Part one of this story described what prescriptive analytics is, where it fits into the analytics cycle and who is using it. Here in part two, learn how organizations are pioneering prescriptive analytics to launch a cycle of continual improvement.
By submitting your personal information, you agree that TechTarget and its partners may contact you regarding relevant content, products and special offers.
Whether the business or prescriptive analytics project is big or small, prescriptive analytics practitioners agree that it's imperative to understand the business process well enough to describe it accurately and to frame the business problem in need of addressing.
"I advise clients that they start with the business problem they are trying to solve and then find the relevant data," Gartner's Kart said. Some of that data may be in house; it may be necessary to purchase other data from outside parties, but the project starts with the problem.
For Tom Doub, CEO at the Nashville, Tenn.-based nonprofit Centerstone Research Institute (CRI), his work on prescriptive analytics was fueled by an acute business pain point. CRI is a research affiliate and IT resource for Centerstone of America, one of the nation's largest nonprofit providers of community-based care to people with mental-health issues.
"We were running through dollars faster than we could afford," Doub said.
Mental-health care has a tendency to "surround patients" with treatment, Doub said -- a psychiatrist, a therapist, group therapy sessions and home visits. "All of those pieces together can get expensive, and they really shouldn't be used unless they are ultimately necessary to get the clinical outcome that you need," he said.
"Getting people to understand the analytics life cycle and to trust the analytics is hard."
Martin Jonassen, head of business analytics, IUM Nordic
The approach is incentivized by the healthcare industry's traditional fee-for-service business model. As health care moves away from a fee-for-services model to one in which physicians are paid based on patient outcomes, however, the value equation for organizations like Centerstone becomes outcome-divided-by-cost. Faced with budget constraints and a needy population, the provider had to find ways to treat patients more efficiently without compromising their health.
"We developed an algorithm using data within our production environment. We measured outcomes and developed data for people coming in and said, 'For each individual coming in, what would be the optimal course of treatment for them?" said Doub, whose Ph.D. work was in clinical and quantitative psychiatry.
The algorithm, using SPSS software from IBM, prescribes optimal treatment plans for patients based on their diagnoses and characteristics. However, unlike in instances where the prescription might be automated -- a recommendation engine telling viewers what movies they might want to watch next -- the prescription in this case would be arbitrated by the human expert.
"The physician would always be the filter for that and could say, 'I have some other information here that is not in the algorithm that makes me want to go in a different direction,'" Doub said. "It's really about providing support for decision making, so they are not inclined to overtreat."
Top-down support crucial
"Inbound" prescriptive analytics projects solutions that automate how people do their work are always a challenge in his experience, said Mofibo Books' Jonassen, a longtime user of SAS Institute’s predictive and prescriptive analytics software. At the large broadband company where he was recruited to help mitigate churn, for example, the call center agents handling incoming customer calls had to pull off a tricky negotiation: They needed to address the issue of churn but the company didn't want them to plant the idea of defecting in the mind of a customer who wasn't already thinking about it.
The analytics challenge was to identify customers in real time who might churn and -- based on the customer's profile and past behavior -- create personalized scripts designed to ward off that behavior and ideally sell the potential churner even more services -- a Web TV on top of Broadband, say, or a music service.
Implementing the technical solution was no mean feat, Jonassen said, starting with the collection of all the available churn data across the company. "There's always a lot more data" than initially surfaces. The data had to be batched in real time to feed the predictions engine. The analytics had to be embedded in Salesforce.com, the call center agents' business application.
Even so, the technology, he insists, was the easy part. "Once the machinery and the modeling is up and going, it runs by itself," Jonassen said. "Getting people to understand the analytics lifecycle and to trust the analytics is hard."
Prescriptions have consequences
If a prescriptive strategy is taken, inevitably there will be consequences that could alter the future in a way that requires additional prescriptive analytics.
CRI's Doub is in the midst of a project that could revolutionize patient care. His group is now working on a project that takes the initial prediction model "several steps further down the road" by looking at patients' treatments and progress over time in an effort to pinpoint when a patient starts to get better.
"We are using machine learning strategies and technologies to model that process and simulate the thought process of a physician and the course of healthcare delivery," Doub said. So, rather than serving up a one-time decision about which direction to go in for a patient's care, the team set out to model the decisions made across the course of an individual's treatment and optimize the sequence depending on how the patient responds along the way.
"What we found, particularly in a traditional fee-for-service environment -- those incentives are pretty strong -- was that people tended to get treated past the point where they were actually getting better," Doub said.
Slowing down or stopping treatment at that point could obviously save money, as would changing the treatment strategy for a patient who is not making any progress. The early results are promising. Using actual clinical data, the researchers found they could increase the outcomes by 42%, relative to the baseline outcomes from treatment as usual, while at the same time reduce the cost of the work and effort required to achieve those outcomes by 58%. "It gives you a little bit of a glimpse to the potential waste and opportunity to improve that exists in the traditional healthcare landscape," Doub said.
The group has gotten a grant to do more research on the algorithm to validate that it works with different patient populations, he said. "But the other piece we're looking at is how do we take this to market? This is an algorithm that would appear to have some real utility in the healthcare environment, so we're looking for partners who are anxious to test these things in different clinical settings."
You heard it here.