Early clinical intervention and medical resources allocation require precise predictions on patient risks. The practical value of traditional risk assessment methods in clinical decision-making is restricted because solutions commonly rest on single physiological information and are complicated to interpret. The model combines state-of-the-art ensemble learning and explainable AI techniques with clinically interpretable composite indices to propose an explainable multi-system composite scoring model for patient risk assessment. To capture complex physiological couplings between multiple organ systems, three new combined indices have been developed: the Physiological Risk Index (PRI), Multi-System Early Warning Score (MS-EWS) and Critical Threshold Composite (CTC). The system employs the Synthetic Minority Over-sampling Technique (SMOTE) for class imbalance processing and includes interactionbased feature engineering. To increase prediction robustness, a stacked ensemble model using Random Forest, XGBoost and LightGBM classifiers trained with 5-fold cross-validation for meta-learner optimization is employed. Performance evaluation shows that the proposed model outperforms individual base learners, achieving 98% accuracy on an independent test set, as well as existing state-of-the-art models. The integration of SHAP and LIME ensures both global and instance-level explainability. Moreover, misclassifications mainly occur between closely related categories, thereby maintaining clinical safety and highlighting key composite features that influence model decisions. The findings suggest that the proposed model presents an effective, interpretable and clinically reliable multi-class health risk prediction approach. In the future, this study plans to focus on the validation of multi-center dataset as well as real-time clinical deployment and more improvement in explainability feature engineering.