In recent years, proactively predicting sports injuries using multimodal wearable sensor data has emerged as a key research focus at the intersection of sports medicine and artificial intelligence. To address the limitations of existing approaches in modeling cross-modal interactions, generalization, and mechanism-level interpretability, this study proposes an integrated framework for proactive sports injury risk prediction that combines self-supervised modeling, dynamic modality fusion, causal inference, and interpretability analysis. Specifically, the method designs structure alignment and feature enhancement modules and incorporates a Transformer-based cross-modal attention mechanism to accurately capture the dynamic relationships among multimodal physiological signals such as heart rate, electromyography (EMG) , and inertial measurement units (IMU) . A self-supervised training strategy based on masked reconstruction and contrastive learning is employed to improve robustness and generalization in scenarios with limited samples and heterogeneous temporal data. Furthermore, by integrating SHAP and LIME algorithms, the method achieves feature contribution visualization, while a structural causal model systematically reveals the causal relationships among training load, recovery rhythm, and injury risk. Experiments on several publicly available sports sensor datasets demonstrate that the proposed method outperforms mainstream approaches in terms of injury warning accuracy, generalization, and interpretability, and thus holds greater potential for real-world applications. In summary, this work provides an efficient, interpretable, and mechanism-driven technical pathway for risk perception and scientific intervention in sports health, advancing data-driven methodologies in the field of sports medicine.