Elnur - stock.adobe.com

News Stay informed about the latest enterprise technology news and product updates.

Nvidia brings federated learning to COVID-19 patient data

Researchers from 20 healthcare institutions built an AI model to predict the supplemental oxygen needs for COVID-19 patients. And they did so without ever sharing patient data.

Twenty healthcare institutions from around the world have trained an AI model to accurately predict whether a COVID-19 patient will need supplemental oxygen -- an effort they accomplished through collaboration.

Researchers at organizations including the University of Wisconsin-Madison School of Medicine and Public Health and NIHR Cambridge Biomedical Research Centre started with an AI model initially developed by researchers at Massachusetts General Brigham Hospital. They then partnered with GPU chipmaker Nvidia to launch the EMR Chest X-ray AI Model (EXAM) federated learning initiative. Together, they trained an AI model, which is not in clinical use, to predict whether an emergency room patient with COVID-19 symptoms will need supplemental oxygen hours or days after an initial exam.

Participants trained the AI model using the Nvidia federated learning framework, a machine learning technique that uses data residing in separate servers rather than data in a centralized location. Each of the 20 participants trained the AI model on internal data and then shared their model parameters or how the local model learned to make predictions on their internal data. That information was used to tweak a global model, which was then sent back out to the 20 participants for more training. The iterative process was repeated until the models converged.

The federated learning technique helps to both preserve patient data privacy, as healthcare facilities keep their data on premises, and expose the AI model to a massive, diverse data set, which broadens the model's predictive ability, making it applicable to a wider range of healthcare facilities.

The Nvidia federated learning framework opens the door to faster development of trustworthy and more widely applicable AI models for healthcare, said Fiona Gilbert, professor of radiology at the University of Cambridge.

"This federated learning approach is a terrific way forward because it safeguards the patient's data and it allows multiple institutions to take part at the same time," Gilbert said. "It allows us then to make something we think will be much more globally useful compared to just developing something in our own institution with our own groups."

The Nvidia federated learning framework allowed 20 healthcare facilities to work together to build an AI model for predicting COVID-19 patient oxygen needs.
Twenty healthcare facilities participated in a federated learning initiative to build an AI model for predicting COVID-19 patients' oxygen needs without ever sharing data with each other.

Nvidia federated learning framework

Gilbert and her research team were already pulling together a COVID-19 patient data set when they decided to participate in the federated learning initiative.

The Cambridge area didn't have large numbers of COVID-19 patients, she said, which meant the facility didn't have a data set big enough to train an AI model like EXAM. The federated learning initiative gave Gilbert and her team a way to share their expertise and data while benefiting from the model training at other healthcare organizations.

A separate server hosted on AWS stored the global AI model, but each of the 20 facilities got a copy of the model to train on its own patient chest X-ray data, vitals and lab values. All facilities used Nvidia Clara, a platform for building AI models for healthcare, to train its local models and participate in the EXAM federated learning initiative.

Gilbert felt the experience showed the value of federated learning for building evidence-backed AI models and hopes to apply the practice to other conditions, such as breast cancer. She also noted that training an AI model on data sets from institutions around the world helps remove potential model bias that could be incorporated when using only limited or constrained data sets.

This is the first step of demonstrating that this is a very good way of working -- it's a model that's super useful.
Fiona GilbertProfessor of radiology, University of Cambridge

"This is the first step of demonstrating that this is a very good way of working -- it's a model that's super useful," she said. "Often, an institution makes an AI tool for, say, a mammogram, but you can't use it in someone else's institution because you're using a different type of machine to acquire the breast X-rays or it's not relevant internationally because they do something differently in America. If you're building tools using this federated approach, you don't get that bias from only building something from one country or one institution."

John Garrett, assistant professor and director of informatics at the University of Wisconsin-Madison's radiology department, said he and his team were also already collecting COVID-19 patient data and were intrigued by the idea of working with other organizations to train an AI model on that data.

"I was excited about an opportunity to not only get involved in an initiative outside of just our own health system, but take a model that's been previously developed and proved that it works at many sites and can perform even better when those data sets are used, but to also repurpose some of the data we were collecting anyway," he said. "That's something I feel like is always an underappreciated aspect of projects like this, is the chance to take that effort and turn it into something reproduceable and reusable."

The Nvidia federated learning framework will be made publicly available later this month as part of Nvidia Clara on the Nvidia GPU Cloud (NGC), which is designed for machine learning projects.

Dig Deeper on Clinical data analytics software and systems

Start the conversation

Send me notifications when other members comment.

Please create a username to comment.

SearchCompliance

SearchCIO

SearchCloudComputing

SearchMobileComputing

SearchSecurity

SearchStorage

Close