Not only do the majority of people with diabetes live in low- and middle-income countries, per the World Health Organization, but treatment also lags in those under-resourced locales, more often leading to death or disability. To more effectively diagnose and treat patients in developing countries—and even here in Wisconsin—healthcare organizations have turned to programs that disperse community health workers to reach a wider swath of patients.
But devising a plan for dispatching those workers isn’t as simple as sorting a list of patient names by their addresses; with limited staff and resources, each visit essentially comes at the expense of another. And each patient, depending on their health and state of mind, responds best to a certain visit cadence.
Healthcare optimization researchers from the University of Wisconsin-Madison, the University of Texas at San Antonio, the University of Ottawa and the Indian School of Business have developed a framework for generating community health worker deployment plans, factoring in patients’ estimated compliance while balancing the tradeoffs between reaching new patients and caring for existing ones. They published their work in the journal Operations Research in May 2026.
“Healthcare is a huge cost,” says Yonatan Mintz, an assistant professor of industrial and systems engineering at UW-Madison and senior author on the paper. “Being able to use this kind of AI and optimization approach really allows us to scale up these types of programs and use our resources more wisely and reach more communities.”
Katherine Adams (PhDISyE ’24) led the work as a PhD student working with Mintz and Justin Boutilier, a former UW-Madison assistant professor who’s now at the University of Ottawa. In addition to her two advisors, Adams collaborated with Sarang Deo, a professor of operations management at the Indian School of Business, and leveraged data from NanoHealth, a health-tech company in India.
The researchers developed new methods to produce visit schedules for community health workers in lower-income, urban environments in India, finding they could ultimately reduce fasting blood glucose levels by up to 25% with the same resources. Their framework integrates behavioral modeling with operational planning to improve treatment efficiency and effectiveness.
“When you have limited resources, visiting someone will often mean you have to not visit someone else,” says Adams, who’s now a postdoctoral researcher as part of a bridge to faculty program at UT San Antonio. “But at the same time, we do have different heuristics that will prioritize equity a little bit more or others that will prioritize efficiency a little bit more. So, the planner in charge of this could still pick the heuristic based on the goal that they’re most concerned about.”
The team’s framework takes into account individual patients’ health data as well as their motivational states—how apt they are to enroll and stay enrolled in a treatment program, considering potential negative factors such as the social stigma of drawing attention through repeated visits (the “nosy neighbor” effect) and the cognitive burden of learning a new treatment program and making behavioral changes. It establishes optimal visit cadences for each patient, reducing treatment program dropout rates while smartly allocating resources.
“It’s really key to not just think of a population as a population, but to think of it also as a collection of individuals,” says Mintz. “Because everyone has different needs, and it’s important to have approaches that incorporate them into the system.”
While the team initially applied its approach to a setting on the other side of the world, Mintz sees clear use cases right here in Wisconsin and the Midwest, including substance-use disorder treatment in rural areas and tribal communities, mental healthcare for farmers, healthcare in urban centers, and more.
“If you have a chronic condition, it’s not like treat it once and you’re done,” says Adams. “It places a much larger burden on the healthcare systems, which are already insufficient. So that’s why trying to plan something in this way is so important.”
Funding for this research came from the National Library of Medicine (training grant NLM 5T15LM007359 to the Computation and Informatics in Biology and Medicine Training Program), the UW-Madison Global Health Institute (seed grant) and the American Family Funding Initiative.
Top photo courtesy Trinity Care Foundation/Flickr (Attribution-NonCommercial-NoDerivs 2.0 Generic license)