Mental health conditions are often accompanied by physiological stress responses that are expressed through involuntary facial dynamics. Automated analysis of such facial motion patterns offers a promising, non-invasive pathway for continu-ous mental health assessment. In this work, we propose a latent variable model-ing framework for mental health–oriented prediction that uses pain-induced facial expressions as a monotonic proxy for stress-related affective states. The frame-work explicitly models facial motion using dense optical flow and spatiotemporal motion features, enabling the capture of fine-grained temporal variations that are difficult to infer from static appearance alone. To effectively learn long-range tem-poral dependencies in facial motion sequences, we introduce a Transformer-based temporal encoder driven by self-attention mechanisms. Given short facial video sequences, optical flow magnitude maps are extracted between consecutive frames and structured as temporal motion representations, which are then processed by the Transformer to emphasize psychologically salient motion patterns. The proposed model is evaluated on the BioVid heat pain dataset and compared against a motion-based baseline and multiple deep spatiotemporal learning architectures. Experimental results demonstrate that self-attention based temporal modeling of facial motion leads to consistent performance improvements, underscoring the relevance of explicit motion dynamics for mental health– related inference. This study highlights the potential of motion-aware, attention-driven frameworks for pain-induced mental health assessment.