via Robust Frequency-Calibrated Virtual Channel Synthesis Using Four Frontal Electrodes

Abstract:

While low-channel wearable electroencephalography (EEG) offers significant appeal for continuous, long-term monitoring, relying on just four frontal electrodes yields a limited and spatially skewed perspective of widespread scalp activity. To address this, we introduce FAVC-Net, a compact network designed to generate 13 unmeasured EEG channels by utilizing data from Fp1, Fp2, F7, and F8. This model integrates several advanced components: shared multi-scale source encoding, source-state embeddings, target-conditioned signed source-block mixing, attention refinement via GATv2, skip fusion that maintains attention consistency, and weak Welch power spectral density calibration.

Instead of viewing the conversion from sparse to dense EEG as a simple waveform-matching exercise, our framework simultaneously prioritizes amplitude accuracy, spectral distribution, channel-frequency texture, and resilience against noisy wearable inputs. Evaluations on the PRED+CT dataset demonstrate that FAVC-Net achieves the optimal balance between waveform and spectral performance, outperforming both neural networks and interpolation baselines. While improvements in the time domain were slight, the model significantly lowered log-spectral distance by 30.09% and PSD KL divergence by 37.98% compared to the best non-FAVC alternative. Furthermore, under wearable-like source perturbations, the system maintained spectral integrity and avoided spectral collapse. These findings validate virtual EEG channel generation as a dual-domain augmentation challenge, highlighting that generated posterior and parietal channels should be understood as frequency-calibrated representations based on sparse frontal data, rather than as direct physical recordings.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC