AI, machine learning can drive better hospice utilization


David Klebonis, chief operating officer of Palm Beach Accountable Care Organization, speaks at the HIMSS22 conference in Orlando.

Photo: Jeff Lagasse/Healthcare Finance News

ORLANDO – More than 1.5 million Medicare beneficiaries were enrolled in hospice care for at least one day during 2018, a 17% jump in about four years. While hospice care is useful and compassionate, especially when focusing on quality of life for terminally ill patients, there’s a problem that looms: At least 14% of Medicare recipients enrolled in hospice stayed there for more than 180 days.

Hospice stays beyond six months can result in substantial excess costs to healthcare organizations under value-based care arrangements. Clearly, something needs to change.

That was the message delivered by David Klebonis, chief operating officer of Palm Beach Accountable Care Organization, during his session, “Driving Appropriate Hospice Utilization With Explainable AI,” at the HIMSS22 conference in Orlando.

“Humans are just really bad at determining when other human beings are going to die,” he said. “This includes the most trained physician experts.”

On the flip side, 21.9% of hospice episodes last between one and seven days. That’s a problem as well: The industry considers both long and short stays as failed prognoses. … Right off the bat, evidence shows we fail this process 42% of the time. That’s what gravitated us toward this program. It’s a problem a lot of people seem to be having difficulty with.”

The program Klebonis is referring to is his team’s efforts to develop interpretable machine learning models that can predict hospital overstays to drive appropriate hospice referrals. It’s a timely initiative given the fact that hospice use continues to grow.

“About 1.5 million Medicare beneficiaries enroll in hospice every year,” he said. “Of the Medicare patients that die, only 50% are on hospice. We have this great service, we know outcomes are better on it, yet 49% slip through the cracks and they don’t get a hospice referral before death.

“Every time we fail on determining a prognosis on the back end, the patient is seven times more expensive than the patient you made the right decision on,” said Klebonis. “Seventy-two percent of all hospice costs come from patients with greater than 180 days length of stay.”

To address the issue with AI and machine learning, PBACO worked with its vendor partner to define when interventions should occur; that’s what the model was built around. It made sense for one of the intervention points to be at the point of referral, so if it was a PBACO physician making the referral, patients were put through the referral engine. Another intervention typically takes place after a patient graduates to a longer length of stay, at which point a “re-review,” as Klebonis put it, gives the referring provider additional information so they can make better decisions moving forward.

“It’s important that you don’t just give a provider a decision,” said Klebonis. “If you tell a doctor, ‘This computer said you should change this,’ you’re going to have low-percentage adoption. We spent a lot of time on training, on physicians agreeing with the model and having confidence. The idea is that when we produce something that’s going to spur an intervention, it’ll be in a language the physician speaks, and we’ll speak very specifically about that patient. That doctor will agree, and you’ll have a better chance at changing their behavior. You’ve got to build trust with physicians.”

The machine learning program trains its algorithms on locally representative populations, using all data sources available. Ownership of the predictive models helps to monitor accuracy and identify anomalies, and the models can be retrained as necessary. Importantly, explainability is built in at the level of individual risk protections, and the machine learning competency is built in-house.

That, said Klebonis, was a better option than choosing a generic AI and machine learning platform. Custom software built in PBACO’s data is more accurate and explainable.

“Ultimately the goal of machine learning is to bring together components and be able to create a list for your interventions,” he said. “The components are what populations you are going to define. Each model we built was significantly different. We wanted to remove noise, remove points that won’t define outcomes.”

As for the ROI, PBACO was able to facilitate a 29% reduction in long hospice stays, with a cost savings of about $47,000 per patient – good for about $2.1 million in annual savings.

“This has been a very helpful project,” said Klebonis.

Twitter: @JELagasse
Email the writer: [email protected]


Source link