At UPMC, Turbo-Charging Quality Improvement Efforts through Data Analytics
At a time when the leaders of patient care organizations are facing intensifying pressure to shift away from a dependence on volume-based payment and to plunge into value-based care delivery, some U.S. hospitals, medical groups, and health systems are helping to lead the way into a future of continuous clinical improvement and of clinical transformation. That topic—of organized continuous quality improvement—was the subject of the fourth-quarter 2018 Healthcare Informatics cover story. Numerous leaders of pioneering organizations were interviewed for their insights into the health system change focused-quality improvement movement that has been emerging across the U.S. healthcare system.
Among the leaders interviewed for that cover story was Oscar Marroquin, M.D., a practicing cardiologist and epidemiologist, who is helping to lead a team of clinical data experts at the vast, 40-hospital UPMC health system in Pittsburgh. Dr. Marroquin and his colleagues have been busy harnessing the power of creating and nurturing purpose-specific teams focused intensively on the management of data to power performance improvement, particularly in the clinical area. Marroquin’s team, of about 25 data specialists, was first created five years ago. Of those, half are IT- and infrastructure-focused, and, says Dr. Marroquin, “The rest are a team of folks dedicated to data consumption issues. So we have clinical analysts, data visualization specialists, and a team of data scientists who are applying the right tools and methods, spanning from traditional analytical techniques to advanced computational deep learning and everything in between. Our task is to use the clinical data, and derive insights”—and all 12 clinically focused data specialists report to him.
And that work—“allowing people to ask questions to generate opportunities”—has paid off handsomely. Among the advances has been the creation of a data model that predicts the chances of patients who are being discharged, being readmitted. The model, based on the retrospective analysis of one million discharges, is also helping case managers to more effectively prepare patients for discharge, specifically by ensuring that patients being discharged are promptly scheduled for follow-up visits with their primary care physicians. “If those patients are seen within 30 days of discharge,” he notes, “there’s a 50-percent reduction in their 30-day rate of readmission.” The program is now active in six UPMC hospitals.
Below are excerpts from the interview for that cover story that Healthcare Informatics Editor-in-Chief Mark Hagland conducted with Dr. Marroquin this summer.
From your perspective, what does it really mean to be data-driven, in the pursuit of continuous quality improvement and clinical transformation?
From my perspective, I’m very passionate in that I feel that you really can’t do any of the things that doing in terms of moving towards value, without having a robust data infrastructure, a robust strategy on how to use data and analyze it, and without then deriving evidence for how you’re going to transform your organization. There are a lot of buzzwords involved in all of this, but the only way to get from a buzzword to a true action of transformation, is if it’s data-driven.
I’m a cardiologist by training and an epidemiologist; I still practice. Over the past five or so years, I’ve been asked to oversee how we’ll derive insights from clinical data in our system; in other words, this work is around anything related to big-data analytics, with those analytics being used to help our clinicians. In order to do that, we’ve had to do many different things, including more intelligently aggregating our data, and focusing on specific analytical purposes. They’ve been structured as databases for transactional systems, but not with the intent of improvement.
So we’ve spent a lot of time creating a purpose-built environment for analytics. That’s involved a lot of technical work, to create tables, what we call our consumable layers, for analytical purposes. And we’ve created a team whose only job is to do analytics. We’ve had folks in the past managing back-end databases, who have generated reports, but that doesn’t lead to a sustainable way of using data. And so we have a team that is dedicated to maintaining the warehouse and consuming the data.
With regard to the team of 25 data analysts, do all of them report to you?
The 12 who do data consumption report directly to me; the others have a dotted-line report to me. They sit on our infrastructure team within our IT Information Services Division. Both teams are part of the Clinical Analytics Team. Data analytics—Health Services Division. Integrated team. Two sides of the same coin.
When did these teams come together?
There have been different phases. The analytics program development started in 2012, and we learned a lot of lessons. A lot of the work early on had to be dedicated to technical issues—identifying data sources, etc. That was a pretty labor-intensive process. We really got enough aggregated data to use it consistently in 2015, so from 2015 on, we’ve had this structure of teams dedicated to doing this as I’ve described.
Can you share a few examples of key advances that your team has made so far?
When asked what our team does, I tell folks we do work at the higher level in three different buckets. The first bucket is the entry point for the majority of projects. Not everybody in the system necessarily knows which questions to ask.
We allow people to ask questions to generate opportunities. Off of that, two things will happen. One, hypotheses can be generated, and so we can do hypothesis testing, we can do comparative effectiveness studies, we can formal testing of hypotheses. Also, when insights get generated, one can say, oh boy, there’s a lot of heterogeneity in this population, why is one group more at risk? So we can identify who is at high risk of a condition, and who within the high-risk category is at high risk of developing specific conditions? And the third level or bucket, we apply machine learning and AI tools to develop models that allow us to do a variety of things, from more precise phenotyping of our populations; we can build predictive models to identify patients at various levels of risk. And we also use these models to do unsupervised learning, where we can start to generate hypotheses. So most people in this space love to talk about the latter part, the predictive modeling, and we have done a fair amount of work there, with things like identifying patients at highest risk of rehospitalization after 7 or 30 days of discharge, and we’re using that in our hospitals to guide clinicians. There are resources everywhere.
So we developed a model derived out of retrospective analysis of one million discharges, and we’ve prospectively verified that the model allows us to identify patients at the highest risk of readmission, so our case managers can help us identify plans to help those patients transition from hospitalization to post-acute care in a more effective way. And we see that if they’re seen within 30 days of discharge, there’s a 50-percent reduction in their 30-day hospitalization if they get in to see a clinician. So we make sure that the patient has an appointment made and is ready to see their doctor once they’re discharged, to address any issues.
When was that program put into place?
We spent a lot of last year validating the data. And then this year, we started rolling this out to our hospitals in a phased approach, so throughout 2018, we’ve been deploying this to our hospitals, and we’ve trained and educated different hospitals to use the model, and we’re actively following patients to measure the impact of the tool. And as an epidemiologist, I’m always cautious about declaring victory too soon. We’re seeing good trends, our smaller hospitals are seeing decreases in patients coming back early.
So you don’t have any metrics to share yet?
We have three hospitals, smaller ones we started the program with first, that different units, have shown that this program has had an impact. The numbers are still small enough that I have reservations about absolute certainty. But already, we’re using the program in 20 of our hospitals.
What would your advice be for CIOs, CMIOs, and other healthcare IT leaders, as they consider these kinds of initiatives?
If we all are serious about transforming the way we care for patients, we need to do it in a data-driven way. There has to be a philosophical belief and commitment to do that. Number two, as a result of the institutional commitment and philosophy, then there has to be a team that’s dedicated to this work. I don’t think this is achievable in an ad hoc way, when people just have time. And three, it’s not for the faint of heart; it takes time and effort, but if you have the philosophical belief and institutional commitment, it’s doable. If I say to myself, I don’t ever want to leave my house and get drenched because I wasn’t prepared for a storm, then I need to check the weather app before I leave my house. In medicine, we haven’t yet taken that approach, but the data and analytics are there to guide us in helping us to make decisions, and making it a part of the everyday decision-making process. And in the same way I use examples of rehospitalization prediction, we also do condition-specific predictive analytics, around patients with asthma, kidney disease, etc., so there’s a lot of work going on there. And the message I give clinicians is, there will be companies that say they don’t’ sell you the predictive models they’ve developed; but in our experience, the models have to be a part of an organic process that leads to the building of the models. Clinicians won’t feel alienated, disenfranchised, or threatened, if you bring them in and engage them from the beginning.
