Federated Learning Benchmarked for Mortality Prediction on Clinical Data

A recent study evaluates five strategies for federated learning (FL)—FedAvg, FedProx, FedAdagrad, FedAdam, and FedCluster—focused on predicting in-hospital mortality with the MIMIC-IV dataset. The analysis involves 466,351 admissions from five care units, establishing a realistic non-IID (non-independent and identically distributed) environment, complemented by an 11-item laboratory panel from the first 24 hours. With a mortality rate of 1.98%, the research utilizes AUC-ROC and AUC-PR as its main threshold-independent metrics instead of F1. Over 50 rounds and five random seeds, FedProx records the highest AUC-ROC (0.897) and mean AUC-PR (0.230), with paired t-tests validating its significant AUC-ROC advantage over other methods. Conversely, FedCluster (0.280) slightly outperforms FedProx (0.273) in F1 score. This study underscores the necessity of assessing FL in the context of heterogeneous and imbalanced clinical data.