Researchers Assess Federated Learning Models for COVID-19 Diagnostics

A team of researchers led by the University of Minnesota is examining federated learning (FL) techniques to evaluate their utility for diagnosing COVID-19 based on chest radiograph data.

FL models, a type of machine learning (ML), are a privacy-focused method for training artificial intelligence (AI) algorithms. The use of AI in healthcare has piqued significant interest in recent years, but data privacy concerns present challenges to AI use.

Traditional AI methods require large datasets. There are practical, ethical, and legal challenges to the data collection necessary for a robust dataset. For instance, training AI models typically requires that patient-related data be shared with a central repository, but data sharing across institutions may require patients to forfeit their rights to data control.

Conversely, FL models allow local data samples to be held on decentralized devices or servers. By training multiple AI models independently on separate computers, researchers can ensure that the models share only learned model weights and not any input data.

“Federated learning is an important future solution for AI in healthcare,” said Christopher Tignanelli, MD, an associate professor at the University of Minnesota Medical School, in the press release. “As all machine learning methods benefit greatly from the ability to access data that provides closer to a true global distribution, federated learning is a promising approach to obtain powerful, accurate, safe, robust and unbiased models.”

This ability to access data is especially useful during public health emergencies, such as the COVID-19 pandemic.

To determine FL’s potential utility in COVID-19 diagnostics, the researchers compared the performance of various federated and AI models using a previously described COVID-19 diagnostic model. They found that the FL models may not improve generalizability compared to other AI algorithms without resulting in poor internal validity, but these FL models may offer an opportunity to develop both internally and externally validated algorithms.

Outside of this application, the researchers indicate that the use of FL can provide multiple potential benefits in healthcare, including improving medical image and text analysis, creating better diagnostic tools for clinicians, advancing collaborative and accelerated drug discovery, decreasing cost and time-to-market for pharmaceutical companies, and addressing rare disease cases where no single institution has enough data to train models.

“We’re proud to be among the first teams implementing and further refining federated learning in real-world healthcare settings, with the strong support of industrial partners including Nvidia and Cisco,” said Ju Sun, PhD, an assistant professor at the University of Minnesota College of Science and Engineering, in the press release. “Data is the oil for modern AI, and federated learning makes the perfect oil refinery to advance AI for healthcare.”

This is the latest application of FL in medical research. Last month, a research team from the University of Pittsburgh Swanson School of Engineering received a $1.7 million National Institutes of Health grant to develop an FL-based approach to achieve fairness in AI-assisted medical screening tools.