The UW – Madison School of Medicine and Public Health helped create an Artificial Intelligence (AI) model for predicting supplemental Oxygen need for COVID-19 patients. The model was devised by researchers at NVIDIA and Massachusetts General Brigham Hospital to help physicians determine the best course of treatment for patients reporting to the ER with COVID-19 symptoms.
After establishing the model, the researchers began a federated learning initiative with 20 hospitals in eight different countries, including UW. Each institution provided data from chest X-rays, vitals, and lab results to construct the model’s algorithm. UW was chosen because it has the infrastructure to provide the curated datasets needed for this model. Out of the 20 sites, UW provided the second most data, helping to rapidly improve the tool so that it could be used during the current pandemic.
Department of Radiology Assistant Professor John Garrett was the Principal Investigator at UW and helped facilitate the school’s involvement in this project. In addition to success in creating the AI model, Garrett believes UW’s participation was significant because it provided a valuable opportunity to participate in a federated learning project, allowed UW to network with other institutions, and set the groundwork for future trials with AI models.
“I see this type of work as a significant step towards more broad acceptance of clinical AI tools. It can be challenging to build and then trust a tool like this with a single site’s data to validate, but traditional ways of performing multi-center studies require years of work. Federated learning lowers the barrier to this type of collaboration without sacrificing security or HIPAA compliance. I think it represents a way for researchers to much more rapidly iterate and develop new tools,” says Garrett.
The recent surge of COVID-19 cases in Wisconsin has highlighted how valuable AI models are in the current world. “The major implication of the current surge has placed an emphasis on the value of these tools which are now validated and ready to test clinically. Although we will need to work with the IRB and other compliance groups before rolling this out in clinical practice, there is clearly a need for tools like this to help us utilize our resources efficiently,” he says.
Learn more about the AI Model here.