Machine Learning
FedProx Outperforms Other Federated Learning Strategies for Hospital Mortality Prediction
A new benchmark compares five federated learning approaches for predicting in-hospital deaths using fragmented clinical data while preserving privacy.
Researchers have conducted a systematic comparison of federated learning strategies designed to predict patient mortality in hospitals while keeping sensitive medical data private. The study highlights how different algorithmic approaches handle the real-world messiness of clinical information spread across multiple care units.
Why Federated Learning Matters for Healthcare#
Traditional machine learning systems for mortality prediction require centralizing patient data—a significant privacy and regulatory barrier in healthcare. Federated learning sidesteps this by allowing separate institutions to train models locally and then combine insights without sharing raw patient records. However, clinical data from different departments often varies dramatically in composition and outcomes, making unified model training difficult.
Testing Five Approaches on Real Hospital Data#
The researchers evaluated five distinct federated learning methods—FedAvg, FedProx, FedAdagrad, FedAdam, and FedCluster—using 466,351 hospital admissions from the MIMIC-IV dataset. They deliberately split cases across five care units to simulate realistic, non-uniform data distribution. Each record was enriched with laboratory results from the first 24 hours, and mortality occurred in roughly 2% of cases, creating a highly imbalanced prediction problem.
Over 50 training rounds, FedProx emerged as the strongest performer, achieving an AUC-ROC of 0.897 and area under the precision-recall curve (AUC-PR) of 0.230. Statistical testing confirmed its AUC-ROC advantage over competing methods was statistically significant. FedCluster performed best on a different metric (F1 score at 0.280 versus FedProx's 0.273), underscoring a key finding: no single strategy excels on all measures.
Centralized Models Still Have an Edge#
A traditional centralized baseline—training on all data in one location—achieved AUC-ROC and AUC-PR values of 0.929 and 0.312 respectively, notably outpacing FedProx on both fronts. This suggests that privacy-preserving federated approaches come with a modest performance trade-off, though the gap is not dramatic.