Self-supervised learning for inflammatory response pattern recognition in physical education and exercise interventions

This study proposes PhysicalForwardPersistent, a self-supervised learning framework for recognizing inflammatory-response patterns associated with physical education and exercise interventions. The method operates on structured multivariate physiological sequences containing inflammation-related biomarkers, physiological measurements, and exercise descriptors, separating label-free representation learning from labeled downstream evaluation. PhysicalForwardPersistent integrates three components: a Physical Consistency Encoder that learns stable representations under physiologically admissible perturbations, a Forward Consistency Projector that models temporal latent-state evolution, and a Persistent Relation Extractor that preserves stable relationships among physiological profiles. The framework was evaluated on four datasets: the Physical Activity Inflammation Response Dataset, Exercise Induced Cytokine Patterns Dataset, Self-Supervised Learning Inflammation Dataset, and Physical Education Intervention Biomarker Dataset. These datasets include approximately 1,000 participants and 8,000 physiological samples, with 14–24 input variables and 2–5 temporal observations per sequence. Using participant-level separation and repeated runs, PhysicalForwardPersistent achieved accuracies of 89.23%-90.23% and macro-F1 scores of 88.34%-89.34%. Compared with the strongest baseline, it improved Accuracy by up to 1.78 percentage points and macro-F1 by up to 1.89 percentage points, with significant improvements after multiple-comparison correction. Ablation experiments confirmed that all three components contributed positively. These results demonstrate the value of jointly modeling representation stability, temporal dynamics, and relational structure for inflammatory-response pattern recognition, while further biological and prospective validation is required before practical deployment.