Healthcare & Bio AI
Federated learning model predicted COVID-19 oxygen needs and ventilation or death risk across 20 institutes, with an average AUC above 0.92

A federated learning model called EXAM predicted the future oxygen needs of symptomatic COVID-19 patients from vital signs, laboratory data, and chest X-rays, according to a 2021 paper in Nature Medicine. The model trained on data from 20 institutes without moving patient data. It reached an average area under the curve (AUC) above 0.92 for outcomes at 24 and 72 hours after patients arrived at the emergency room. An AUC is a common measure of how well a model separates patients who have an outcome from those who do not. Compared with models trained on one site's data, the federated model improved the average AUC across all participating sites by 16% and increased generalizability by 38%.
What did the study measure?
The study focused on symptomatic COVID-19 patients who came to emergency rooms at the participating institutes. It used three kinds of input: vital signs, laboratory data, and chest X-rays. The model predicted future oxygen requirements at 24 and 72 hours and predicted mechanical ventilation or death at 24 hours.
What did the model achieve?
The model reached an average AUC above 0.92 for outcomes at 24 and 72 hours. Compared with models trained on one site's own data, the federated model improved the average AUC across all participating sites by 16% and increased generalizability by 38%. Generalizability means how well the model applied to data from other sites. At the largest independent test site, for predicting mechanical ventilation or death at 24 hours, the model had a sensitivity of 0.950 and a specificity of 0.882.
What is federated learning?
Federated learning is a method for training AI models with data from multiple sources while maintaining data anonymity. The sites do not exchange patient records. In this study, the approach let the researchers collaborate across 20 institutes without moving sensitive health data.
What does AUC mean?
AUC, or area under the curve, is a common score for how well a model separates patients who have an outcome from those who do not. A higher AUC means better separation.
What do sensitivity and specificity mean?
Sensitivity is the share of patients who needed mechanical ventilation or died whom the model correctly flagged at 24 hours. Specificity is the share of patients without that outcome whom the model correctly cleared.
Limits and what's next
The study used data from 20 institutes around the world and focused on symptomatic patients with COVID-19. That scope means the results may not apply to people without symptoms or to other conditions. The model used vital signs, laboratory data, and chest X-rays, so the findings are tied to those inputs. The abstract reports this as one study with an evaluation at one independent test site for the ventilation or death result. It does not describe a broader replication of the model across other external sites. The authors write that the work sets the stage for broader use of federated learning in health care, but the reported performance still comes from this one study.
Source: Ittai Dayan et al., “Federated learning for predicting clinical outcomes in patients with COVID-19,” Nature Medicine, 2021.