Vancouver doctors developing a new AI tool to predict COVID-19 infection

An open-source artificial intelligence tool is being developed to predict the presence of a COVID-19 infection in patients— and it’s happening right here at UBC.

Dr. Savvas Nicolaou, professor of radiology at UBC and director of emergency and trauma imaging at Vancouver General Hospital (VGH), and Dr. William Parker, radiology resident at UBC and VGH, are co-leading a project to develop an open-source AI application which may be able to predict the presence, severity and complications of COVID-19 infections. The tool utilizes X-ray and CT scan images as training data.

The project is being funded by the UBC Community Health and Wellbeing Cloud Innovation Centre (UBC-CIC) and is in partnership with Amazon Web Services, Amazon’s cloud computing and database storage platform.

Radiologists, residents and UBC medical students alike are collecting and analyzing CT scans from COVID-19 patients from around the world. These collected images will be used to create and train the AI model which will predict the presence and severity of COVID-19 and other lung diseases. Currently, there is no existing tool that predicts the disease’s severity and clinical impact in various patient groups.

“The medical challenges with COVID being such a new disease … is that it generates very unique fingerprints of what pneumonia looks like, or [what] lung infection looks like,” said Dr. Kendall Ho, VGH emergency physician and UBC-CIC academic director. “And so when we look at those X-rays or CT scans, we can pick up in our eyes what might be some of the characteristics.”

The lungs of COVID-19 patients appear white and hazy in these images, in comparison to healthy lungs which appear more transparent. Thousands of images have already been collected, but according to Ho it will take tens of thousands before a project relying on AI becomes completely accurate.

Once the project is fully developed, it will pilot at VGH with the aim of being used in routine diagnostic procedures to increase the accuracy of COVID-19 diagnoses.

Though the community has high hopes that the pandemic will not return, it is likely that there will be multiple future outbreaks of COVID-19 before we have a vaccine developed which will help establish herd immunity. That’s where the AI model comes in.

According to Ho, the image database will help in diagnosing the disease more accurately, recognizing its presence more quickly and providing patients with optimal care based on their stage of illness. This will assist in them being able to recover in a more predictable way.

“It’s fantastic to be able to get a group of scans of patients with different diseases at different stages of disease, so that we can then identify not only the pattern that’s very typical of COVID-induced pneumonia, but also help us to understand the stage of disease as it evolves,” said Ho. “[This will] not only help us to manage the patient better, but also help us to predict the disease and how severe they can get.”

Some patients have been admitted to the emergency room with abdominal pain, stroke and acute chest pain, which are atypical symptoms of COVID-19. However, after analyzing their CT scans it was clear they were infected with the virus due to the presence of prognostic indicators, or characteristics of a patient infected by a disease. These patients all had white, hazy lungs.

“We have excellent support around UBC with different faculties, different students wanting to step up and support different faculty members who want to support this project. So it becomes an exceptional learning experience for all of us,” said Ho.