AI Triage Model Predicts COVID Oxygen Needs
via Shutterstock
Among the first steps when COVID-19 patients are hospitalized is a blood oxygen test used to determine whether they require supplemental oxygen.
Given the “wildfire” nature of the pandemic (a more apt description of the pandemic than the standard “wave” metaphor, argues one epidemiologist), AI researchers have plenty of data upon which to train models that can help triage coronavirus patients. This week, Nvidia (NASDAQ: NVDA) said it has developed along with healthcare partner Mass General Brigham Hospital an AI model for predicting the level of oxygen required by new patients.
The EMR CXR AI Model, or EXAM, was developed in collaboration with 20 other hospitals as part of Nvidia’s federated learning initiative. The EXAM partnership and other AI model development are managed under the graphics chip maker’s Clara healthcare platform.
CORISK, the model upon which EXAM is based, was developed at Mass General Brigham, combining medical imaging with health records. The oxygen prediction workflow includes chest x-rays and vital statistics like blood pressure, blood oxygen levels and respiratory rate. The model then predicts whether a patient will require standard hospital “room air” ventilation, low- or high-flow oxygen or a ventilator. The triage tool also helps physicians determine whether patients should remain under observation or be transferred to intensive care units.
The partners said the federated learning training framework achieved a level of accuracy of 0.94, the goal being 1.0, “resulting in excellent prediction for the level of oxygen required by incoming patients.”
Nvidia said the EXAM model would be released as part of its Clara platform in the next several weeks.
The federated learning framework is designed to protect patient data while allowing diagnosticians to share medical imagery, patient vital signs and lab results used to train a local model. Only a subset of the resulting model weights is shared with a global model, Nvidia noted. The EXAM model was then trained using distributed data from patient data sets across Asia and Canada as well as North and South America.
Among the participants in the AI model training effort was the U.S. National Institutes of Health (NIH). Each medical institution used Nvidia’s Clara to train local models and fold their results into EXAM. They also employed secure, in-house servers to protect patient privacy, then used a separate, dedicated server hosted by Amazon Web Services (NASDAQ: AMZN) that supported a deep neural network used to train the global model.
Each participant received a copy of the model to train on its own data set.
Nvidia previously said it is working with NIH clinicians and data scientists under cooperative R&D agreement to generate new AI models targeting COVID-19 patients.
Clara federated learning runs on Nvidia’s EGX edge platform.
