SEATTLE — Computer scientists and doctors at the University of Washington have found a way for artificial intelligence to deter a concerning incidence in healthcare: medication errors.
Errors in administering injectable medicine in particular, account for an estimated 1.2 million hospitalizations a year in the U.S., according to a study published in the American Health & Benefits journal.
UW researchers say they've found a way to flag such errors using AI.
As AI becomes increasingly ingrained in our lives, their research demonstrates it could play an important safety component in the ICU or an operating room.
"The goal is really for patient safety," said Dr. Kelly Michaelsen, an associate professor of anesthesiology at UW Medicine.
The study involved video recorded from an AI camera worn on the head of a healthcare provider as they prepare a syringe for medication.
In one video example, the powerful anesthetic known as propofol is shown properly labeled, as detected by the AI system by displaying text flashes.
In another staged example using different medication to demonstrate a label mismatch, the AI system provided test flashes showing the discrepancy. It flagged the error in real time.
"We know that the medication from the vial to the syringe, it's possible that things could get switched up, and the patient could get the wrong drug. This is important because maybe they get a drug that they're allergic to," Michaelsen said.
Mismatches in medication could also lead to dangerous complications.
Shyam Gollakota, a professor at UW's Paul G. Allen School of Computer Science & Engineering, co-authored the study with Michaelsen.
He thought about testing AI in this scenario, while considering some grim numbers: an estimated 5 to 10 percent of all drugs given, are linked with errors, according to a study.
"So the question we ask is, can we have an additional pair of AI 'eyes' looking over and trying to do an additional check to make sure that the drugs are being given correctly and making sure that no errors, unintentional errors, have actually crept into the system," Gollakota said.
Their research, published in the NPJ Digital Medicine in October, showed their AI achieved 98.8 percent specificity at detecting vial-swap errors.
They tested the AI algorithm while recording staged mismatches, plus large amounts of video data gathered from operating rooms until eventually, the program was trained to find discrepancies.
"It ultimately uses all the visual cues of the vial or the syringe. So it looks at the color, the text, the orientation, the size," Michaelsen said.
Still, the researchers had to find a way for the AI to recognize labels even when they're partially covered by a hand or if the lighting is dim. Through deep learning, their AI program was indeed, able to find the mistakes, according to Gollakota.
"So the way to think about it is even, for example, you're buying cereal and even if half the words in the cereal are covered because of, let's say you're buying Cheerios and let's say 'Che' in the cereal are covered because someone has their hand on it. You can know the context of the colors around it; you basically know that it's Cheerios, right?" Gollakota said.
The camera system has been used for research and is not available in market.
There are no specific details on how the errors would alert the healthcare provider, whether it's an auditory alert like a beep, or a visual cue through special glasses, as an example.
The first step was to prove, the AI can work.
"Having another person standing there is kind of a cumbersome solution; using AI to do this automatically in the background I think, is a really neat solution," Michaelsen said.