While healthy lungs would usually show up as dark, these were showing the white haziness doctors describe as “ground glass” on chest x-rays and CT scans, resembling the splintering of a car windshield, and in some cases, a near total whiteout.
These patterns in the lungs were familiar to Dr. Savvas Nicolaou, a professor of radiology at the University of British Columbia and the director of emergency and trauma radiology at Vancouver General Hospital in Canada. In the 2000s, he was part of a team that analyzed chest x-rays of SARS patients. When COVID-19 started spreading in January, he teamed up with Dr. William Parker, a radiology resident at the University of British Columbia. Parker is a cofounder of SapienML, a company that has developed software to anonymize medical imaging data, along with Nicolaou and engineer Brian Lee.
The three of them had previously worked together on an artificial intelligence model trained on chest x-rays. “We thought, well, can we see the findings of COVID-19 in that model?” says Parker. They put out a call to collect as many chest x-rays and CT scans as possible to start building an open source AI model analyzing how the disease displays in the lungs, with the goal of developing an alternative way to diagnose patients besides existing tests.
Among those who answered the call to help Parker, Nicolaou and Lee build these CT scan diagnostics? Amazon.
In January, Nicolaou called Dr. Shez Partovi, a friend from his radiology residency 25 years ago, who now leads the healthcare life sciences, genomics and medical devices division at Amazon Web Services. He told Partovi about their plan to “use AI to empower physicians across the world in helping combat this disease” and the project took off from there.
Though most people know Amazon as a deliverer of goods, its biggest source of operating profit (67% for the fourth-quarter) comes from its colossal cloud computing business line, known as Amazon Web Services. The operation, estimated by one Wall Street analyst last year to be worth half a trillion dollars, offers a number of services to its customers, including storage, web hosting, and — of particular interest to researchers fighting the COVID-19 pandemic — machine learning applications.
In March, AWS announced its global Diagnostic Development Initiative, offering an initial $20 million in cloud credits and technical support to help accelerate research and development of coronavirus diagnostic tools. Instead of giving direct cash grants, the selected projects are getting a break on what they would normally have to pay AWS as a credit to their account. The tech support varies on a project basis, but includes access to AWS specialists. The program is supported by a technical advisory group of scientists and public health experts yet to be disclosed, except for Steve Davis, co-chair of the World Health Organization’s digital health technology advisory group. At launch, there were 35 global research institutions, startups and businesses involved in the project. AWS has since received more than 45 additional applications from customers, which are being evaluated.
Nicolaou and Parker’s team is the first public diagnostic customer, says Teresa Carlson, vice president of Amazon Web Services Public Sector. Other partners include the Chan Zuckerberg Biohub, a nonprofit research collaboration founded by Facebook CEO Mark Zuckerberg and his wife Priscilla Chan.The BioHub is utilizing AWS to optimize machine learning models with genomics data to estimate how many cases of a disease are in the population beyond what confirmed test results indicate. Estimates of the scale of pandemics can contribute to infectious disease research, and inform public health planning and preparedness.
Another participant is Beijing’s ETComm, which provides telemedicine services for medical institutions in China to remotely diagnose cardiovascular disease. The company completed more than 18,400 remote diagnoses of COVID-19 complications via its electrocardiogram reading platform built on AWS. Doctors at the University of California San Diego have received AWS credits for a clinical research study using artificial intelligence to speed up the diagnosis of pneumonia in COVID-19 patients based on chest x-rays.
The AWS diagnostic initiative was conceived in January after Carlson’s unit, which works with governments, educational institutions, nonprofits and NGOs in over 180 countries, was flooded with calls from customers asking to partner on projects surrounding COVID-19. Her hope is that by Amazon forming this consortium to power many different projects at once, participating institutions will ultimately choose to work together and share their findings globally, speeding the fight against the disease.
Many health systems have struggled to keep pace with the virus sweeping the globe, and public health officials, doctors and scientists have called for more and better testing. But even when patients do get access to tests, there are lingering questions about their overall accuracy and the rate of false negatives (that is, a patient has the virus but the test says she doesn’t).
The accuracy concerns in early accounts of COVID-19 lab tests “likely result from manufacturing and procedural failures,” says David Boyle, chief scientific officer and co-lead of the diagnostics program at global public health nonprofit PATH. Part of the problem is the speed at which some of the tests, which analyze for the presence of the virus in a patient sample, were fast-tracked by the federal government due to the nature of the crisis. Additionally, he says, labs are facing “a huge backlog of specimens for testing that stresses the human, material and logistical components of a laboratory system.”
The coronavirus pandemic and the urgent need for new diagnostics gets to the heart of a longstanding global public health problem — a lack of funding and coordination around diagnostics compared to issues like vaccines, says Steve Davis, who co-chairs the World Health Organization’s digital health technical advisory group and is a member of the AWS initiative’s technical advisory group.
Historically, one reason why there are so many difficulties with testing procedures during an epidemic is simply that “people have focused so much on the cure,” Davis says. But COVID-19 is changing that attitude, with more people realizing the value of diagnostics, he says.
Back in Vancouver, Nicolaou and Parker hope that by matching patterns, such as the percentage of the lungs with the ground glass pattern, doctors could help better diagnose patients and link the severity of lung damage to different stages, from hospital admission to ventilation, and, in the worst case, death. The team is compiling what Nicolaou says will be the largest dataset of COVID-19 positive images from around the world. While other local centers might have more data, they won’t have the range from different continents, which so far includes North America, Europe, Asia and Australia. A recent study of more than one thousand COVID-19 patients in China found CT scans were better at detecting the disease than commonly used polymerase chain reaction diagnostic tests.
Each image must be labeled by a human and then fed into the model to train the algorithm. All of the images are stored in the Amazon Web Service cloud, known as S3, or simple storage service. So far the team has tagged 1,000 images and there is a backlog of thousands more. There are currently three different artificial intelligence coding teams developing models: SapienML, the University of British Columbia and Amazon. Other support for the project comes from the University of British Columbia Cloud Innovation Center and the Vancouver Coastal Health Research Institute.
‘We're not releasing this as a for profit venture. We're releasing this as a humanitarian effort,” Parker says. The team is hoping to release the open source AI model within three months, so that other researchers and companies can start using it. Then they will work towards predictive modeling for patients and correlating their analysis with the existing laboratory diagnostic tests.
The ultimate goal is to go beyond an alternate way to diagnose patients, says Nicolaou. His and Parker’s dream is to use the gathered data to enable doctors to create simulations, where a patient’s CT scan could be used to virtually model how they might respond to certain therapeutics, thus improving care. With the virtual models, “you would actually start to see [the patient’s] response...before you deploy it in reality,” says Nicolaou.
Developing new ways to fight the pandemic is at the core of Amazon’s initiative, says Carlson. “Hopefully we can speed up the point of care and diagnostics very quickly over the next one to two years with this initial $20 million investment — but this is just a start,” she says. “If we burn through it sooner, we’ll come back to the table.”
Maneet Ahuja, Forbes Staff
Katie Jennings, Forbes Staff