Recent years have witnessed the successful development of health-related Internet-of-Things (IoT) devices, i.e., electroencephalographic (EEG), paving a ground-breaking way for better understanding the functionals in our brains. Despite massive research on EEG, there lacks an effective way to interpret complex EEG signals due to the shortage of informative EEG data, the challenge in capturing sophisticated connectivity patterns of EEG signals, and unfavorable results of the inherent noise associated with data collection. In this article, novel heterogeneous-domain deep learning is proposed to address these issues. Especially, we first propose a new scheme to extract the multilevel latent features using hybrid networks and provide two pathways for reconstructing the EEG signals. As a core part of the proposed method, the scheme considers the complex dependencies among adjacent EEG channels, spatiotemporal connectivity, and signal denoising. In addition, a novel consistency regularization method is proposed to enhance information sharing among the multilevel latent feature obtained from the labeled and unlabeled EEG samples, which is beneficial for both the information transfer and accelerating the training. Finally, we provide comprehensive case studies on the Lomonosov Moscow State University EEG data set, demonstrating that the proposed methods achieve superior performance than all the competing ones over a wide range of experimental settings.