Robust Federated Learning for Intensive Care Unit Mortality Prediction Across a Heterogeneous Hospital Network
Le résumé fourni par la source
Federated learning (FL) allows clinical institutions to train a shared model without pooling patient records, but itdoes not by itself prevent leakage through exchanged updates; because this study uses neither secure aggregationnor differential privacy, this exposure is discussed as a limitation. FL robustness under corrupted or adversarial dataalso remains underexplored. This study’s AI contribution is a robustness-benchmarking methodology for federatedaggregation (FedAvg, FedProx) under label noise, input noise, and Byzantine attacks; its engineering application isfederated in-hospital mortality prediction across a real, heterogeneous multi-hospital ICU network. Using the eICUCollaborative Research Database (94 hospitals, 83,909 intensive care unit (ICU) admissions), this study developsand stress-tests the corresponding FL framework. A lightweight logistic regression model (392 parameters) is se-lected over larger alternatives on efficiency grounds, paired with focal loss for class imbalance, and evaluated undercentralized training, FedAvg, and FedProx, which perform comparably overall with a small, statistically detectableedge for FedProx under a subset of adversarial conditions. Robustness depends strongly on perturbation type: inputnoise (up to 1.0 standard deviation, SD) causes negligible degradation; three Byzantine attacks range from gradual(norm-matched random update) to severe (scaled model-replacement); and label noise causes a sharp threshold-levelcollapse beyond 20% corruption that is substantially recoverable via validation-set threshold recalibration, indicatinga calibration failure rather than a loss of ranking ability. The model requires about 17.7 MB of theoretical communi-cation payload over 60 rounds, achieves sub-millisecond inference, and approaches APACHE IVa (Acute Physiologyand Chronic Health Evaluation) clinical-severity-score performance.
Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.