From preventive care tools to advances in nurse scheduling, artificial intelligence innovations are accelerating in the health field, helping providers improve interventions and cut costs.
AI technologists are still working to combat certain barriers to progress, including a lackluster appetite by policymakers to chart development and overcoming public fears around replacing human-provided care with machine-learning interventions.
Deepesh Chandra, chief analytics officer of Cincinnati-based Bon Secours Mercy Health (BSMH), said a key to quelling public concerns around AI is demonstrating its usefulness.
“As an industry, we are now at a point where people are more open to AI because it can prove its value,” Chandra said. “Before, it was just a sales pitch.”
BSMH recently started applying AI models to predict patients potentially at risk of not receiving needed care. Such risk factors include past medical history, check-up no-shows, and hierarchical condition category (HCC) scores, a standard in the health field to calculate patient risk based on clinical and demographic variables.
The predictive technology has alerted BSMH to a few thousand patients who were generally “under the radar” or regarded as “higher-risk” for a significant adverse health event in the future, Chandra said.
“The signals weren’t that prominent for a human being to pick,” he said. “But the AI algorithm actually picked up those signals and highlighted it for a case manager to review.”
Each prevented emergency room visit translates to roughly $5,000-$15,000 in savings, he noted. Because BSMH’s AI team continues to hone these technologies and apply them more efficiently to more cases, the total amount of savings often compounds.
Chandra estimated AI’s full impact in the above case, conservatively, could be $25 million-$30 million annually.
Sioux Falls, S.D.-based Sanford Health started developing a predictive analytics nurse scheduling system called Leveraging Analytics to Mobilize (the workforce) and Prepare (for the future) (LAMP). The tool has unburdened scheduling managers of some of the painstaking manual work of tracking nurses’ hours and scheduling shifts.
For each scheduling manager who has used it, LAMP has decreased the time spent doing administrative tasks – like updating spreadsheets – by 40 hours per month, said Sanford Health Chief Nursing Officer Erica DeBoer.
DeBoer says LAMP has three main goals associated with streamlining the nurse scheduling process: Improving employee satisfaction; increasing operational efficiency through quickly predicting and shifting employees to meet staffing needs; and ensuring employee retention and adequate staffing levels.
Giving scheduling managers their time back “means that they could focus on mentoring, they could focus on supporting their frontline teams, and those things that are going to add more value to our patients,” DeBoer said.
A General Electric Healthcare innovation implemented by Cleveland-based University Hospitals helped deliver approximately 150 endotracheal tube placements during the height of the COVID-19 pandemic, according to a May 2021 University Hospitals press release.
Known as Critical Care Suite 2.0, the technology helped to assess where, exactly, ventilator tubes should be placed in patients when as many as 1 in 4 intubations had resulted in mispositioned endotracheal tubes. The technology also helped recognize critical patient conditions like a collapsed lung.
Since then, University Hospitals has gone a step further in developing AI to recognize worsening COVID-19 conditions in patients, integrating a number of vital signs to diagnose patients more accurately and predict who is getting the sickest, said University Hospitals Chief Quality and Clinical Transformation Officer Peter Pronovost.
This iteration of the Critical Care Suite is helping clinicians determine when COVID-19 in a patient might be accompanied by other conditions like sepsis and pneumonia, and guide therapy appropriately, Pronovost said.
“We have to link prediction to prescription if you’re really going to get value out of it, because just telling me someone’s sick just doesn’t necessarily add a whole lot of value,” he said.
Barriers to Developing AI in Health
Though AI innovations continue, AI technologists in the health field are wrestling certain financial and policy-related hindrances.
The AI market generally sells technology to the health industry on a “financially prohibitive” case-by-case basis, Pronovost said. Therefore, it behooves health companies to work with tech vendors that allow the companies to pay once for the data work so more iteration can be done on various use cases.
Vendors sell AI use cases such as nurse scheduling or operation room scheduling. For each use case, 90% of the overall purchase covers manipulation of data already held by the customer, Pronovost said. The other 10% covers the actual function that the AI performs.
“Because of the way the health systems are buying these, the cost is what’s limiting the number of places you use it, because if you’re spending $2 million or $3 million … every time I want a use case, it’s hard for a health system to go above 10 or 20 use cases,” Pronovost said. “You’re spending ridiculous amounts of money.”
University Hospitals has partnered with AI healthcare technology company Edgility, where University Hospitals pays for the data work only once, and therefore has more freedom to direct finances toward relevant use cases, Pronovost said.
Public Perception and Policymaking
Pronovost and Chandra noted that trust barriers around automation in health are steep.
It’s sometimes difficult to bridge the gap between innovation and the human aspect that good care requires, as illustrated by the 2019 example of a California hospital delivering end-of-life news to a 78-year-old patient via a machine.
Part of the reason for high trust barriers is that “healthcare is personal” with direct impacts on patients’ livelihoods, Chandra said.
Applying AI in important, care-adjacent health areas is pivotal to demonstrating AI’s utility in the field and building trust among the public, Chandra said.
“When it brings efficiency in the mix, when it brings how we can run things faster, cheaper, better,” he said, “all of that is a pretty significant adoption part.”
The Food and Drug Administration had approved 521 AI and machine-learning-enabled medical devices as of Oct. 5 and expects the trend to continue as the technology continues to advance, according to the agency’s website.
If the prediction holds up, it will spur public confidence about AI’s efficacy and safety in the health field, Chandra said.
“In the next five years, we are going to see that uptake happening more and more, but it is going to take a much longer cycle for it to reach what we accept as ‘everyday AI’ in patient care,” he said.
The time is ripe for policymakers to set guidelines for how AI developers should prove effectiveness and transparency in their health models, Pronovost said, recalling an interaction with a vendor who demonstrated an ultimately weak and faulty palliative care model.
Pronovost is currently working on a tip sheet to help health systems evaluate AI purchases, and he hopes it can serve as a blueprint for policymakers to craft future regulation.
“These are going to be health systems that don’t have big engineering departments that have to make decisions on how to use these,” Pronovost said. “Really practical tools are what the field needs, like a simple checklist that has the elements that you need to look at” when making buying decisions.
