How UC San Diego Health, AWS implemented an AI imaging algorithm to detect COVID-19 in 10 days

Earlier this year, Michael Hogarth, MD, clinical research information officer at UC San Diego Health, sat in on a presentation about how radiologists could use machine learning decision support to identify pneumonia in X-rays images; the presenter ended his session by mentioning the algorithm could also detect COVID-19.

Albert Hsiao, MD, PhD, an associate professor of radiology at the University of California San Diego School of Medicine, worked with Brian Hurt, MD, of the UC San Diego School of Medicine radiology department to develop the machine learning algorithm allowing radiologists to use AI to augment their readings. The algorithm was trained in 22,000 notions to spot pneumonia in chest X-rays and then produce color-coded maps that showed the probability of pneumonia.

His team then began to look at X-rays of five patients in China and the U.S. who had tested positive for COVID-19 and published their results. Over the next several weeks, the health system was busy responding to the pandemic and when Amazon Web Services reached out to Dr. Hogarth, asking how they could help, he remembered Dr. Hsiao's presentation.

The two organizations decided to work together and deploy the algorithm systemwide. It took just 10 days for UC San Diego Health to implement the algorithm and begin collecting data to study the results. The technology for chest X-rays is portable and cost-effective, and clinicians can return results faster than other COVID-19 tests.

"We had the infrastructure, it was just a matter of scaling it up," said Dr. Hogarth. "On the first day we hit 500 images. Our process slowed down and AWS showed us how to quickly add more servers by just clicking a few buttons."

AWS gave UC San Diego Health service credits to roll-out the program and study the results. After six weeks, the health system reported 15,000 images, and now the number is closer to 20,000 images studied. There is a lot of data gathered on algorithms that can detect issues in clinical images, but they are rarely implemented in real-world environments because it's challenging to integrate into the EHR, and healthcare organizations need a culture that accepts machine learning decision support.

While the algorithm is still considered investigational, Dr. Hogarth said he is excited by the early results. In one recent case, a man arrived at the emergency room with diarrhea and a fever. His clinicians did not expect him to have COVID-19 and his chest film was initially read as "fine"; however after it was put through the algorithm, the system indicated COVID-19 was present. The patient then tested positive for the coronavirus.

"The algorithm clearly had an impact in that case," said Dr. Hogarth. "It was a great example of a real-world implementation using integration between the cloud and our system."

UC San Diego Health also has a young investigator using artificial intelligence to provide real-time feedback to the electronic ICU clinicians about sepsis. The AI technology is designed to identify the indications that someone may be developing sepsis.

"This young faculty member has one of the best algorithms around and it can successfully predict sepsis early on," said Dr. Hogarth. "He is providing the same technique to COVID-19 patients that will likely need intubation as well as ventilators and dialysis. It can predict if someone is getting in trouble before we do."

More articles on artificial intelligence: 
Microsoft AI health director: How AI is fueling intelligent health systems
Mayo Clinic to implement AI-powered predictive patient triage platform
GE, University Hospitals Cleveland partner on AI lung-imaging tech

© Copyright ASC COMMUNICATIONS 2020. Interested in LINKING to or REPRINTING this content? View our policies by clicking here.

 

Featured Webinars

Featured Whitepapers