Kent Deighton / Harvard Chan School
Dr. Matlin Gilman, ’26, studies the impact of policy decisions on real-world health and builds artificial intelligence tools to advance healthcare.
After 10 years in health policy research in academia, nonprofit organizations, and the federal government, Matlin Gilman arrived at Harvard’s TH Chan School of Public Health with a firm belief that, in his words, “rigorous policy analysis and cutting-edge data science can be powerful tools for improving population health and promoting health equity.”
As a doctoral candidate in population health sciences at Harvard University’s Kenneth C. Griffin School of Arts and Sciences, Gilman has carved out a unique academic path to turn that idea into reality. The Department of Social and Behavioral Sciences at the Harvard Chan School focuses on how to use advanced statistical methods to analyze the health impacts of policy decisions. As a cross-enrollment student at Harvard’s John A. Paulson School of Engineering and Applied Sciences (Harvard SEAS), he focused his research on artificial intelligence (AI) and machine learning technologies to help hospitals and clinicians improve patient care.
Measuring the effectiveness of abortion bans
Reproductive health care changed dramatically in June 2022 when the U.S. Supreme Court overturned federal abortion protections. That same year, 13 states banned abortion, and four more states have since enacted bans. In her doctoral dissertation, Gilman set out to measure how these states’ abortion bans affected reproductive, maternal, and infant health outcomes in the two years following the Supreme Court’s decision.
“Access to reproductive health care has changed dramatically in some states and not at all in others,” Gilman said. “This kind of variation allows us to study cause and effect, but only if we can build a reliable picture of what would have happened in the absence of prohibition.”
To build this picture, Gilman used Bayesian modeling, a statistical approach that dynamically integrates new data to improve predictions. She built a model that predicted state-level reproductive health outcomes if abortion access were upheld, taking into account each state’s pre-ban trajectory and broader national trends, such as economic changes and the COVID-19 pandemic, that the model indirectly captured. “The gap between model predictions and actual reproductive health outcomes was the estimated effect of the ban,” Gilman explained.
Gilman found that in states with abortion bans, birth rates increased more than would be expected without the ban, especially for Hispanic and black women and for women whose highest educational attainment was a high school diploma. In contrast, Gilman said, “We did not see a significant effect for women with a college degree. This suggests that workarounds available to some people, such as expanding telemedicine or traveling out of state for abortion care, may be unaffordable to others.”
Gilman’s study also found that neonatal mortality (death within the first 28 days of life) increased in prohibition states, primarily due to deaths from severe birth defects. Mortality rates for black infants in these states were also higher than expected. Professor Gilman found no statistically detectable impact on maternal mortality, but said maternal deaths are so rare that small changes are difficult to detect. “The full impact of abortion bans on maternal health may take years to be realized,” he said.
Building AI tools for healthcare
While at Harvard Chan School, Gilman’s interests expanded beyond measuring policy effectiveness. He became interested in exploring how complex medical information can be made more accessible and useful to those who need to act on it, whether it’s hospitals trying to understand their own performance or clinicians looking for the best available evidence.
Gilman decided to complement his public health research with data science research. He was cross-enrolled with Harvard SEAS and focused on machine learning, architectures behind large-scale language models, and engineering.
As one of his projects at Harvard SEAS, he analyzed Medicare’s value-based purchasing program. The program adjusts hospital reimbursement by up to millions of dollars based on clinical outcomes, safety, patient experience, and efficiency. He built a machine learning model that predicts whether a hospital would receive higher or lower reimbursement based on its performance, and identified the most important factors in that prediction.
“This model shows what individual hospitals should focus on to improve performance under the program. Patient experience and efficiency drove the prediction more than other areas,” he said. “This is especially valuable information for safety-net hospitals, which often face financial instability.”
In another project, Gilman devised an AI-powered web application that allows clinicians to effectively and efficiently search published medical literature and stay informed of current research findings. This is a huge undertaking with thousands of new studies being published every day. He built this tool with his SEAS classmates at Harvard University. It uses large-scale language models to understand the meaning of user questions, searches indexed medical research databases for relevant studies, and generates synthetic answers that include direct citations. Each cited study is labeled with its conflict of interest status, how widely it was cited, and how recently it was published. Users can filter results by any of these attributes and switch between clinical and research modes to tailor the response to their needs.
“The literature is huge, and no one can read it all,” Gilman says. “We’ve built tools that synthesize evidence, link to sources, and allow users to decide which research to prioritize, so the answers are useful and verifiable.”
Looking to the future
As he nears completion of his doctorate, Gilman is exploring roles in academia, health systems, and technology. His ultimate goal is to use data to improve health outcomes and enable well-being for all communities.
“We have more health data than ever before, but data alone cannot improve outcomes,” he said. “Someone has to do the careful work of understanding what it means and making it useful. That’s the part that interests me the most.”
