A new artificial intelligence laboratory at the University of Texas Medical Branch is tackling one of healthcare's most stubborn problems: getting doctors to actually trust AI predictions. The DIVA-AI Lab, launched through a $300,000 Rising STARs award, is developing methods to integrate fragmented medical data,images, clinical notes, genetic information, and wearable device readings,into a single, interpretable view of each patient.

Why Are Doctors Skeptical of AI in Medicine?

Despite rapid advances in artificial intelligence, clinical adoption remains sluggish. The barrier isn't technology itself, but two interconnected challenges that have plagued healthcare AI for years. First, medical data is messy. Patient information lives in different formats across different systems: a CT scan here, a lab result there, genetic sequencing data somewhere else. Combining these heterogeneous sources into a model that actually learns something useful has proven extraordinarily difficult.

Second, even when AI models perform well, they often function as "black boxes." A model might predict that a patient has a 78% risk of complications, but clinicians have no way to understand why. Without interpretability, doctors understandably hesitate to act on AI recommendations, especially when patient safety is at stake.

How Is the DIVA-AI Lab Solving This?

Dr. Vibhuti Gupta, an associate professor in the Department of Biostatistics and Data Science at UTMB's School of Public and Population Health, received the Rising STARs award on April 28 and is now leading the effort. His team is building a technical pipeline called Multimodal Hybrid AI, or MoHAI, which uses two different types of machine learning working in tandem.

The architecture works like this: deep learning handles the complexity of diverse biomedical data, including missing values and variations in which signals matter most for each patient. Then, traditional machine learning methods deliver the interpretability that clinicians require. Rather than treating these as competing approaches, the lab uses them sequentially, with each handling the job it does best.

"I believe impactful multimodal AI in healthcare requires three essential components: deep learning to extract rich representations from complex and heterogeneous data, machine learning methods that provide interpretable and clinically meaningful insights, and close collaboration with clinicians, patients, and domain experts to ensure clinical relevance, usability, and responsible deployment," said Dr. Vibhuti Gupta.

Dr. Vibhuti Gupta, Associate Professor, Department of Biostatistics and Data Science, UTMB

The lab will draw on massive national data repositories to train its foundation model, including the All of Us research program, TriNetX, the Medical Information Mart for Intensive Care (MIMIC), The Cancer Genome Atlas, The Cancer Imaging Archive, the Surveillance, Epidemiology, and End Results program, and the UK Biobank, alongside local resources like the UTMB Moody Brain Health Initiative.

What Data Sources Will the Lab Use?

• Medical Images: CT scans, X-rays, MRIs, and other imaging data processed through specialized encoders trained specifically for visual information.
• Clinical Notes: Unstructured text from patient records, physician observations, and medical histories converted into machine-readable representations.
• Genomic Data: Genetic sequencing information and biomarker data that reveal disease risk and treatment response patterns.
• Wearable Sensor Data: Continuous physiological signals from devices like smartwatches and fitness trackers that capture real-world patient behavior.
• Tabular Records: Structured data including lab results, vital signs, medication lists, and demographic information.

Each data type passes through what researchers call a "modality-specific encoder," essentially a specialized AI model trained to understand that particular kind of information. The outputs are then fused into what the field calls a "unified patient embedding," a single mathematical representation that captures the essence of a patient's health status across all available data sources.

Gupta's path to this work was unconventional. He began as a computer scientist, earning his PhD from Texas Tech University in 2019 with a dissertation on big data stream analytics. But during his doctoral work, he took on healthcare projects on the side, and the experience fundamentally shifted his career direction. "I chose to focus on healthcare research over industry because of its potential for real-world impact," he explained. "I was motivated by the possibility of applying artificial intelligence and data science to some of society's most pressing health challenges, where even small advances can make a meaningful difference in patients' lives".

That conviction led him to a postdoctoral fellowship at the University of Michigan, where he worked on a project examining whether continuous data from wearable devices, paired with survey responses from a mobile health app, could identify post-transplant risks early in patients undergoing bone marrow transplantation. The same project examined the strain on family caregivers, who support transplant patients around the clock and often carry their own physical and mental health burden.

How Will Clinicians Actually Use This AI?

The DIVA-AI Lab's approach includes a crucial third layer: the foundation model becomes a feature extractor that disease-specific models can fine-tune for particular clinical tasks. These downstream models are simpler, more interpretable, and grounded in the rich representations the foundation model has already learned. This is where clinicians would actually interact with the system.

For example, a cardiologist might use a fine-tuned model for patient risk stratification in heart failure, while an oncologist uses a different fine-tuned model for cancer subtyping. Both leverage the same underlying foundation model, but each is tailored to its specific clinical context and designed to be understandable to its users.

The lab will also evaluate the foundation model for bias and completeness across different patient populations and healthcare settings, addressing a critical concern in healthcare AI: that models trained on data from wealthy, urban hospitals may perform poorly for patients in rural or underserved areas.

Gupta's earlier work at Meharry Medical College, where he served as founding assistant professor in the newly launched School of Applied Computational Sciences, gave him experience navigating academic research funding. He served as a peer reviewer for the National Science Foundation and studied the patterns that separated funded proposals from rejected ones. His first grant as principal investigator came in 2021, a pilot award from the American Cancer Society to develop deep learning methods for detecting virus integration sites in tumor genomes. In 2023, he received an NIH AIM-AHEAD consortium development award of roughly $1.27 million to build a multimodal framework for prostate cancer risk prediction.

The DIVA-AI Lab represents an expansion of an existing UTMB facility. Dr. Suresh Bhavnani received his own Rising STARs award years earlier and used it to establish the original DIVA Lab, focused on visual analytics methods that translate complex statistical models into forms interpretable to clinicians and researchers. Gupta's lab adds multimodal foundation modeling to that work, with the renaming to DIVA-AI signaling the new scope.

As healthcare systems worldwide grapple with the promise and peril of artificial intelligence, the DIVA-AI Lab's focus on interpretability and data integration addresses real clinical needs. The question is no longer whether AI can outperform humans on narrow tasks, but whether AI can be integrated into clinical workflows in ways that clinicians understand, trust, and can act upon with confidence.