Monitoring psychological health through physiological cues is crucial for non-contact human-device interaction frameworks. This study utilized a proprietary dataset encompassing synchronized multi-modal data from 45 student participants. We introduced a tailored convolutional graph neural network (ConvGNN) to assess an individual’s psychological well-being, leveraging physiological indicators like respiration wave patterns (RWP) and heart wave patterns (HWP). The intrinsic relationship between RWP and HWP signatures was dynamically learned via an adjacency matrix during neural network training. Enhanced discriminative feature extraction facilitated the classifying of six psychological states encoded in the softmax layer. Our ConvGNN model outperformed existing methods, as demonstrated on widely-used psychological datasets such as ASCERTAIN, DREAMER, and our in-house dataset. Psychological health classification accuracy was validated using ground truth data from subjective evaluations (SAM score) and objective evaluations (RGB facial expression).