Nonnegative tensor decomposition imposes nonnegative constraints on its latent factors, providing a part-based tensor representation that can extract meaningful and convincing information. This approach has been used widely across applications like signal processing, neuroscience, and other areas. For multi-block tensor group analysis, including multiple-subject or multiple-modal medical data, traditional single tensor decomposition fails to maintain feature comparability or explore the coupled information across tensors. This study introduces a novel coupled CANDECOMP/PARAFAC tensor decomposition method using the non-negativity constraints and the alternating proximal gradient strategy, termed CoNCPD-APG. The proposed algorithm enables the group analysis of two or more tensors that are fully- or partially-coupled, allowing for the simultaneous acquisition of shared, individual information, and core tensors. Experiment results of synthetic and real event-related potential data confirm the effectiveness of the proposed coupled tensor decomposition algorithm in discovering meaningful latent patterns and relationships from/among complex multi-block tensors.