Abstract: This study focuses on the core contradiction in clinical prediction and resource allocation for critically ill patients: data from a single hospital is limited, making it difficult to train reliable models with strong generalization ability; meanwhile, directly sharing raw medical data faces strict privacy and compliance restrictions. To address this issue, we first constructed a federated learning-based dual-task prediction system to achieve multi-center collaborative modeling while protecting privacy. In the mortality risk prediction task, we compared various data augmentation methods and machine learning models to select the optimal combination suitable for imbalanced medical data classification tasks. In the hospital length-of-stay prediction task, we leveraged the federated learning framework to aggregate distributed medical knowledge from multiple hospitals, building a global prediction model, and validated it against local models at each hospital and centralized models using multidimensional metrics. Experimental results show that the new method effectively balances data privacy protection with multi-center collaborative modeling, achieving predictive accuracy close to centralized training. For mortality risk prediction, the ADASYN-MLP combination demonstrated significant advantages, with the federated learning global model achieving higher F1 score (0.30) and recall rate (0.76) than single-hospital local models, effectively addressing the sample imbalance problem caused by the 'rare event' nature of critical illness. For hospital length-of-stay prediction, the federated learning global model significantly outperformed local hospital models in mean absolute error (MAE = 7.13) and coefficient of determination (R² = 0.69), demonstrating superior predictive accuracy and generalization ability.