<p>Understanding the psychological effects of martial arts training requires models that can bridge the gap between observable physical behavior and subjective cognitive states. This study proposes a deep learning framework that explicitly uses questionnaire-derived emotional and attentional labels as the primary supervision signal to evaluate the psychological impact of explosive martial arts actions. Rather than treating questionnaires as auxiliary or post-hoc validation tools, we integrate them directly into model training and evaluation, enabling a human-centered interpretation of motion dynamics. The proposed architecture, combining a Trajectory-Aware Spatiotemporal Perception Network (TASPN) with Dynamic Viewpoint Induced Optimization (DVIO), processes multimodal video data while being guided by psychological insights encoded in participant questionnaires. Our results on questionnaire-annotated subsets show strong predictive performance for both emotion regulation and attention assessment tasks, outperforming baselines and confirming the viability of fusing subjective human reports with spatiotemporal modeling. This integration offers a principled pathway for psychologically-informed sports analytics, real-time feedback, and individualized cognitive profiling in high-intensity physical training environments.</p>