Individual EEG Super-Resolution Via ADMM-based Coupled Matrix Decomposition Towards Long-Term Brain Monitoring

Super-resolution (SR) reconstruction of electroencephalography (EEG) is mandatory in neuro-science and engineering applications demanding fine-grained spatial information while high-density EEG devices do not suit. The successes of EEG SR routinely rely on the excessive high-resolution (HR) ground truth in order to properly handle the spatio-temporal characteristics of EEG, which cannot be guaranteed in the scenarios of long-term brain monitoring due to the insufficient HR ground truth and the subject's individuality. Aiming at this pitfall, this study proposes an ADMM-CMD approach (Alternating Direction Method of Multipliers- based Coupled Matrix Decomposition), which simultaneously operates on the initial HR ground truth and the low-resolution (LR) EEG requiring SR with the individual spatio-temporal relations preserved. First, the CMD model is constructed to transform the initial HR ground truth and the target LR EEG to the latent source space with common mapping pattern, where functional connectivity measure applies to highlight the EEG's spatial characteristics. Second, the ADMM algorithm iteratively solves the CMD model to derive the mapping matrix and more importantly the latent sources embedding the temporal characteristics, thus to support the process of EEG SR. The experimental results on EEG datasets of Autism Spectrum Disorder (ASD) & Typically Development (TD) and Motor Imagery (MI) indicate that: (1) ADMM-CMD performs effectively in EEG SR reconstruction with the decrease in normalized mean squared error by 0.1%similar to 7.5%, the increase in signal-tonoise ratio by up to 2.1 dB, and the improvement in Pearson's correlation coefficient by 4.3%, and (2) the reconstructed SR EEG by ADMM-CMD demonstrates superiority to the LR alternatives in ASD classification, especially when limited HR ground truth is available.