Federated learning for predicting clinical outcomes in patients with COVID-19
Problem
Predicting future oxygen needs of symptomatic COVID-19 patients using data from many hospitals without sharing patient data.
Why it matters
Early prediction of oxygen requirements helps clinical triage while data anonymity removes barriers to multi-site collaboration.
What the AI did
A federated learning model, EXAM, was trained across 20 institutes on vital signs, lab data and chest X-rays to predict patient outcomes.
Results
Average AUC >0.92 for 24 h and 72 h outcome prediction; 16% average AUC improvement and 38% generalizability increase over single-site models; sensitivity 0.950 at largest independent test site.
Impact
Outperformed single-site models and generalized better across participating sites.
Record
| Venue | Nature Medicine |
|---|---|
| Published | 2021 |
| Work type | article |
| Authors | Ittai Dayan, Holger R. Roth, Aoxiao Zhong, Ahmed Harouni, Amilcare Gentili, Anas Zainul Abidin, Andy Liu, Anthony Beardsworth Costa et al. |
| Citations | 823 (OpenAlex, at import time) |
| References | 76 |
| Model or method | EXAM (electronic medical record chest X-ray AI model) |
| Organization | not stated in source |
| Primary source | https://doi.org/10.1038/s41591-021-01506-3 |
| Open access | hybrid |
| Keywords | federated learning, COVID-19, chest X-ray, oxygen prediction, multi-site, clinical AI |
Importance rating
significant — thousands to millions affected
COVID-19 patient triage during a pandemic; 20 institutes globally; better prediction supports clinical decisions and demonstrates federated learning for healthcare.