Skip to content Skip to navigation

Moving Beyond Predictive to Prescriptive Analytics

August 31, 2015
by David Raths
| Reprints
UW Tacoma, MultiCare prototype ‘readmission score as-a-service’
Center for Data Science team. Photo courtesy University of Washington Tacoma

A partnership between data scientists at the University of Washington Tacoma (UWT) and MultiCare Health System, the South Puget Sound’s largest hospital and healthcare system, is fine-tuning algorithms to better predict which chronic heart failure (CHF) patients would be most susceptible to readmission within 30 days, with the goal of understanding which adjustments to treatment plans have the most impact on preventing readmissions.

David Hazel, managing director of the Center for Data Science at UWT, said the partnership engaged the strengths of both organizations. MultiCare had been working to bring down readmission rates for patients with CHF. “What they were doing wasn’t working to bring down rates as fast as they wanted to,” he said. “We have a strong team of data scientists who understand how to look at healthcare data and run experiments. They have a really strong clinical team.”

He said that in developing the models, it was key that the UWT researchers had access to MultiCare cardiologists. “We could go there once a week and do rounds with the cardiologists and get feedback on types of data points we were seeing and thought might be relevant and get their clinical input,” Hazel said.

The researchers started with the data in MultiCare’s data warehouse and also got access to the state of Washington’s claims data on all inpatient CHF patients in the state. “We have to understand what is happening to patients in the hospital setting as well as what is getting billed and reported,” Hazel said. Once the models were developed, they had cardiologists at Northwestern University, UW Medicine and MultiCare study them and they did a retrospective analysis in which the solution was deployed against all of MultiCare’s heart failure patients for the last three years to see if its predictions for which patients would readmit meshed up with the patients that actually came back. He said the algorithm has proven very good at predicting which patients were traditionally ranked as medium risk, but readmitted, and by flagging those patients, clinicians can apply some more resources and attention and have an impact.

To advance the work with the algorithm, Hazel and his team are launching a new spinout company called KenSci Inc. Pilot projects are being run at MultiCare Health System, UW Medicine & Madigan Army Medical Center.

Their goal is to create a “readmission score as-a-service” that would enable providers to use the bank of predictive models for many chronic conditions, not just heart failure.

The current project status is prototype, he said. “We are working on physician validation in a clinical setting. Once we are able to show that it works well and gets the kinks out, then the next step would be a full-on EHR implementation.”

Hazel said the tool pulls data from MultiCare’s data warehouse and from HL7 feeds from Epic and other EHRs, comes up with the insights and then publishes those using an application programming interface that can be pulled back into the cardiovascular dashboard inside Epic, or the care team flow sheet, or another external dashboard. “The way the tools are oriented in Epic vs. the other reporting and analysis tools that the hospital uses is completely different,” Hazel said, “so you can have the same underlying insights and surface them up into multiple experiences. Then you can have the information the clinicians need in the work flow that they are used to when they are dealing with the patient.”

He added that getting the relevant data out of the EHR was a huge challenge. “It turns out that a lot of the attributes we needed for the models were not contained in discrete fields in Epic at the time we started, just in the notes,” he explained. “We could have developed a series of protocols that would do natural language processing and extract relevant attributes out of those free-text fields,” he said. The other approach was to convince MultiCare that it should alter its Epic implementation to start storing these discrete values so they are easier to track. “Because every Epic installation is different, those customizations are pretty standard. It is just a matter of getting them prioritized,” he said. “We got buy-in from senior leadership to do that. And it is helpful for them, too, because if they are trying to run their own reports on certain things, if it is just in the notes, it is very hard to do any type of analysis on it but if they have a discrete field, it is easier.”

Hazel noted that one of the key points about readmission is that most of the drivers don’t happen in the hospital setting. They involve complications, or whether or not patients adhere to post-discharge instructions. Because the providers working with patients tend to be siloed, there are gaps in insight. His research team has worked to develop a “Risk-o Meter” to expose complex models to a layperson so they can understand that if they are compliant, it has a huge impact on the likelihood of re-admitting.

He said there are also ways to tie more data sources together to give clinicians a better 360-degree view of the patient. “If a risk is trending up, they can start targeting interventions to reduce risk at that point vs. waiting for them to come back for a follow-up appointment,” he said. “We are moving beyond predictive to prescriptive — not only identifying what is going to happen, but how you can change what you are doing to influence that future outcome.”