A hybrid deep learning framework for privacy-preserving sepsis prediction in distributed ICU environments using federated learning simulation

Abstract

Sepsis remains one of the leading causes of intensive care unit (ICU) mortality in the United States, and timely clinical decision support is essential for improving patient outcomes. Although Electronic Health Record (EHR) systems provide access to large volumes of clinical data, strict patient privacy regulations and institutional data governance policies limit data sharing across hospitals, hindering the development of robust, generalizable predictive models. To address this challenge, this study proposes a privacy-preserving federated learning framework for sepsis prediction using a hybrid deep learning architecture that integrates one-dimensional Convolutional Neural Networks (1D-CNN), Bidirectional Long Short-Term Memory (BiLSTM) networks, and Transformer attention blocks. A MIMIC-IV-style ICU dataset was partitioned across three simulated hospital clients, and Federated Averaging (FedAvg) was employed to coordinate global model optimization. An explicit adversary/threat model (e.g., honest-but-curious server, colluding clients) was not formally defined; whether a malicious server or client could infer patient-level information was not assessed. The proposed framework was evaluated using accuracy, AUROC, precision, recall, F1-score, and specificity metrics, with SHapley Additive exPlanations (SHAP) utilized to ensure clinical interpretability. Experimental results demonstrated that the federated hybrid model achieved an accuracy of 93.6%, an AUROC of 0.959, a precision of 0.871, an F1-score of 0.759, and a specificity of 98.24%. Although a centralized baseline model achieved an AUROC of 0.997, the federated approach delivered competitive predictive performance while maintaining strict data privacy. Furthermore, the total communication cost across all federated training rounds was only 36.09 MB, highlighting its practical efficiency for distributed healthcare environments. SHAP analysis identified Sequential Organ Failure Assessment (SOFA) score, lactate level, white blood cell (WBC) count, creatinine, and procalcitonin as the most influential predictors of sepsis risk. These findings demonstrate feasibility on synthetic MIMIC-IV-style data; validation on real multi-institutional ICU records is required before scalable, secure deployment across distributed healthcare systems can be claimed.