For years, scientists have searched for the hidden clues that explain why two people with the same disease can have dramatically different outcomes. Why does one patient respond well to treatment while another grows sicker? Why are some communities consistently hit harder than others?

Associate Professor,
School of Public Health
University of Minnesota School of Public Health Associate Professor Sandra Safo is using artificial intelligence to help answer those questions. By combining biological data with information about patients’ lived experiences, her research aims to uncover why people from different backgrounds experience complex diseases so differently—and how that knowledge could lead to more individualized care.
Her latest work focuses on Chronic Obstructive Pulmonary Disease (COPD), a condition that causes inflammation and damage in the lungs and is the fifth leading cause of death nationwide. Despite its prevalence and severity, researchers still don’t fully understand why the disease progresses so differently for people, particularly across race, gender, and socioeconomic status. For instance, Black Americans are more likely to end up hospitalized and suffer from severe breathing challenges. Women’s symptoms often worsen at a faster rate, especially if they’re smokers.
When Safo realized disparities exist across COPD, she saw an opportunity to leverage her expertise to investigate why they exist across populations. Because this problem isn’t exclusive to COPD, she knew her research could help address the inequities that drive people’s experiences with many diseases, including diabetes, heart conditions, and different kinds of cancer. Ultimately, her research could help move healthcare beyond a one-size-fits-all model toward treatments that are more personalized—and more equitable.
“This method was motivated by COPD, but it was built so that it can be applied to pretty much any data set to assess disease risk and progression,” said Anja Logan, a biostatistics doctoral student who worked with Safo on the project.
Untangling Risk Factors
Safo and Logan developed a new deep-learning model to better understand how COPD outcomes are shaped by both biology and a person’s broader circumstances. The model analyzes biological data alongside demographic information, socioeconomic status, access to healthcare, environmental exposures, and other key factors.
Using AI, the model identifies biomarkers—measurable indicators in the body, such as proteins, genes, or blood counts—that are linked to disease risk and progression. Through its ability to analyze so many layers of data, their model can identify biomarkers that are connected to both the general population and to different subgroups with different histories or exposures.
Most previous research has focused primarily on the overall population, often overlooking how disease risk differs across demographic groups or how factors like income, neighborhood conditions, and access to healthcare shape health outcomes. These influences—known as the social determinants of health—are essential to understanding why disparities exist, Safo says.
“Failure to integrate these contextual factors limits our ability to understand the actual risk factors and disease progression in disadvantaged populations,” said Safo.
Banishing Bias in the Data
The researchers rigorously tested their model to ensure it could accurately identify connections between biomarkers and different patient populations. They also designed it to minimize algorithmic bias, helping ensure that its predictions are reliable across diverse groups.
This approach, known as precision equity, aims to make healthcare predictions more accurate and fair for every patient, regardless of their background.
Logan is now developing user-friendly software with the goal of creating a plug-and-play tool that researchers and clinicians can easily use. Over time, the team hopes the model will help improve diagnosis and treatment not only for COPD but also for many other diseases, including several types of cancer.
“Our goal is to build computational models that can disentangle what’s relevant for the general population and what’s specific to certain groups while reducing bias in AI, “ Safo says, pointing to a future that promises more accurate and equitable risk assessment, more effective treatments, and individualized care for complicated diseases.


