Cutting edge federated learning tools to help medical industry adapt to new realities

7/27/2020 Aaron Seidlitz, computer sciences

Written by Aaron Seidlitz, computer sciences

For years, Illinois Computer Science professor Sanmi Koyejo focused his research in artificial intelligence on two areas—federated learning tools and biomedical image analysis. Together, he believes, these advancements could alter the way healthcare entities react to more data.

The right people have noticed this work, and support for Koyejo’s research has grown as his workgroup tries to help the world react rapidly to the COVID-19 pandemic.

To further this effort, Koyejo’s research group recently earned an award from C3.ai Digital Transformation Institute (DTI) to alter the way healthcare entities make data-driven decisions.

The goal driving work at DTI, according to Thomas M. Siebel, CEO of C3.ai and Illinois CS alumnus, is to join “leading scientists, researchers, innovators, and executives from academia and industry” to “accelerate the social and economic benefits of digital transformation.”

By furthering these projects, the potential is there, according to Siebel, to “change the course of a global pandemic.”

“Just looking at the list of projects from DTI and understanding what they’re working on, inclusion with that group is so exciting,” Koyejo said. “We have a great team, and I’m looking
Sanmi Koyejo
Sanmi Koyejo
forward to the future beyond this initial work. It’s inspiring to think about how computer science can impact the world.”

To begin this process with healthcare entities, Koyejo’s team is working on the way they use data. The title of their project awarded by DTI is “Secure Federated Learning for Clinical Informatics with Applications to the COVID-19 Pandemic.”

The workgroup includes Koyejo as primary investigator, alongside:

  • Co-Investigator: Dakshita Khurana, professor at Illinois CS
  • Co-Investigator: William Bond, OSF HealthCare and Jump Simulation
  • Senior Personnel: George Heintz, Healthcare Engineering Systems Center, University of Illinois at Urbana-Champaign
  • Senior Personnel: Roopa Foulger, OSF HealthCare

“Our work is forward looking. Our work is hoping to improve some of the tools at decision-makers’ fingertips to help them now, and, perhaps even more importantly, down the road,” Koyejo said. “This could affect ongoing care, preparedness, and what care looks like in the future. We know, or expect, that contagious disease outbreaks will continue, and part of our goal is to be more prepared next time.”

By understanding that the bread and butter of machine learning is data, federated learning trains machine learning models on distributed and un-shareable data.

As Koyejo says, “one thing federated learning does, is, instead of taking data to computation, it takes computation to data.”

This advancement comes at the same time industry has begun taking advantage of more data sources. Cell phones are becoming more powerful. Edge devices have processors included. Even hospitals are turning to cloud computing sources or powerful in-house computation.

“If data is the bread and butter of federated learning, a corollary of that statement is the more data you have, the better,” Koyejo said. “That’s because more data means your work is more representative. For individual hospitals, the benefit to federated learning is enhancing automated decision support systems.”

Of course, by expanding data sources and methods to analyze the data, security also becomes an issue. And in healthcare, security is always an issue; as health systems seek to keep patient data private.

That is why collaboration with his colleague at Illinois CS is exciting to Koyejo. Khurana's research emphasizes cryptography and privacy/security. This will provide new “protocols to enable machine learning on distributed private data.”

Their project will use public data from the C3.ai data lake, as well as private date from research partner OSF Healthcare.

“Our primary goal is to enable multiple institutions with their own private medical data to collaboratively train accurate models on their combined datasets,” Khurana said. “Privacy regulations prevent healthcare providers from pooling their data in one place to perform better clinical inference. Cryptography, via a tool called secure multi-party computation, can nevertheless enable these institutions to compute arbitrary functions of their joint data without having to share it with each other or with a third party.”

The tools this research team will produce come from a combination of secure data and improved prediction quality.

“To the extent we’re successful, this can help decision makers make the difficult choices," Koyejo said. "This is about which patients need further treatment in the hospital. These tools can also help with decisions about discharging patients or who needs the limited medical equipment in a time of need.

“Beyond that, the scope of this grant will build toward something more. Democratizing healthcare information is about removing barriers to information and improving disparities in the quality of care.”


Share this story

This story was published July 27, 2020.