Explainable Deep Contrastive Federal Learnig System for Early Prediction of Clinical Status in Intensive Care Unit

Abstract

Early prediction of a patient's clinical condition in the Intensive Care Unit (ICU) is essential for improving treatment outcomes, reducing mortality, and supporting timely medical interventions. However, traditional centralized machine learning approaches face significant challenges related to patient data privacy, limited data sharing between hospitals, and lack of model transparency. To address these issues, this project proposes an Explainable Deep Contrastive Federated Learning System for Early Prediction of Clinical Status in Intensive Care Units. The proposed framework combines federated learning with deep contrastive learning to train a robust predictive model collaboratively across multiple healthcare institutions without transferring sensitive patient data. Contrastive learning enables the model to learn meaningful and discriminative feature representations from heterogeneous clinical records, improving prediction accuracy even when data distributions differ across hospitals. Additionally, Explainable Artificial Intelligence (XAI) techniques such as SHAP or LIME are integrated to provide transparent explanations for each prediction, helping clinicians understand the key factors influencing patient outcomes. The system processes electronic health records, vital signs, laboratory test results, and demographic information to predict clinical deterioration at an early stage. By preserving patient privacy, enhancing model generalization, and providing interpretable predictions, the proposed framework supports reliable clinical decision-making while complying with healthcare data protection requirements. Experimental evaluation demonstrates improved predictive performance, enhanced privacy preservation, and greater trustworthiness compared to conventional centralized and non-explainable machine learning approaches, making the proposed system suitable for real-world ICU monitoring and intelligent healthcare applications