ALBANY — Albany-based health insurer CDPHP is enlisting the help of RPI researchers to better provide coverage to its highest-needs members.
The effort will rely on artificial intelligence to analyze patient information and treatment histories to better understand what works and why, then look for ways to create small subgroups for whom a plan of treatment can be built. The goal is better outcomes for the chronically ill at a lower cost for the insurer.
The RPI team is led by Kristin Bennett, a professor of mathematics and director of RPI’s Institute for Data Exploration and Applications. She said the growing complexity of modern medicine and increasing amounts of data about patients have created the opportunity for such an approach to benefit both patient and insurer.
“You want to discover in the data how different patients have different outcomes and different needs, automatically,” she said. “Health care is getting more and more complex … you can’t make guidelines that are simple enough any more.”
Bennett said part of the challenge is grouping patients together in a way that generalizations about their ailments and treatment are meaningful — if the group is too small, the conclusions are unique to its members, and can’t be applied to other groups.
She is joined by Malik Magon-Ishmail, a professor of computer science, and Jason Kuruzovich, an associate professor of management. Working with them is Tanvir Khan, CDPHP’s chief analytics officer.
Khan said the health insurer already has its own machine-learning model in operation.
“It’s not that our methods aren’t working,” he said. “It’s working, but RPI ... is showing more promise.”
Often, the doctor or nurse discharging a high-needs patient after a particular episode doesn’t have a full grasp of all the factors in that patient’s condition, just an understanding of what put him or her in the hospital this time, Khan said.
The goal of the CDPHP-RPI project is to create a system that will suggest the best course of health management to reduce the chance of future crises.
“It would still be a doctor’s choice,” Bennett said.
She calls this the cadre model of machine learning. Standard machine-learning and deep-learning models can predict which patients may require extensive care but cannot explain why, or suggest ways to reduce the problems.
“Really the exciting part is being able to work with all this data and be able to impact health care in my community,” she added.
Khan said there’s a risk of false or misleading results because human perception and judgement are involved in generating the data that the computer will analyze. One of the goals is to refine the program’s algorithm to correct for that.
There is not, however, a risk to the confidentiality of CDPHP’s 366,000 members: The medical records being analyzed for this project have had identifying information stripped from them, and they aren’t being exported from CDPHP’s servers.
The end result of the project will be put to work for CDPHP and its members with the highest level of medical need.
“In fall, we expect it to be deployed,” Khan said.