Identifying Patients Needing Help With Squark AutoML At UPMC

UPMC, or the University of Pittsburgh Medical Center, is a $23 billion nonprofit health care provider and insurer in Pennsylvania, New York, Maryland and with several locations internationally. It is the largest nongovernmental employer in Pennsylvania, with over 92,000 employees, 40 hospitals, 800 doctors’ offices and outpatient sites, and more than 4 million insured members through its Insurance Services Division.

UPMC has a number of world-class medical specialty areas, one of which is living donor liver transplants (LDLT). Nationally, LDLT is a small specialty, with only about 500 transplants per year across all centers, but UPMC has the largest living donor practice in the U.S. Within the U.S., there is a long waiting list for liver transplants, but many patients who need one aren’t familiar with the living donor option, even though it tends to have more successful outcomes. UPMC has advocated to help patients reduce their wait for cadaveric organs by considering LDLT as an alternative.

UPMC clinicians and administrators concluded that the provider would be performing a valuable service if they marketed its living donor transplant services to those who might benefit from them. It’s a narrow group of individuals, however, and they weren’t sure how to target such a small population. Many doctors aren’t even aware of the living donor transplant option, so they knew that educating patients was going to be an important part of the process.

UPMC and Squark

The UPMC Marketing Intelligence group, spearheaded by Jake Collins, Director of Marketing Intelligence, did have an idea how to reach the potential patients. He has a background in data analysis and knew that UPMC could reach out to potential patients while maintaining their privacy in adherence with HIPAA guidelines. The Marketing Intelligence group at UPMC  maintains a robust national consumer database with a large number of variables on each individual. He knew that under typical identification models, UPMC would have to reach out to about ten thousand people to get one possible transplant candidate, but that they could reduce that number considerably with some targeting based on the data and statistical analysis.

Identifying people outside of UPMC’s traditional footprint with a need for liver transplants using 1,500 different variables would normally be very labor-intensive, but Collins had recently brought Squark, a “low code predictive analytics tool”, into Marketing Intelligence. Squark, which was co-founded by my friend Judah Phillips, employs an easy-to-use approach to automated machine learning (AutoML), which isn’t true of some other providers. Some vendors of AutoML are primarily for data scientists, but Squark is focused on democratizing semi-automated prediction so that business users (ideally with a bit of quantitative training) can use it. It’s also a lower-cost option than many AutoML tools.

Collins commented that Squark made the cost of creating this model so low you could think about doing it anywhere. He said that in addition to needing more monetary resources, he used to need a data scientist and six months to do the kind of analysis they did for liver transplant prospects. Now, based on the initial work done by the Marketing Intelligence team, anybody on the team with a basic understanding of statistics can set up and run machine learning models, and much more quickly. In grad school (at Carnegie Mellon, noted for its quantitative focus) Collins ran a regression model on 114 variables. Doing this required that he write code, examine the outcomes in a highly manual process, and it took him months to complete. Now, with Squark, he can do models with 1000 variables, and runs them in an hour with the same accuracy. The team can now do hundreds of different analyses in a few months; they can try out all sorts of ideas.

The modeling approach Collins and his group use examines all the variables for various geographies and age groups, along with the known liver disease outcomes for those same groupings, within an AutoML model. He’s trying to find the variables that are most influential on the outcomes at the group level. The resulting model is then used to score each individual person across all geographic areas based on their likelihood of having the condition. Collins generally takes the best performing deciles, and then uses a partner company “to go from a person in the database to a person in the digital space,” as he put it. Then they target and retarget informative ads on all platforms.

Collins is also comfortable with the idea of using an AutoML tool like Squark on an ongoing basis. He feels it is empowering his “citizen data scientists” to do more complex analysis with the same level of accuracy on behalf of UPMC and its patients. He has a UPMC-employed data scientist “check the math” before they embark upon a campaign, to ensure that the models are being run and interpreted correctly.

Overall, Collins says, the team has gotten phenomenal results from the liver transplant campaign. Instead of reaching out to 10,000 people to get one interested patient, UPMC only has to reach out to 75 now. There was a tremendous response to the call center and UPMC has added 50% more traveling patients to its liver transplant pipeline through this and other tactics.

Beyond Liver Transplants

The liver transplant program was the first test case for the outreach approach, and now Collins’ group is pursuing similar approaches in other areas. They are working on a program for pancreatic cancer, collaborating with a UPMC physician who uses a lot of advanced statistical analysis in his own work. They also have some musculoskeletal programs in mind that they feel could help more patients. 

Collins says he believes that no other healthcare providers are doing this or even collecting data of this magnitude to support it. But the executive leadership, especially the Chief Marketing Officer of the hospital system, are highly supportive and are willing to spend on national campaigns in areas where UPMC has strong capabilities. Collins noted, “It’s rare for anyone to do a truly national marketing campaign in healthcare, but where we have something unique to offer that benefits patients, we consider it.”

“The data and the models help us deliver the care at the right time and the right place to the right person in most effective way possible,” Collins said. “We don’t use our own patients’ individual data and we don’t violate privacy, that is sacred.” Most importantly, he concluded, “The patients who get transplants or other successful treatments are very glad we reached out to them.”