SMARTFEDICU: PERSONALIZED FEDERATED DEEP LEARNING FOR IN-HOSPITAL MORTALITY RISK FORECASTING

Abstract

In-hospital mortality prediction is a critical task in intensive care units (ICUs), where timely risk assessment can guide clinical decisions and improve patient outcomes. However, sharing patient data across multiple hospitals for model training raises significant privacy and regulatory challenges. To address this issue, this paper proposes SmartFedICU, a personalized federated deep learning framework designed to predict inhospital mortality across multi-center ICU datasets without compromising patient confidentiality. The proposed system enables distributed model training across participating hospitals, ensuring data privacy by keeping sensitive patient records local while aggregating model parameters globally. SmartFedICU incorporates a two-phase learning strategy: global knowledge extraction through federated aggregation and local personalization using adaptive fine-tuning. Deep neural architectures such as Long Short-Term Memory (LSTM) networks and attention mechanisms are employed to capture complex temporal and physiological relationships from patient monitoring data. The personalization component adjusts the global model to reflect local patient population characteristics, enhancing predictive accuracy for heterogeneous healthcare centers. Experimental evaluation on multi-center ICU datasets demonstrates that SmartFedICU outperforms conventional centralized and non-personalized federated models in terms of accuracy, sensitivity, and F1- score. Moreover, the proposed system ensures robust generalization while preserving data privacy and regulatory compliance. The results confirm that personalized federated learning is a viable approach to advancing precision medicine and real-time mortality risk forecasting in critical care environments.