NVIDIA & Mass General Brigham Hospital Federated Learning Project Predicts COVID-19 Patient Oxygen…

Original article was published by Synced on Artificial Intelligence on Medium


NVIDIA & Mass General Brigham Hospital Federated Learning Project Predicts COVID-19 Patient Oxygen Need Using 20 Days of Data From 20 Hospitals

In its interim treatment guidance for hospitalized COVID-19 patients, the World Health Organization recommends clinicians begin supplemental oxygen therapy immediately. But there is no protocol for determining if a symptomatic patient presenting at an Emergency department requires immediate oxygen therapy, and if so, in what dosage.

To develop a robust AI model for predicting patient oxygen need and levels that would generalize to as many hospitals as possible, tech giant NVIDIA joined forces with non-profit hospital and physicians’ network Massachusetts General Brigham and 20 hospitals around the world in the federated learning-based initiative “EXAM” (EMR CXR AI Model).

Mass General Brigham Scientist Dr. Quanzheng Li developed the original model, CORISK, which combines medical imaging and health records to triage suspected patients showing COVID-19 symptoms.

Federated learning is a privacy-preserving technique that makes it possible for AI algorithms to learn from a vast range of data located at different sites, eliminating security risks associated with traditional data pooling methods. In this clinical setting, federated learning allowed researchers from various hospitals to collaborate on developing the model without directly sharing sensitive clinical data.

In just two weeks, the global collaboration achieved a model with .94 area under the curve (with an AUC goal of 1.0), resulting in excellent prediction for the level of oxygen required by incoming patients,” reads a NVIDIA blog post. The company says EXAM, which involved hospitals and patients from North and South America, Canada, Europe and Asia, is the “largest, most diverse federated learning” initiative yet.

Under the NVIDIA Clara Federated Learning Framework, participating hospitals use chest X-rays, patient vitals and lab values to train local models, as a centralized server would maintain the traditional global model. Each hospital sends updated versions back to the server, keeping all sensitive data within their own secure infrastructures.

To bring more hospitals onboard and further improve the model, NVIDIA will release its federated learning model in a few weeks as part of NVIDIA Clara on the NVIDIA GPU Cloud (NGC).