December 3, 2020
‘All hands on deck’: How Duke University and AI for Health raced to create a COVID-19 solution for patients
The 7-month-old was crying. Amanda Randles picked her up, held her and continued talking on the video meeting call from home. Nearby, her other 7-month-old baby started to wail. Randles put one baby down, picked up the second and cradled him as he quieted down. The rotation of the crying twins continued during the meeting.
It wasn’t an ordinary work-from-home session for Randles, Duke University assistant professor of biomedical sciences, and one of the nation’s experts in high-performance computing and biomedical engineering.
She and a group of two dozen tech and science professionals from the North Carolina university and from Microsoft AI for Health had been virtually introduced to each other on that April day for the first time – and time was in short supply. They were racing to submit a proposal to the U.S. Food and Drug Administration (FDA) for emergency use authorization of a Y-shaped device called a ventilator splitter and resistor system to help COVID-19 patients, as a growing number of hospitals continued to face a shortage of ventilators.
Randles’ colleague, Muath Bishawi, developed the device, while Randles and her team at Duke created the software for the ventilator splitter resistor system using airflow simulations so it could be customized for patients.
That customization is crucial for making sure patients get the correct amount of oxygen. Though the FDA gave emergency approval allowing two patients to share one ventilator during the pandemic, medical professionals believed it was not optimal, and in some cases, could cause more harm than good.
For a shared ventilator to work well patients needed, among other factors, to be approximately the same weight and have similar lung compliance – a measure of the lungs’ ability to stretch and expand. Finding two such similar patients on a hospital ward can be a game of chance. The ventilator splitter resistor system makes it a safer option.
“In real life, you don’t want to split ventilators. They’re not built to be split,” says Randles. “It is really only in these dire circumstances that you should even consider splitting them.”
In addition to the voluntary work done on the project by Randles and her team at Duke University, including Mike Kaplan, a medical student, Simbarashe Chidyagwai, a graduate student in biomedical engineering, as well as the Duke Office of Information Technology, others jumped in to help on an unpaid basis. A software developer created a mobile app for doctors to quickly make decisions about matching patients and ventilator settings. Another company stepped up to do the 3D printing of the ventilator splitters themselves.
“Everyone was doing this completely because they thought it was important and they wanted to help,” Randles says. “Everyone was motivated to get this out to doctors as quickly as possible.”
Duke University and Microsoft were initially paired through the COVID-19 High Performance Computing Consortium, created in March by the White House Office of Science and Technology Policy (OSTP). The consortium, whose members include Microsoft, aims to provide COVID-19 researchers worldwide with access to the world’s most powerful high-performance computing resources to fight the virus.
Geralyn Miller, Microsoft AI for Health senior director, is on the consortium’s technical review committee, which reads and reviews COVID-19-related research proposals. She says when Randles’ project came through, it was “incredibly unique” among the hundreds of proposals that were submitted because of its “immediate lifesaving measures.”
“It’s really great that there’s a lot of scientific research happening on the virus itself. It needs to happen,” Miller says. “But this one, this is one that really is focused on saving lives. It was the only one that I’d seen that talked at all about immediately beneficial therapeutic intervention.”
Miller learned about and empathized with Randles’ chaotic schedule, which includes overseeing a research lab at Duke called Randles Lab and caring for a toddler daughter as well as the twins. Miller, a mother of five, was hired at Microsoft in 1998 as a software developer, a time when few women were in that profession.
“Women are always doing a professional juggling act,” Miller says. “There’s always some element of that happening whether or not it is as intense as it has been with COVID-19, with people juggling family obligations and home schooling and the general things that come with living through a shutdown.”
Randles and the Duke team learned on a Monday they were matched with Microsoft AI for Health to provide the high-performance computing power with Microsoft Azure. By Thursday, they had that first virtual meeting with Microsoft experts in various areas including fluid dynamics, Azure infrastructure and networking, high-performance computing and AI for Health.
