When Dr. Jonathan Michel partnered with a pediatric neurologist on an analytics-heavy endeavor, he figured everything in healthcare would be automated and digitized in short order. The actionable information gleaned was just too valuable and powerful.
But that was almost 30 years ago. It's only now that Michel, director of Data Science at the University of Virginia Health System in Charlottesville, Virginia sees the healthcare industry catching up to where he thought it would be in the early 1990s. Adoption of digitization and automation in healthcare has generally lagged behind other industries.
Michel's colleague, Dr. Bommae Kim, the health system's senior data scientist, said there's a higher barrier to adoption in healthcare than in other sectors.
"Users have a very wide range of different levels of understanding in data or science," said Kim. "Healthcare is very labor intensive. It's not like a tech company. Being labor intensive is at the very core of healthcare services. Nurses, case workers, care coordinators -- you've got a very wide range of different people."
On top of that, digitization in healthcare has made only a sluggish process. A lot of work has traditionally been done on paper, and when electronic health records were brought into the equation, that created some barriers as well. It reduced clinicians' efficiency because they were often dealing with EHRs that were ill-suited to their workflows.
"That was causing inefficiencies," Michel said. "I'm not sure that's all gone at this point, but it's gotten better. The result of that is in the late 2000-aughts, when it became more widely adopted, now we're seeing the benefits of that and the data that flows out of them."
Now, for the past three years, Kim and Michel have been closely engaged with the general medicine department in an effort to use digitization and automation to slash 30-day readmission rates. The work was important both from a clinical and a financial perspective, because the Centers for Medicare and Medicaid Services uses a 30-day readmission rate as one of the metrics by which reimbursement levels are determined.
By building a strong relationship with a core set of clinicians who have implemented the data driven analysis and modeling employed by the health system, the system has seen a 2% absolute drop in readmissions that has held steady since implementation. That translates to a 15% relative reduction amongst the overall patient population.
Doing everything in-house, the system was able to employ new technology and processes to develop machine learning algorithms that are very accurate in predicting readmission risk. By determining high or low risk, clinicians can select an appropriate intervention track.
The changes took some communicating, since staff wasn't used to this type of analysis. But Kim knew from the start that communication would be part of the process. It's a necessary part of shepherding in a widespread culture change.
"Bommae did a really nice job of walking them through the descriptive results, the predictive results, and the vocabulary we needed to speak to them in a way that resonated with them. Once we did that it was like, 'OK, we got it. We understand what the value of this is, and we trust that it's unbiased, and we trust that it's not going to lead us in the wrong direction.'"
It took about five years to develop the predictive analytics model, and it has gained a lot of adoption in that time, which makes it easier to implement into clinicians' workflows. The model gives clinicians a risk level for a patient, and while clinical results were unsatisfactory when the risk level was manually lowered, the results were successful when they were limited to only raising the risk level if they deemed it appropriate.
There are a couple of reasons for this.
"Clinicians tend to underestimate patient risk -- that's one thing," said Kim. "But another reason is that data always has limitations. Our data is limited to our patients. What that means is, if we get patients from another hospital … we don't have their data. Because we don't have data, the model will just predict the risk for that patient is low, but that's not true. Patients can have better estimates."
That one tweak -- only allowing clinicians to upgrade risk -- made all the difference, and now Kim and Michel are spreading the word about how providers can implement a similar approach.
They'll next speak about the issue at the HIMSS20 conference in Orlando, Florida in a session entitled "Making Prescriptive Analytics Work for Clinicians," Wednesday, March 11 at 1 p.m. in Room W304A at the Orange County Convention Center.