How AI Is Preventing Fatal and Expensive Hospital Falls

By Pragati Verma, Contributor

In December 2014, El Camino Hospital, a nonprofit health system located in Silicon Valley, California, linked patients’ bed alarms, nurse call-light data, and medication records to a machine learning and predictive analytics system to identify patients who were at high risk of falling. When the patient got to a certain threshold, the system would alert a nurse, reminding them to check on the patient for the next 12 hours.

Implementation of the technology was a response to a pressing issue. “About half of our patients were at high risk of falling,” Cheryl Reinking, chief nursing officer at El Camino Hospital, explained, “so fall prevention became a patient safety focus area for us.”

Today, the results of this technological solution are evident. According to Ian Christopher, co-founder and chief technology officer of Qventus, the healthcare technology startup that created the decision management platform to predict and prevent falls at El Camino, “the hospital data shows, fall rates have reduced by 29 percent.”

Attacking the Root Problem

According to the Department of Health and Human Services’ Agency for Healthcare Research and Quality, between 700,000 and a million people fall in hospitals every year. In addition to being dangerous, these falls can be costly: Centers for Disease Control and Prevention (CDC) estimates an average cost of $30,000 per patient for an injury caused by a single fall.

Before working with Qventus, Reinking recalled that El Camino Hospital had invested time, money, and staff in fall-prevention efforts. In addition to installing a call-light system with bed alarms, the hospital used sitters—volunteers who sit with patients to provide patient support and alert medical staff when they see a problem—to monitor patients. When these options didn’t successfully prevent falls, the hospital realized it needed a more proactive approach and decided to re-evaluate where they went wrong.

Traditional fall prevention methodology, according to Christopher, can be overly simplistic. For example, Morse Fall Scale—a commonly used method to assess a patient’s likelihood of falling—simply scores patients on six parameters and gives a static assessment that does not capture changes in the patient’s condition or treatment.

“It can be a very noisy measure and might end up tagging too many patients as highly likely to fall,” he explained. “Instead of marking 30 to 40 percent of the patients as likely to fall, we are able to identify the top 5 to 10 percent who need more care.”

“Data in healthcare is not the most pristine, clean, and beautiful thing in the world.”

—Ian Christopher, co-founder and chief technology officer, Qventus

For Christopher, the ability to prioritize at-risk patients and bring the focus to those who need attention is rooted in AI technology. The software, developed by Qventus, pulls data from electronic health records, then looks at nurse call-light and bed-alarm data, and finally combines it with other real-time information, such as medication and vitals, recorded by a nurse.

When certain data elements line up, such as patients that set off the bed alarm and call light more than a certain threshold—which can vary depending on the ward, age, and medications given to a patient—it sets off a trigger. The system then sends out an alert to the nursing station, identifying that a particular patient is at a high risk of falling for the next 12 hours.

Yet this data-driven software, Christopher explained, wasn’t easy to develop. The first challenge was to extract data from health records and hospital call-light systems. “Getting data out of hospital systems can be incredibly challenging,” he said, commenting on how they had to “do a lot of manual work.”

What’s more, for this project, there were no APIs to help software applications talk to each other without user intervention. Once they could access the data, Christopher and his team had to comb through the dataset and clean it. “Data in healthcare is not the most pristine, clean, and beautiful thing in the world,” he lamented.

The next step was to build and train a model that would predict the likelihood of a patient falling, but the ability to make meaningful predictions came with its own set of complications.

“We had to pay a lot of attention to underlying factors and ensure that it didn’t pick up any frivolous information,” Christopher said. “For instance, we didn’t want a model that would predict something like everyone who had a pain medication in unit 12 and called at 10 am is likely to fall down in the next 12 hours. Our predictions needed to be more specific and meaningful.”

Turning Data into Action

Since the Qventus team was delegating complex decisions to the AI algorithm, they worked hard to ensure that there was no bias in how the AI software made predictions. Typically, bias can creep into machine learning algorithms from parameters unconsciously designed to promote a certain outcome or match some preconceived notions, such as older people are more likely to fall. There can also be limitations in datasets, such as overrepresentation of certain populations.

“We began by developing a model internally and went back to the drawing board after talking to people actually working at the hospital,” he recalled. “It took a few iterations before we got the prediction process right.”

It’s this iterative process of building models and aligning them with the real problems people face that can transform organizations like El Camino Hospital. Or, as Christopher puts it, it’s this process that “turns data into action.”

As for Qventus, which received $30 million in a Series B funding in a round led by Bessemer Ventures Partners in May, it is time to expand their platform. “We are exploring a number of solutions, from streamlining patient flow to improving pharmacy management and building staff schedules, to helping simplify how hospitals operate and relieve operational bottlenecks and challenges in emergency departments, perioperative areas and the pharmacy,” Christopher explained. In the end, of course, it comes down to how technology can best serve the people in need.