Virtual EEG channel inference from four frontal electrodes for wearable EEG augmentation

To develop and evaluate a compact method for inferring virtual EEG channels from four frontal electrodes, with attention to both waveform fidelity and spectral-topographic usability.

FAVC Net estimates 13 target channels from Fp1, Fp2, F7, and F8 using shared multi-scale encoding, source-state embeddings, target-conditioned signed source-block mixing, GATv2 refinement, attention-consistent skip fusion, and a waveform-spectral reconstruction objective. Training was task agnostic. Evaluation on PRED+CT used five independent subject-level splits and included neural and interpolation baselines and checkpoint-only source-perturbation tests.

The complete FAVC-Net configuration achieved waveform fidelity comparable to the strongest waveform baseline, alongside lower spectral errors. Relative to the strongest non-FAVC comparator for each metric, log-spectral distance and PSD KL divergence were reduced by 30.50% and 38.94%, respectively. Under wearable-like perturbations, the most consistent advantages concerned spectral fidelity and channel-frequency texture, with condition-specific benefits in spectral-collapse measures.

FAVC-Net provides montage-compatible deterministic conditional representations for low-channel EEG analysis. These representations transfer learned multichannel structure without adding independent measurement information beyond the recorded sources.