“It was all-hands-on deck, let’s get this up and running,” says Miller.
The goal of Duke’s work was “to provide physicians with the entire parameter space, so that no matter what patient weight or what lung compliance they ran into, that data was already there and would provide them with the guidance they needed to determine how best to use this ventilator splitter,” Randles says. “To do that, we needed to pre-compute all the different combinations of parameters, and that led to the need to run hundreds of millions of simulations that, in total, required almost one million compute hours.
“It’s something that was just not possible in the conventional computers and clusters that we had on campus at Duke, at least in the turnaround time that we needed,” she says. “It would have taken months to complete if we had just done it on local resources at Duke.”
Three days after that Thursday meeting, the initial work had been done, with 800,000 hours of compute time logged in 36 hours.
The team then spent the next three days finalizing the project.
“There were emails flying back and forth over the weekend at 2 and 3 in the morning from people trying to track everything down,” says Randles. “The whole team from across the three institutions did not sleep and got things done.”
Randles long has known about not getting sleep and about getting things done. She and her husband, a biochemist, have honed their work-at-home routine during COVID-19 to make it as manageable as possible.
“We’ve done a lot of one person gets the kids for an hour-and-half, and then the other person gets the kids for an hour-and-a half,” she says. “And you kind of trade off at these weirdly small-time intervals because there’s so many meetings in between that we needed that kind of flexibility and granularity to get on to virtual meetings that we both needed to attend.”
Already highly accomplished in her field, the Association for Computing Machinery presented Randles in 2017 with the Grace Murray Hopper Award, given annually to an outstanding young computer professional for a single recent major technical or service contribution.
Randles created computer code that can model the entire arterial system at a subcellular resolution, something that can help show areas in the body where vascular disease may occur. She named the code HARVEY – after William Harvey, the 17th century physician who was the first doctor to accurately describe how blood was pumped around the body by the heart.
MIT Technology Review also named Randles as one of 2017’s “Innovators under 35” and a “visionary” for her work on HARVEY, which has continued to evolve. To make HARVEY more accessible to those in the sciences and medicine, Randles and the Duke team recently introduced a graphical user interface, called Harvis, described in a study published in the Journal of Computational Science.
The ventilator splitter project was something very different for Randles. But it was also something she very much wanted to do when colleague Bishawi asked for her help.
“When we first went on lockdown for COVID-19, I didn’t think there was any way of contributing,” she says. “I wanted to do something useful. With the ventilator splitter, I thought it was exciting that there was actually a way that we could concretely have an impact.”
That immediacy and impact to make things better was among the reasons AI for Health was so interested in the project. Part of the AI for Good initiative, AI for Health is a $60 million, five-year program to empower researchers and organizations with AI to improve the health of people and communities around the world. The program was developed in collaboration with leading health experts who are driving important medical initiatives.
Since April, AI for Health has awarded more than 150 grants to COVID-19 projects around the world.
“One of the things we’re really focused on in AI for Health is societal impact,” Miller says. “Amanda’s project maps very well into our strategy for COVID-19, where one of the areas we’re thinking about is allocation of resources, how to do things like allocate ventilators and ICU beds and PPE (personal protective equipment).”
The FDA has not yet approved the emergency use authorization of the ventilator splitter resistor system, but Randles finds some comfort in that.
“It’s positive that we keep getting put on the back burner,” she says. “That means they don’t need it right this second.”
But if they do, plans can quickly be put into place.
“This work is incredibly important,” Randles says. “The number of cases of COVID-19 are rising. There’s still a finite number of ventilators available, and as we’re starting to push that capacity, doctors need more options, more capabilities and they need more data. And we’re providing them information that’s critical and will hopefully improve patient outcome and patient care.”
Top image: Researchers at Duke University created ventilator splitters that are 3D-printed, and can help make sharing of ventilators safer. Photo credit: Duke University