Federated Knowledge Distillation for In-Hospital Mortality Prediction in ICU Patients with Acute Kidney Injury
Year: 2026
Type: conference-paper
Abstract: Acute Kidney Injury (AKI) is a critical ICU complication, with in-hospital mortality reaching 18.7% among AKI-positive patients in this cohort. Data-privacy restrictions across hospitals prevent centralized model training, motivating privacypreserving alternatives. This paper proposes a Federated Knowledge Distillation (FedKD) framework for in-hospital mortality prediction in ICU patients with AKI. Using the PhysioNet Challenge 2012 dataset, a KDIGO-based labeling pipeline identifies 6,075 AKI-positive patients, of whom 4,021 are distributed across four federated clients and 2,054 form an external hospital cohort (Set-C) reserved for zero-shot validation. A teacher network (90,241 parameters) trained via FedAvg is distilled into a compact student (14,337 parameters,$\mathbf{6. 3} \times$smaller). Averaged over 20 seeds, the FedKD student achieves a mean test AUROC of$0.8302 \pm 0.0177$, which is statistically comparable to the FedAvg teacher$(0.8344 \pm 0.0161; p=0.0199)$, while delivering markedly higher sensitivity on the held-out test set$(\mathbf{0. 7 4 8 3}$vs.$\mathbf{0. 5 3 6 4})$and generalizing to the unseen hospital cohort with zero-shot AUROC 0.8182. These results show that knowledge distillation substantially improves mortality recall, the clinically decisive metric, while enabling privacy-preserving, communication-efficient federated inference without a corresponding loss in discriminative performance.
Institution: Chittagong Independent University
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7213537014