The medical mistake harms millions of Americans each year and costs billions
RESEARCH TO WATCH
Researchers at the University of Michigan College of Pharmacy and College of Engineering are collaborating to solve the problem of medication dispensing errors and patient safety issues related to pharmacies.
The team will develop artificial intelligence and imaging-based error-detection methods that train AI to prioritize avoiding mistakes that cause the most harm to patients. The systems the researchers plan to design will study how AI can communicate effectively with humans in pharmacy work flows.
Why it matters
Every year, more than 4 billion prescriptions are dispensed in the U.S. with an error rate of 0.06%-or about 2.4 million incorrectly dispensed medications. Those errors result in about 1.5 million emergency room visits and lead to extra medical costs related to treating drug-related injuries. For hospitals alone the errors cost about $3.5 billion a year. The errors saddle some patients with lost wages and poorer health, and cost the economy in lost productivity and additional health care costs.
How much and for how long
The $1.3-million grant from the National Institutes of Health's Institute on Aging will fund the project titled, "Preventing Medication Dispensing Errors in Pharmacy Practice with Risk-sensitive Artificial Intelligence." The project began in May 2026 and will conclude in May 2030.
The work will build on a previous R01 grant, the NIH's most common grant for independent projects. That project trained AI models to identify the pills that were supposed to be in a given bottle with 99.5% accuracy. In the .05%, it was unclear if the researchers taught the AI model to communicate so that a human could double-check the medication dispensed. The researchers also worked with pharmacists to measure trust between humans and the AI model.
Who's involved
Corey Lester is an assistant professor of clinical pharmacy and a medication safety expert who uses pharmacy informatics techniques to improve the delivery of healthcare.
"A dispensing error is more than just the wrong pill in a bottle-it can expose a patient to preventable harm while leaving their underlying condition untreated," he said. "As pharmacies shift from direct vial inspection to reviewing digital images of filled prescriptions, one of pharmacy's most critical safety checks is being redesigned. That creates a major opportunity to improve safety and workflow, but only if AI is built to support the way pharmacists actually make decisions.
"Most AI models treat every mistake the same, even though some medication mix-ups carry far more serious consequences than others. We are developing risk-sensitive AI that accounts for those clinical stakes and communicates them clearly to pharmacists. We don't just want a system that says, 'this looks wrong.' We want one that helps pharmacists understand when a potential error is especially risky and deserves closer attention.
"The goal is to amplify pharmacists' clinical judgment-not replace it-while reducing cognitive burden and building appropriate trust in AI-assisted verification."
Raed Al Kontar is an associate professor of industrial and operations engineering. His research focuses on personalized, collaborative and distributed data analytics with an emphasis on effectively integrating knowledge from diverse data sources.
"Not all medication errors carry the same consequences, yet most AI systems optimize for accuracy as if every mistake were equally harmful," he said. "Our research is developing a new generation of cost-sensitive and uncertainty-aware AI that focuses on preventing the most consequential medication errors.
"A key scientific challenge is that traditional cost-sensitive learning frameworks do not naturally extend to modern deep learning, requiring fundamentally new machine learning methods that explicitly incorporate the asymmetric costs of clinical decision errors. We hope this work will establish new principles for designing AI systems that better reflect real-world clinical risk."
Xi Jessie Yang is an associate professor of industrial and operations engineering. Her research is focused on human-autonomy and human-robot interaction. She seeks to understand the underlying mechanisms governing these interactions and to propose design solutions facilitating them.
"Another goal of this project is to model how pharmacists trust and interact with cost-sensitive AI. The goal is not simply to make pharmacists trust AI more. Too much trust can lead someone to accept an incorrect recommendation while too little trust can cause them to disregard a useful warning.
"We want AI systems that clearly communicate their uncertainty and the potential consequences of an error, helping pharmacists know when they can rely on the system and when a prescription warrants closer examination."
