CLN - Feature

Artificial Intelligence Is Poised to Transform Point-of-Care Testing

In both research and practice, advanced software is being deployed to develop new kinds of instruments and help clinical laboratorians monitor quality.

Jen A. Miller

Artificial intelligence (AI) and machine learning (ML) continue to become more prevalent in healthcare. The AI-associated healthcare market is on track to reach $6.6 billion this year, according to Accenture. AI applications could create opportunities for $150 billion in healthcare cost savings by 2026. And the use of these technologies has even expanded into the field of point-of-care testing (POCT). As more tests move out of the core laboratory and are performed by nonlaboratory professionals, AI and ML are being used to make POCT cheaper, faster, and easier to check for quality control.

“We generate a lot of data with POCT and in our core laboratories,” said James Nichols, PhD, DABCC, FADLM, professor of pathology, microbiology, and immunology and medical director of clinical chemistry and POCT at the Vanderbilt University Medical Center. “We haven’t even really scratched the surface on what we can do with that data.”

Cheaper and Faster POCT in More Settings

There’s no lack of enthusiasm for advancing POCT, either among healthcare professionals or even patients, who increasingly expect that most tests will soon be as simple and reliable as home pregnancy tests or blood glucose meters. But translating that ease and price point to other kinds of testing, especially molecular testing, hasn’t been easy. Some new innovations show that AI and ML can help.

Aydogan Ozcan, PhD, Chancellor’s Professor and Volgenau Chair for Engineering Innovation at the University of California (UC), Los Angeles, has been developing disposable vertical flow assays that offer cardiovascular risk stratification using high sensitivity C-reactive protein (NPJ Digit Med 2020;3:66) and an immunoassay for early-stage Lyme disease detection (ACS Nano 2020;14:229-40).

In about 15 minutes, the tests show a reaction across immunoreaction spots on a sensing membrane. The user then captures an image of the spots on a smartphone, and a deep neural net that has been trained to recognize patterns processes the image to indicate the concentration of C-reactive protein, or a positive Lyme test.

Nam K. Tran, PhD, associate clinical professor in the department of pathology and laboratory medicine at UC Davis, recently led a study that looked at a range of variables for predicting acute kidney injury in burn patients and found that AI and ML help make these predictions quickly and accurately. “Humans are typically only able to integrate about six to seven variables at one time, while machines integrate a wide range of variables at once,” he said, adding that tests that automate test interpretation using AI could be more widely distributed, since their users would not need to have in-depth technical backgrounds.

This could bring down costs and make tests easier to use in a variety of settings, including in patients’ homes. “You can imagine lots of tests that need to be administered extremely inexpensively and in field conditions,” said Ozcan. He hopes that within about a decade these tests will be “at the level of a CVS test, where patients will be able to buy it off the shelf and activate it at home, in the same way that today’s glucose monitors are operating.”

AI/ML Maintains POCT Quality Control

Glucose meters are also a good example of where AI started, said Nichols. “They used to be all manual, so you had to apply the blood from the fingerstick to test a strip, wipe it, then insert it into a glucose meter and read the result,” said Nichols. Then the user had to record the reading.

In the next generation of glucose meters, the result was stored in the meter itself, and uploaded to laptops carried around the hospital. Laboratories could use that data to calculate a hospital-wide group mean, each nursing unit’s group mean, and a group mean for each meter. That could help identify which meters weren’t working correctly.

“That was really the start of AI and utilizing big data to prove competency of operators, to prove that meters were giving results that were close to the other meters in the hospital,” he said. “Jump forward many years and all our POC devices tend to have data-management features. We have large amounts of data coming from individual hospitals. Let’s group all that together.”

Clinical laboratorians should also get accustomed to using AI and ML to maintain quality control in devices used outside of the hospital and operated by nurses, emergency medical technicians, and others who don’t have the training that laboratory professionals do. “POC devices are being used in physician offices and clinics, but they're also being used in helicopters and with visiting nurses,” Nichols said. The devices are thrown in the backs of trunks and get tossed around in ambulances. “They’re used in cold and wet conditions, in dust, wind, and sand. How do we ensure quality beyond just the device failing and giving an error message?”

AI also could be applied to the pre-analytic phase of testing, Tran noted, which would be similar to how the automotive industry uses it for vehicle quality control. “AI could be used to evaluate the quality of specimens, including but not limited to examining how specimens are collected and detecting interferences in samples. Machine vision could also be used to watch operators perform a test.”

Hurdles to New POCT Technology

Just because a technology is new and fancy doesn’t mean it’s also worth it—or that it even does what it says it’s supposed to do. In 2012, the University of Texas MD Anderson Cancer Center partnered with IBM Watson on an AI-enabled “Oncology Expert Advisor.” It didn’t work, and cost the healthcare system $62 million, according to news reports.

That’s why new technology applied to POCT must be validated the same way any other test is and assessed on a continual basis so that bias doesn’t intrude and make POCT less accurate, experts say. Regulations need to keep in step, too. “We have to follow CLIA requirements, but these same regulations must evolve as AI becomes more accepted—AI is not going away,” Tran said.

The evolving technology also needs to be accepted by the clinical laboratory community, which could take time. “In reality, yes, any tech can fail. But it’s going to come to a level of maturity and robustness with checks and balances and eventually work so seamlessly that you won’t notice anything,” Ozcan said. He pointed to mobile and online banking. “Twenty to 30 years ago, people would say, 'No way, I wouldn't trust online banking,'" he said. Now Americans routinely deposit checks via their smartphones.

AI and ML in POCT doesn’t need to be a threat to central laboratories either, said Tran. “I see this as less of a central laboratory versus POCT comparison, and more of a look at how the central laboratory and POCT exist together,” he said. “In the end, AI/ML is a tool, and it is here not to replace our most valuable resource—the humans—but to optimize these limited resources.”

Plus, AI and ML can’t do everything. While they analyze data quickly and automatically for things they were trained on, they “may not be able to predict things that fall outside of their training, hence the value of the laboratory professional remains,” Tran said.

Jen A. Miller is a freelance journalist who lives in Audubon, New Jersey. @byJenAMiller

Advertisement
Advertisement