Improving Hospital Hand Hygiene: The Solution May Lie with AI
A pilot test inside Lucile Packard is a joint project of Stanford Medicine and Stanford Engineering
August 12, 2020 | By Laurie Flynn
Hand hygiene of hospital staff is notoriously difficult to monitor, yet few behaviors are as essential to patient health. A trial inside Lucile Packard Children’s Hospital Stanford has revealed that an algorithm using artificial intelligence is at least as good as human observation for detecting use of hand-washing stations. It could be even better.
Researchers from Stanford’s Partnership in AI-Assisted Care (PAC) installed sixteen depth sensors in the ceiling above wall-mounted hand hygiene dispensers outside patient rooms. The system captures three-dimensional silhouettes of health care workers based on their distance from the sensor and, using AI computer vision, accurately determines whether they use the handwashing dispenser when they enter and exit the room.
The study revealed that the AI system was at least as accurate as human observation. Some hospitals use the “secret shopper” approach, where they dispatch individuals to covertly observe hand-washing
compliance, while most rely on individuals positioned nearby for long stretches of time. But such human-reliant approaches have obvious disadvantages. Neither approach is capable of observing when health care workers forget, for example, if they get sidetracked after they’ve washed their hands, or when a large medical team enters a patient’s room at the same time.
Perhaps its most obvious benefit is that the AI-based system being tested at Lucile Packard is on alert around the clock. “It doesn’t get tired or distracted,” said Amit Singh, MD, a Clinical Associate Professor of Pediatrics at Stanford with a deep belief in the potential for information technology to improve care. “It’s like a security guard that never takes a break.”
The trial at Lucile Packard is the first of its kind to systematically compare human observation of health care hand hygiene with an AI-based surveillance system using depth sensors.
PAC is a joint project of Stanford Engineering’s AI Lab and the Clinical Excellence Research Center (CERC) that is studying ways to apply AI technologies such as computer vision and machine learning to improving the quality of health care delivery while reducing its cost. The results of the observation, which took place between March and August 2017, were published in the July 26, 2020, issue of the Journal of the American Medical Informatics Association.
If you consider that proper hand hygiene is critical to preventing hospital-acquired infections, the impact of using AI to improve compliance could be huge. The Center for Disease Control estimates that 1 in 31 hospitalized patients acquires at least one health care-associated infection every day. But Singh is quick to point out that while poor hand hygiene often takes the rap for infections that patients contract in hospitals, it is very often not the cause. Still, identifying a site’s actual degree of compliance with hand hygiene protocols can help hospitals more accurately identify the real causes of infections and address them, he said.
Today, the most accurate method of monitoring handwashing is simple human observation, an approach considered “subjective, expensive, and discontinuous,” the paper concluded. “Consequently, only a small fraction of hand hygiene compliance is observed.” Other technologies have been tested in hospitals but each one has considerable disadvantages. Trials involving radio frequency ID have revealed the technology is neither as accurate as human observation nor as efficient as other options. Video cameras are the most obvious, but they require that other people review the recordings, raising privacy concerns and requiring additional labor. The depth sensors do not reveal the identities of those being observed, and the analysis is completely automated. “You can’t outsmart it,” Singh said.
The trial at Lucile Packard is the first of its kind to systematically compare human observation of health care hand hygiene with an AI-based surveillance system using depth sensors. Once the sensors were installed, the PAC research team then trained a machine learning algorithm to detect hand hygiene dispenser use in the collected images. The algorithm’s accuracy was then compared with simultaneous in-person observations of people using the hand hygiene dispensers, and the concordance rate was calculated.
The study is a critical step in PAC’s development of novel ways to apply AI-based monitoring to improving patient care and safety, but it’s just the beginning. “The potential is really limitless when you are monitoring something continuously and automatically,” Singh said. “It’s time for hospitals to start investing in autonomous monitoring and discovering the benefits.”
“Automatic detection of hand hygiene using computer vision technology”
Amit Singh, Albert Haque, Alexandre Alahi, Serena Yeung, Michelle Guo, Jill R Glassman, William Beninati, Terry Platchek, Li Fei-Fei, Arnold Milstein
Journal of the American Medical Informatics Association, ocaa115, https://doi.org/10.1093/jamia/ocaa115