Multicenter Privacy-Preserving Federated Models for Predicting Postoperative Delirium and Acute Kidney Injury in Older Patients
Abstract
Postoperative delirium (POD) and acute kidney injury (AKI) are serious complications in older patients undergoing surgery, yet predictive model development is often constrained by single-center data limitations and privacy concerns that preclude centralized data sharing. To address these challenges, we retrospectively evaluated a simulated federated learning (FL) framework using multicenter datasets partitioned by hospital source, without sharing raw patient data across centers. A total of 7,216 non-cardiac, non-neurosurgical patients aged 65 years or older were included across five centers, with four contributing training data and one serving as an external validation site. Using a multilayer perceptron architecture, we implemented three federated algorithms and benchmarked them against local learning models (LLMs) and centralized learning models (CLMs). For POD, federated learning models (FLMs) achieved internal area under the curve (AUC) values of 0.725-0.726 and external AUCs of 0.700-0.701. For AKI, internal AUCs reached 0.780 and external AUCs ranged from 0.740 to 0.741. FLM performance was statistically comparable to CLMs (p > 0.05). These findings support federated learning as a feasible privacy-preserving strategy that achieved discrimination comparable to centralized learning for multicenter prediction of postoperative complications in older patients.
Read the paper
Related studies
- Active Learning during Federated Learning for Object Detection
- Fairness-aware federated learning
- Decentralized Federated Learning
- Building a Personally Identifiable Information Recognizer in a Privacy Preserved Manner Using Automated Annotation and Federated Learning
- Preservation of Patient Level Privacy: Federated Classification and Calibration Models
- Efficient explainable federated learning
Go deeper in Evidence Hub
Evidence Hub adds a study data sheet (design, sample size, evidence level, key numbers), an AI summary and a chat grounded in the paper's text.