uf college of medicine office of research
Four faculty members are advancing AI-driven research funded by the institute.
By Manny Rea
The UF Artificial Intelligence and Informatics Research Institute endowed its inaugural seed fund award on an exclusive few projects from institute-affiliated faculty. Awarded to interdisciplinary teams advancing innovative, AI-driven research, four College of Medicine faculty members are contributing to biomedical research with strong real-world impact. From advancements in medical imaging to glucose monitoring systems, faculty discuss their AIIRI funded projects:
Next-Generation 3D Vision-Language AI for Real-Time Medical Image Understanding
What problem are you solving?
Modern medical imaging (CT, MRI) is inherently three-dimensional, yet most current AI tools are limited to narrow, task-specific applications that cannot interact naturally with clinicians. This creates a critical gap: physicians still cannot “talk” to imaging data or receive real-time, comprehensive insights from a single system. This work is urgent because healthcare is producing massive volumes of complex imaging data, and without more intelligent, interactive AI, we risk delays in diagnosis, inefficiencies, and missed opportunities for precision care.
This award emphasizes interdisciplinary work. How does combining experts from different colleges allow you to tackle this problem in a way a single specialist couldn’t?
This project is only possible through deep collaboration between the College of Medicine and Herbert Wertheim College of Engineering. AI experts design efficient, scalable models capable of understanding 3D data, while radiologists ensure the system learns clinically meaningful patterns and produces actionable outputs in real-world care. A single discipline cannot solve this: engineers alone may build powerful models that lack clinical relevance, and clinicians alone cannot develop the computational frameworks required. Together, we create a system that is both technically advanced and clinically transformative.
What does success look like?
Success will be a working proof-of-concept of a real-time 3D vision-language AI system that allows clinicians to query imaging studies using natural language — for example, asking a CT scan to “highlight suspicious tumor regions” and receiving immediate, interpretable results. Demonstrating accurate zero-shot disease detection, prompt-driven segmentation, and clinically meaningful question-answering within a single unified framework will be a major milestone.
Looking 5 – 10 years into the future, what is the ultimate “dream” for this research?
In 5 – 10 years, this research could redefine how clinicians interact with medical imaging — transforming static images into interactive, intelligent systems that function as real-time diagnostic partners. This could lead to faster diagnoses, improved accuracy, and broader access to high-quality care, even in resource-limited settings. More broadly, it establishes a new paradigm for AI: systems that seamlessly integrate visual data and human language, advancing not only healthcare but also the broader field of human-centered artificial intelligence.
AI-Ready Multimodal EEG Dataset for Cefepime-Associated Neurotoxicity
What problem are you solving?
Cefepime is one of the most commonly used antibiotics in intensive care units, but it can cause serious and often unrecognized brain toxicity – including seizures, delirium, and coma – in up to one in four patients who receive it. The problem is that clinicians currently have no reliable way to predict which patients will develop this complication before it happens, and by the time it is recognized, significant harm may already have occurred. This research is urgent because artificial intelligence tools capable of detecting early warning signs in brain wave recordings could transform how we monitor and protect critically ill patients – but building those tools requires a dataset that, until now, has never existed.
Why is cleaning and curating this specific dataset so important for the future of AI? How will making this data “FAIR” (Findable, Accessible, Interoperable, Reusable) help the broader research community?
AI models are only as good as the data they are trained on. Right now, researchers working on drug-induced brain toxicity are working blind, because no structured, well-characterized dataset linking brain wave recordings with drug levels and clinical outcomes has ever been made publicly available. Curating this dataset with rigorous, standardized methodology means that the hard work of data cleaning, expert outcome adjudication, and signal processing only needs to be done once, by a team with direct clinical expertise, rather than being repeated imperfectly by dozens of research groups working in isolation. Making this data FAIR matters because science moves faster when researchers can build on each other’s work rather than starting from scratch. A Findable and Accessible dataset means a research group in Tokyo or Berlin can ask the same questions we are asking at UF, test their own algorithms, and compare results against our benchmarks. Interoperability – formatting the data according to international standards like BIDS-EEG – means the dataset can be combined with other clinical EEG databases to build larger, more powerful models than any single institution could produce alone. And Reusability means the dataset will have scientific value not just for cefepime neurotoxicity, but for any researcher studying drug-induced encephalopathy, ICU brain monitoring, or AI-assisted EEG interpretation for years to come. In short, this dataset is infrastructure – the kind of foundational investment that makes an entire field of research possible
What does success look like? What is the most exciting milestone or “proof-of-concept” result you hope to share by the end of this award period?
