Document Type

Conference Paper

Publication Date

2026

DOI

10.1145/3807503.3819489

Publication Title

Proceedings of the 17th ACM International Conference on Bioinformatics, Computational Biology and Health Informatics

Pages

10 pp.

Conference Name

17th ACM International Conference on Bioinformatics, Computational Biology and Health Informatics, 30 June-3 July 2026, Rende (CS), Italy

Abstract

Accurate prediction of ICU Length of Stay (LoS) is essential for clinical decision-making and healthcare resource management. Graph Neural Networks (GNNs), such as GraphSAGE, offer a natural fit by capturing patient data from Electronic Health Records (EHRs) through graph structures. However, the distributed and sensitive nature of this data raises both privacy and legal concerns regarding the aggregation and training of GNN models. This additionally leads to issues with data imbalance and model robustness. In this study, we perform an analysis of the Federated Graph Neural Network (GNN-FL) framework to enable decentralized learning on EHRs derived from the MIMIC-III dataset. We benchmark several federated aggregation strategies, including FedAvg, FedProx, FedCurv, FedAvgM, FedNova, Clustered FL, and attention-based aggregation, to assess their effectiveness in modeling ICU LoS. Additionally, we perform an extensive robustness analysis against real-world data poisoning threats, including label flipping, feature vanishing, and randomized labeling applied at different corruption levels. We further integrate Local Differential Privacy (LDP) into our framework and evaluate its impact across a range of privacy budgets. This work highlights the viability of secure, privacy-preserving, and robust GNN-based ICU LoS prediction in Federated Learning (FL) environments. From the analysis, we found that the GraphSAGE outperforms the other GNNs by approximately 8% in terms of AUC-ROC in both centralized and FL settings. Among the aggregation strategies, FedAvgM achieved the best AUC-ROC score for both clean and poisoned datasets. The incorporation of LDP further improves robustness, reducing the average AUC degradation by approximately 2% for label flip and 7% for feature vanish at the highest poisoning level.

Rights

© 2026 Copyright held by the owner/authors.

This work is licensed under a Creative Commons Attribution 4.0 International (CC BY 4.0) License.

Original Publication Citation

Dipto, S. M., Banerjee, S., Roy, S., Al Musawi, A. F., Ghosh, P., Shetty, S., & Rana, P. (2026). An investigation of federated GNNs under aggregation, data poisoning, and differential privacy for ICU length-of-stay prediction. In, Proceedings of the 17th ACM International Conference on Bioinformatics, Computational Biology and Health Informatics (pp. 1-10). Association for Computing Machinery. https://doi.org/10.1145/3807503.3819489

ORCID

0000-0003-2704-118X (Dipto), 0000-0001-6455-5134 (Banerjee), 0000-0003-4743-0342 (Roy), 0000-0002-8789-0610 (Shetty), 0000-0001-9199-2479 (Rana)

Download

Share

COinS