SUNNYVALE, Calif. (May. 13, 2019) – Banner Health, a health system based in Phoenix, Arizona, wanted to focus its resources on the patients who would most benefit from early interventions that would reduce future utilization and improve outcomes.
The tricky part, of course, was predicting which patients were most likely to need those interventions, and the cost savings the interventions were likely to generate.
Traditionally, Banner Health relied upon retrospective claims data to inform this analysis. Conventional wisdom has held that the same 5 percent of the population that generated the most utilization and cost one year would be the same 5 percent that would do so the following year. Recent advances in the sophistication of healthcare data analytics, some experts said, have demonstrated the fallacy of this wisdom.
“We recognized that there was room for financial improvement,” said Barrie Bradly, senior director, data science, analytics and reporting, at Banner Health. “Despite previous efforts, we could not seem to make a difference – basically could not bend the needle. From the medical literature and from our own observations, we could see that a relatively small portion of our membership accounted for a high portion of the costs.”
The health system was unable to develop any insights into how it might better identify where to focus efforts for improvement. Without insights, performance could be seen to be driven by chance. But the health system is in a risk-based business for which it is staking its future (at least the potential for future results) on risky things (health).
“Surely a risk-based business and chance-driven insights don’t belong together,” Bradly said. “We needed to either figure out how to develop insights or get out of the risk side of the business. We needed to identify where to find the risk. Historically we believed that the concentration of costs – risk – from a prior year repeated itself in the succeeding year.”
This belief was acted upon by concentrating efforts on those patients who bore the most cost in the prior year. Specifically, Banner Health developed a first-stage algorithm, a decision tree, to identify the highest 5 percent risk in the population. In practice, care managers would then concentrate on the members in that 5 percent list.
“We concentrated efforts on communication; education; filling gaps in care; scheduling screenings, labs and other diagnostic-type tests; and more,” Bradly said. “Our team, the advanced analytics team of data scientists, became involved in the effort at this stage. We did not on the surface accept that last year’s top 5 percent would necessarily be this year’s 5 percent. We conducted a retrospective study and determined that looking year-over-year within our common membership, only 1 percent of the 5 percent top costs remained in the top 5 percent bracket the following year.”
So, only a fifth of all of those that Banner Health had thought would be high-cost ended up being high-cost. If the practice of population health management includes acting today to forestall less healthy outcomes later, then Banner Health should be focusing first on those that are headed toward high-cost, not those that have been high-cost.
“This is a manifestly different approach than traditionally practiced,” Bradly said. “We needed to develop an algorithm that could stratify our membership based on predicted future risk independently of historical costs, though of course costs could contribute to the prediction, just not be the driver of the prediction.”
The team, though knowledgeable, did not have the tools or advanced skills to build such an algorithm. Moreover, the team’s vision was to develop a self-service model – suggesting a software platform that could be used by different individuals for different needs, all related to population health management.
To accomplish this, Banner Health turned to BaseHealth, which specializes in population health management. Banner Health began using the BaseHealth Analytics Platform + Interventional Analytics Solution.
There is a variety of vendors of population health analytics on the market today. Some of these vendors include Conifer Health Solutions, Lightbeam Health Solutions, Lumeris, Orion Health, Relias, Sisense and ZeOmega.
MEETING THE CHALLENGE
Banner Health spent several months with the BaseHealth product development team to customize and validate its approach. Along the way, the health system regularly verified the underlying science by comparing what would the BaseHealth model have predicted versus what actually happened.
Banner Health provided to BaseHealth two years’ of deidentified data on about 100,000 patients. BaseHealth produced a stratified listing of expected costs for the entire membership based on these two years of data to predict the actual costs in the third most current year. Once complete, Banner Health provided to BaseHealth the actual third year data results and then compared how accurately the BaseHealth model predicted future outcomes.
The accuracy results were “a very high (good) c-statistic,” a common statistical measure for such information, Bradly explained.
“According to medical literature, traditional models, such as those used by all others in the industry that we know of, varied between a c-statistic of 0.65 and 0.85, with a concentration around 0.78 or so,” he added. “Incidentally, these seemed to be for studies with quite controlled conditions – such as more regular, clean data than one might otherwise encounter – so perhaps these results are biased to a higher level of accuracy than one would actually encounter.”
The BaseHealth blinded study consistently produced a c-statistic of about 0.9, a much better, more accurate, prediction than other traditional approaches, he stated.
“Once developed and tested, our analysts managed the platform and delivered to the care management team regular membership lists of which members to concentrate on, for what reasons and with what recommended action,” he said. “After using this solution for several months, other opportunities became evident, such as using the powerful clinical disease prediction algorithms and a special risk adjustment engine that BaseHealth had developed to assess RAF aspect on a per member basis for our Medicare Advantage plan contracts.”
Banner Health has identified tens of millions of dollars in potential savings in the area of cost avoidance through intervention – those tens of millions of dollars match the health system’s capacity of handling a care management load of about 2,500 members over the year, so represents just the top stratified 2,500 out of 100,000 Medicare Advantage members.
“We successfully integrated the platform with our FTP data feed with secure cloud-based Banner access to become an entirely self-serve model,” Bradly explained. “We successfully integrated with another third-party vendor, Holon, for an improved high-synergy solution. We identified two entirely new opportunities not originally contemplated representing incremental revenue/cost avoidance lines over our starting ROI expectations.”
Success metrics include the aforementioned 1 percent to 5 percent; the aforementioned blinded starting study; and the RAF insights that identified a further potential savings/incremental or accelerated revenue of tens of millions of dollars.
ADVICE FOR OTHERS
“Consider the implementation capabilities along with the solution capabilities,” Bradly advised. “We determined that we had not adequately positioned our team to optimally use the platform when we delivered it. To that end, we would recommend including an implementation team alongside the product development team.”
As an example, had Banner Health done that, it could have provided to the care management team the core output from the platform much earlier on while the team developed other capabilities that the care management team would not be able to use for some time, anyway.
For the full article, please go to the Healthcare IT News.
Population health analytics help Banner Health better predict high-cost patients
Better predict high-cost patients