SalmaZakzouk168/eICU-FL-Robustness-Pipeline: v1.1.0

Abstract

This repository contains the complete, reproducible pipeline used in Zakzouk, S.; Selim, S.; Said, L.A., "Robust Federated Learning for ICU Mortality Prediction Across a Heterogeneous Hospital Network" (submitted to Machine Learning and Knowledge Extraction, MDPI). The pipeline covers eICU-CRD cohort filtering, feature extraction, centralized architecture selection (logistic regression, MLP, and kernel-size-one CNN variants), class-imbalance handling, federated learning via FedAvg and FedProx, systematic robustness evaluation under label noise, input noise, and Byzantine client attacks, hospital-size stratification, comparison against the APACHE IVa clinical severity score, and coefficient-based feature importance analysis. This repository does not include raw patient data. The eICU Collaborative Research Database (eICU-CRD v2.0) is a restricted-access, credentialed dataset hosted on PhysioNet; its Data Use Agreement prohibits redistribution. To reproduce this analysis, obtain independent credentialed access at https://physionet.org/content/eicu-crd/2.0/ and populate the eicu_data/ folder as described in the README.