Success looks like a moment we can point to and say: a computer, trained only on brain wave recordings and drug level data, predicted which ICU patients would develop cefepime toxicity before their clinical team recognized it. More concretely, by the end of this award period we expect to have three tangible deliverables. First, a fully curated, de-identified multimodal dataset covering 500 or more ICU patients — the first of its kind linking continuous EEG recordings with cefepime pharmacokinetics and expert-adjudicated neurotoxicity outcomes. Second, a validated AI model that can detect EEG signatures of early cefepime-induced neurotoxicity with clinically meaningful accuracy, measured by standard performance metrics including area under the receiver operating curve, sensitivity, and specificity. Third, a peer-reviewed publication describing the dataset, the methodology, and the model’s performance — establishing UF as the group that built the foundation for this entire research area. The most exciting proof-of-concept result would be demonstrating that EEG changes precede clinical recognition of neurotoxicity by hours — because if that is true, it means there is a window of opportunity for intervention that we are currently missing in every ICU in the country. That finding alone would justify a prospective clinical trial and position this work for NIH funding at a scale far beyond this pilot grant.
Looking 5 – 10 years into the future, what is the ultimate “dream” for this research?
In ten years, we envision a world where no ICU patient develops unrecognized brain toxicity from a medication their clinical team prescribed to save their life. The dream is an AI monitoring system embedded in the ICU that continuously analyzes a patient’s brain wave activity, cross-references it with their drug levels and kidney function in real time, and alerts the clinical team hours before neurotoxicity becomes clinically apparent — giving physicians the window they need to adjust the dose, change the antibiotic, or intervene before seizures or coma occur. One of the most sobering realities driving this work is how little we actually know about what happens to these patients after the ICU. We know cefepime can cause encephalopathy, myoclonus, and nonconvulsive seizures — but whether those events leave lasting cognitive or neurological consequences has simply never been systematically studied. The absence of evidence is not evidence of absence. By building the infrastructure to identify and characterize these patients rigorously, this research lays the groundwork for the long-term outcome studies that will finally answer that question. But the impact extends well beyond cefepime. The framework we are building — linking continuous physiologic monitoring with pharmacokinetic data and structured clinical outcomes to train predictive AI — is a template that could be applied to any drug with neurotoxic potential. Chemotherapy agents, immunosuppressants, antiepileptic drugs, sedatives — the ICU is full of powerful medications whose neurological side effects are poorly understood and even more poorly predicted. This research creates the methodological infrastructure to tackle all of them. At the level of AI science, this work addresses one of the hardest unsolved problems in clinical machine learning: how do you build models that are trustworthy enough to influence real patient care? By starting with rigorous data curation, expert outcome adjudication, and transparent benchmarking, we are demonstrating a standard of evidence that the field desperately needs. The ultimate dream is not a paper or a dataset or even an algorithm. It is a critically ill patient who receives the right antibiotic at the right dose, monitored in real time, with their brain health protected as carefully as their infection is treated — and a research community finally equipped with the tools to make that standard of care a reality everywhere.
Developing a highly accurate, non-invasive glucose monitoring system using advanced NIR spectroscopy and machine learning algorithms
What problem are you solving?
Continuous Glucose Monitors (CGMs) have revolutionized diabetes management offering real-time insights into glucose levels, which help maintain optimal control and reduce the risks of complications such as hypoglycemia and hyperglycemia. The invasive nature of the sensor insertion, which involves puncturing the skin, can cause discomfort, irritation, and a risk of infection. There is a critical need for developing non-invasive glucose monitoring.
This award emphasizes interdisciplinary work. How does combining experts from different colleges allow you to tackle this problem in a way a single specialist couldn’t?
Engineering provides the novel and exciting technology, and we in translational medicine understand the process, from regulatory issues to implementation. Our multidisciplinary team brings together a wealth of expertise across various fields to drive the project forward. PI Parisa Rashidi (BME) focuses on bridging the gap between machine learning and patient care, with a particular emphasis on developing AI-driven clinical solutions. Co-I Adam Khalifa (ECE) contributes his expertise in low-power and miniaturized circuit design for biomedical applications, which is crucial for advancing sensor technology. Co-I Mark Atkinson (Pathology) is an internationally recognized authority on type 1 diabetes. Finally, Co-I Desmond Schatz (Endocrinology), the medical director of the UF Diabetes Institute, offers a wealth of knowledge in translating glucose-sensing technologies into clinical practice
What does success look like? What is the most exciting milestone or “proof-of-concept” result you hope to share by the end of this award period?
We anticipate achieving a real-time, edge-computing platform on a custom PCB that processes data for non-invasive glucose monitoring with comparable accuracy to cloud-based models. Over the two-year project, we plan to recruit 30 participants and collect a comprehensive dataset comprising spectroscopy data, smartwatch data, participant-reported information, and CGM readings. By the end of the first year, we aim to develop a transformer model that accurately predicts CGM values based on these inputs.
Looking 5–10 years into the future, what is the ultimate “dream” for this research?
A truly effective non-invasive CGM which will improve the lives of all with diabetes