UC San Diego Health Uses AWS to Implement Artificial Intelligence Imaging Model in 10 Days

When the COVID-19 pandemic hit the United States in the spring of 2020, UC San Diego Health researchers had already developed an image recognition model using machine learning to identify pneumonia in difficult-to-detect cases. Because pneumonia was quickly becoming one of the major indicators of severe infection in COVID-19 patients, Dr. Mike Hogarth, clinical research information officer at UC San Diego Health, asked Amazon Web Services (AWS) for help setting up a system for applying the model in a clinical setting that would enable medical practitioners to use the information in diagnosis and treatment.

When patient files and information are used in healthcare settings, data security is vital. The system for applying the artificial intelligence model would need to enable UC San Diego Health to meet strict rules in compliance with HIPAA to be used in a clinical environment. UC San Diego Health had set up similarly compliant solutions on AWS in the past, and this experience enabled the UC San Diego Health team to build its desired system in just 10 days using AWS.

UC San Diego Health is among the top 15 research universities worldwide. The UC San Diego School of Medicine conducts hundreds of clinical trials each year, and the UCSD Health Services Research Center specializes in the collection and analysis of health outcomes data to support research. In 2018, a team from UC San Diego Health, led by Dr. Albert Hsiao, associate professor of radiology at the UC San Diego School of Medicine, developed a way to use machine learning to detect pneumonia in X-ray images. “We created a probability map with color coding that indicated how uncertain or certain a pneumonia was,” says Dr. Brian Hurt, a resident physician on Dr. Hsiao’s team. According to Dr. Hsiao, “What people typically do is create a model that gives a simple yes or no output, indicating whether it’s pneumonia or not pneumonia. But we felt that producing an image was important to highlight where abnormalities are.” The team published a paper on these results in early 2020.

The UC San Diego Health team had previous experience using AWS to create a secure, HIPAA-compliant environment for its research data. “One of my first jobs at UC San Diego Health was to create an environment where protected health information could be computed upon and moved if needed,” says Dr. Hogarth. “Our team and AWS were having weekly meetings about optimizing it, and when the COVID-19 pandemic was hitting hard, AWS asked how it could help and if there was anything we wanted to do.” Dr. Hsiao’s team’s model immediately came to mind.

The goal was to implement a system that could take in X-rays from a clinical setting, run the model on them, and return results quickly to assist with diagnosis. “A lot of models created in labs like ours are potentially promising but not very useful without actually building them into the clinical workflow,” says Dr. Hsiao. Such an implementation would potentially affect the diagnosis, treatment, and outcomes of COVID-19 patients. “The findings of COVID-19 pneumonia are basically the same as those of any other viral pneumonia,” says Dr. Hsiao. “The model that Brian trained performs well in this population because it’s a good model for detecting pneumonia in general.” The model was useful in two ways. First, if it detects pneumonia in an X-ray image, it prompts a test or retest in the event of a previous—possibly false—negative test result. Second, in patients with known infection, finding pneumonia on an X-ray could indicate the illness’s severity and prognosis, informing treatments.

The UC San Diego Health research team had already set up the model to take in images and return them with a color-coded overlay. It only needed a cloud solution that could connect to the clinical imaging system to receive the images and output them directly into the patients’ files, making the images convenient for medical professionals to access and view. Because the team had already created HIPAA-compliant environments on AWS, it was able to get the project up and running in a mere 10 days. “The first day it was running on AWS, the model processed around 400 X-rays with very few glitches,” says Dr. Hogarth. In the next 6 months after implementation, the model processed over 65,000 X-rays, each in 3–4 minutes.

The model’s ability to provide information to physicians at the point of care is what makes it so useful, and AWS has been critical in making that not only possible but straightforward and simple to maintain. According to Dr. Hogarth, out of an information technology team of 500 people at UC San Diego Health, a single member can verify continued compliance with HIPAA and other regulations in the AWS environment as part of their duties. And when it comes to implementing the model created by Dr. Hsiao’s team in a clinical setting, Amazon Elastic Compute Cloud (Amazon EC2) instances are primarily what are needed. The environment offers the security configurations the team needs and is simple to resize as the compute capacity increases.

A recent paper published by the Journal of the American College of Emergency Physicians Open indicated that implementing this model has impacted clinical decision-making 20 percent of the time. “There are very few things that we know of that really impact clinical decision-making to that extent,” says Dr. Hsiao. The model’s initial accuracy was 86 percent, and the team will soon deploy an even more accurate version that accounts for pneumonia cases often missed when located behind the heart.

Evaluation of machine learning–based clinical decision support within existing point-of-care workflows is important but relatively uncommon. Although evaluation of this image analysis tool is still in its infancy, there is anecdotal evidence that it is having a positive impact. Recently, a 78-year-old patient was admitted with fever and abdominal pain. Doctors had not been considering a COVID-19 diagnosis, but the model showed signs of pneumonia on a chest X-ray. So, they tested the patient, and it came back positive for the virus.

Dr. Hsiao’s team plans on continuing and refining the model, but the idea of evaluating machine learning and artificial intelligence algorithms at the point of care has potential applications across a wide range of other healthcare research as well. “For us, it’s a data and decision support pipeline,” says Dr. Hogarth. “We’ve demonstrated the use of the pipeline with these images, but there could be many other applications too.”