{"paper_id":"16e7a20d-5f3b-4653-b34b-ae2e37c09a30","body_text":"Abstract\nWe present CoPrimeEEG, a neural reconstruction framework that unifies co-prime sub-Nyquist sampling theory with a CRT-guided learning objective for EEG. Two low-rate streams obtained by co-prime decimations feed a dual-branch convolutional encoder whose fused representation is upsampled to reconstruct high-rate EEG while jointly predicting a temporal usefulness mask and canonical bandpower features. We derive a principled loss with four terms: (i) waveform fidelity, (ii) mask sparsity and smoothness, (iii) bandpower supervision in the log-domain, and (iv) a CRT-consistency term enforcing agreement between the reconstruction and its co-prime downsampled counterparts. On real EEG data, CoPrimeEEG achieves state-of-the-art reconstruction quality across MSE, MAE, correlation, SNR, and PSNR while using fewer parameters. The approach provides a practical path to low-power EEG acquisition with high-fidelity downstream analysis.\nCompeting Interest Statement\nThe authors have declared no competing interest.\nFootnotes\ndong.liu.dl2367{at}yale.edu, ywu{at}stat.ucla.edu","source_license":"CC-BY-4.0","license_restricted":false}