Investigators at the Jacobs School of Medicine and Biomedical Sciences are deploying advanced artificial intelligence, including the proprietary SCAI large language model, to modernize trauma triage, protein structure mapping, and patient discharge workflows.
These tools, supported by New York State’s Empire AI infrastructure, aim to synthesize rapidly doubling medical knowledge into actionable, patient-specific data.
SCAI Model Handles Complex Semantic Reasoning
The Department of Biomedical Informatics at the Jacobs School developed the Semantic Clinical Artificial Intelligence (SCAI) model to handle complex semantic reasoning. Published in JAMA Network Open, data showed that SCAI outperformed most physicians on all three sections of the United States Medical Licensing Exam. The system functions by processing 66 million medical facts, ranging from genomic data to clinical guidelines.
Despite its performance, researchers are actively working to isolate and eliminate systemic bias within the model. Doctoral candidate Rachel Stephens tested SCAI against 75 historical case studies from Massachusetts General Hospital, finding that altering demographic markers—such as housing status, race, or LGBTQ+ identity—frequently shifted the AI’s recommended care path.
“If we add that the patient is unhoused, the advanced imaging that was previously indicated is now no longer recommended,” Stephens told the UB Medicine alumni magazine. “The assumption is the patient can’t afford it.”
LLMs Condense High-Pressure EMS Communications
In the Department of Surgery, the Data Augmented Research Technology in Surgery (DARTS) laboratory is using large language models to condense high-pressure EMS communications. Trauma surgeons must often filter “noisy” prehospital audio in real time; an LLM can transcribe and strip nonessential data to extract vital signs and injury mechanisms.
Research on 133 pediatric emergency department activations showed that LLM-assisted triage improves accuracy. When clinicians reviewed an AI-generated summary alongside their own initial assessment, their likelihood of correcting an incorrect triage decision tripled. “What the results do show is that LLMs can match or slightly exceed human accuracy in interpreting EMS communications,” said Peter C. W. Kim, MD, PhD, vice chair for research and innovation in the Department of Surgery.
Researchers Model Dynamic Protein Movements
Beyond clinical triage, investigators are pushing the boundaries of generative AI in molecular biology and neurology. Thomas Grant, PhD ’13, is integrating SWAXS (small- and wide-angle X-ray scattering) data with AlphaFold-derived algorithms to model dynamic protein movements. This approach aims to move beyond static structure prediction, theoretically accelerating drug discovery by identifying how proteins shift shape to function.
In neurology, researchers have successfully used generative AI to visualize cortical gray matter lesions in multiple sclerosis patients, structures previously invisible on standard MRI scans. By analyzing the relationships between multiple image sets, the AI detects subtle tissue discrepancies that indicate disease progression.
“Generative AI is very powerful because it can look between the scans and detect tiny differences between them,” said Michael G. Dwyer III, PhD, associate professor of neurology and biomedical informatics.
Improving Patient Discharge Outcomes
The “Team Alice” research program is targeting the vulnerability of older adults during hospital transitions by using the AI platform Open Evidence to simplify discharge summaries. Rather than relying on standardized templates, the system uses patient-specific medical characteristics to generate tailored instructions. This addresses a critical gap in current AI training, which is often not optimized for geriatric populations with multiple chronic conditions.
“AI is able to summarize large amounts of patient information and format it in a way that’s approachable and individualized for the patient,” said Andrew Baumgartner, MD, assistant professor of family medicine. The team is currently gathering feedback from caregivers to refine the output, ensuring that the information provided is accessible regardless of the format, whether written, audio, or video.
The 2026 Clinical Integration Verdict
- SCAI: A specialized LLM trained on 66 million medical facts; currently used to assist with medication dosing and complex diagnostic reasoning.
- DARTS Laboratory: Leveraging LLMs to reduce prehospital trauma transcripts by 80%, providing actionable summaries for emergency staff.
- CANDO Platform: An AI-driven drug discovery tool currently being used by biotech startups to model compound interactions for cancer and opioid addiction treatments.
- Regulatory Outlook: Researchers are actively developing risk-mitigation strategies for “prompting the machine” to prevent discriminatory medical recommendations.
As the Jacobs School continues to integrate these systems, the emphasis remains on augmentation rather than replacement of human judgment. “AI is really something that augments your decision-making,” said Irshad Ally, MD ’20, a physician with UBMD Primary Care. “It helped me get a starting point and it was reassuring to have that.”
The ongoing refinement of these tools reflects a broader effort to adapt AI to a “broken health care system” that demands both speed and precision. “Physicians have been working in an incredibly broken health care system,” said Baumgartner, “so any potential improvements to the system are worth exploring.”