Northwell Health uses machine learning to reduce readmissions by nearly 24%

Photo: Northwell Health

Reducing readmissions is a major focus for healthcare organizations operating under value-based care contracts.

Clinicians at Northwell Health, the largest healthcare provider in New York State, are applying clinical artificial intelligence to augment their post-discharge workflows and have reduced readmissions by 23.6%. The clinicians studied clinical AI stratified patients for their risk of readmissions, identified the clinical and nonclinical factors driving their risk, and recommended targeted outreach and interventions to reduce patient risk.

Clinical AI versus predictive analytics

The clinicians noted the contrast between prescriptive clinical AI and traditional predictive analytics, and their impacts on patient outcomes.

“Predictive analytics as a whole is a powerful tool using a combination of historical data, statistical modeling, data mining and machine learning in order to predict events and identify patterns,” said Dr. Zenobia Brown, vice president and medical director at Northwell Health, a health system based in Manhasset, New York.

“Despite those powerful insights, predictive analytics are really just a starting place in terms of enacting meaningful change at the population and individual levels.

“Prescriptive analytics, a tool that uses predictive modeling to make specific recommendations across a matrix of potential decision points, adds the ability to operationalize the information given which is key,” she continued. “When orienting clinical teams to prescriptive analytics, I liken it to how we as providers make recommendations based on our understanding of the clinical data and our experience over time, which [lead] us to the ‘right clinical decision.'”

Clinical staff members accept, and the data would support, that the more experienced one is – the more historical information staff has about the pattern of outcomes, given a certain set of circumstances and intervention – the better the outcomes, she explained.

A million different patients

“I ask my teams to imagine how much better their decision-making would be if they had one million times the experiences in that set of clinical data, and the experience of treating the disease one million different ways in a million different types of patients,” Brown said. “This is what prescriptive analytics supports; a way to make decisions in managing the complexity represented by patients beyond the data set that is limited by the human brain.”

The technology supports a hyper-informed recommendation based on a complex matrix of data points specific to achieving the desired outcomes.

“It’s a really exciting time in healthcare right now when it is widely accepted that the factors that influence the overall health of people extend way beyond the strictly clinical risk,” Brown said. “Many believe that social determinants are equally if not more impactful on the overall clinical outcomes.

“We had a really interesting case of a cardiac patient who was in the healthcare field,” she continued. “While diet was discussed as part of his routine care, based on his high education level and clinical background, this would not have been identified as a high-risk area. As it turned out, this particular patient had social isolation, living in a food desert, as well as other nonclinical factors that cause the prescriptive AI to recommend multiple nutrition interventions.”

A gaping hole in self-management

When the recommendation first appeared, the care navigator was perplexed, but when she contacted the patient, she in fact found that this was a gaping hole in the patient’s self-management and ability to recover successfully from surgery. In the clinical domain, typically staff looks at historical utilization, disease severity and acuity to determine the risk.

“In terms of the more typical clinical risk factors, AI-driven recommendations contribute a deeper understanding of the most likely intervention to impact the outcome,” Brown said. “In this example, what has been fascinating is that the order of recommended interventions might be unexpected.

“For instance, in a typical heart failure patient, we would typically prioritize medication reconciliation, education about daily weights, etc., to mitigate the risk of a CHF readmission/exacerbation,” she continued. “In one heart failure case that comes to mind, the AI recommended a nephrology consult as the first most important intervention to accomplish.”

The team might have gotten to a nephrology consult over the course of the patient care plan, but probably not as the first thing, and probably not in time to prevent a readmission, she added.

“Medical providers and people in general are very good at recognizing the patterns with which we are familiar,” she noted. “It’s the ones we don’t recognize, don’t see and can’t prioritize that represent the opportunities to keep patients on the path to wellness.”

Integrating into the clinical workflow

So how does clinical AI integrate into the clinical workflow to augment transitions of care and prevent readmissions post-discharge?

“The first, most important step is for the providers of care to be confident in the technology,” Brown stated. “If they don’t believe it works, or don’t see the value in how it helps their time or helps the patient, there is zero chance of good operational integration. In our case, we had a mature transitional program that was already seeing good outcomes, so it was even harder to convince providers that this would be additive.

“Having said that, an important part of the journey was sharing these cases of patterns that otherwise would have been missed; the ‘good catches,'” she continued. “This reinforced the value of the tool. Also important was making sure the predictions and recommendations were timely, such that the team had appropriate lead time to impact each patient.”

For the team, that meant that the AI/predictive modeling tool was being refreshed multiple times per day, while the patients were still in the hospital, so that the identification of the high-risk patients could happen as far upstream as possible.

“It also allowed for interventions to occur in the hospital that might be more difficult or less timely in the ambulatory setting – specialty consults particularly,” she said. “In terms of how it integrates into the workflow, it’s like another vital sign or lab report. It’s an additional piece of data or information that can be used to connect with the patients in meaningful ways. It does not replace what happens in that provider/navigator/patient relationship, but it can enhance the interactions.”

Brown will offer more detail during her HIMSS21 session, “Applying Clinical AI to Reduce Readmissions by More Than 20%.” It’s scheduled for August 11, from 4:15 to 5:15 p.m., in Venetian Murano 3201A.

Twitter: @SiwickiHealthIT
Email the writer: [email protected]
Healthcare IT News is a HIMSS Media publication.

Source: Read Full Article