University of Kentucky Uses AI to Advance Biomedical Research

At the University of Kentucky, the Center for Applied Artificial Intelligence (CAAI) was launched to bridge the gap between biomedical research theory and operational implementation through automated data sorting, transcription and protocol management. In just two years, the center has led more than 65 AI projects and collaborated with over 40 partners to deliver repeatable services rather than one-off demos.

HOW IT STARTED

The impetus for CAAI stemmed from past projects which had money and data but lacked the expertise to bridge technical and domain knowledge. Previous initiatives, such as the Institute for Biomedical Informatics’ Innovation Core and the AI in Medicine Alliance, revealed that giving researchers funding to explore AI use without a technical partner often resulted in stalled projects.

“What we learned was that just giving people money, and even giving people data and money to do these sorts of projects, pretty much all of them fail,” Cody Bumgardner, director of the center, said. “[Researchers] didn’t have the data that they needed or not in the way that they needed it, and most importantly, they didn’t have people to do the work to kind of bridge the gap.”


CAAI was built to fill that need. Housed within the university’s Institute for Biomedical Informatics, the center provides infrastructure, reusable code and technical staff that allow collaborators to move from concept to prototype quickly and trust data is housed securely.

Bumgardner said collaborations with the center can begin informally, with researchers hearing about CAAI through word of mouth, or publications on work the center has contributed to. Then, researchers can meet with CAAI team members to translate their individual problems into a technical plan, often involving a self-service tool that allows them to experiment on their own time. Having adaptable self-service tools helps researchers get started more quickly, Bumgardner said, so they are not waiting on bureaucratic processes or individuals to create one-off tools.

Aaron Silverstein, a Ph.D. candidate who came to CAAI for help with his dissertation work, followed that pathway. A professor he worked with suggested he reach out to them for help deciding which of more than 200 variables in his research concerning rat gaits after spinal cord injuries were the most important to focus on. A professor made the suggestion, and Silverstein learned how to use a self-service tool called CLASSify after one video call with CAAI staff.

THE CENTER’S SELF-SERVICE TOOLS

CAAI’s reusable platforms remove repetitive technical work from individual projects, and some are simpler than others.

For example, CAT-Talk is an in-house speech-to-text transcription service.

Another tool, LLM Factory, provides a locally hosted platform for users to fine-tune large language models to their own needs, also known as creating custom GPTs. It uses OpenAI training models as a base, and data uploads from individual users for customization, without using those data uploads to train OpenAI’s general models on proprietary research.

Other tools tackle more complicated tasks. CLASSify, the tool Silverstein used, is a web-based tool that lets users train and compare machine learning classifiers on any data that comes in a table. Using the AI, Silverstein said, can help validate variable choices and avoid purely subjective selection.

CLASSify helps with interpreting data collected, not necessarily by forming written conclusions, but by determining which of many variables changed the most, as in Silverstein’s case. If a researcher is working with sensitive health data, it can generate synthetic example data that follows the same patterns, so researchers can share key findings without compromising individual information.

“The most surprising thing was how easy it was to actually use,” Silverstein said. “I do not come from a background with really any programming experience at all.”

SmartState is a platform that automates research protocol compliance processes. Making use of agentic AI, virtual agents guide study participants through protocols and send messages to reduce the need for manual reminders and follow-up. It also documents study progress.

THE CENTER’S IMPACT

CAAI’s projects span surgery, neonatal health, cardiology, public health and more. In neonatal care, CAAI supported machine learning models analyzing data from ventilators and oxygen monitors to predict when pre-term babies are ready to have breathing tubes removed. That helped doctors make more informed decisions, according to an executive report from CAAI. Another project through the center uses ambulance and emergency data to forecast opioid-related incidents across the state, helping inform response.

Operational uptake like this is one way the center measures success, Bumgardner said, along with helping its people grow and continuing to advance research.

“What can we do to actually directly apply to help folks?” he said.