AI That Helps Doctors Think Smarter

A clinical AI tool developed at the University at Buffalo has demonstrated remarkable accuracy, passing a physicians’ licensing exam with flying colors.

But its abilities go beyond providing correct answers. This tool offers doctors a partner in care, helping them to work through complex situations and ultimately make the best decisions for their patients.

Called Semantic Clinical Artificial Intelligence (SCAI, pronounced “Sky”), the system is the most accurate clinical AI tool to date. It outperformed most doctors and all other current AI models on the Step 3, the final and most clinical examination of the United States Medical Licensing Exam.

But while the exam results are striking, the real breakthrough is its capacity to reason through complex medical questions.

According to Peter L. Elkin, chair of the Department of Biomedical Informatics at the Jacobs School of Medicine and Biomedical Sciences, and lead author of the study published in JAMA Network Open, most AI tools function by using statistics to find associations in online data that allow them to answer a question.

“But SCAI is different,” he said. “It answers more complex questions and performs more complex semantic reasoning.”

The UB team built SCAI on natural language processing software they had previously developed, then expanded it with vast amounts of authoritative medical data, including clinical guidelines, medical literature, genomic data and more.

At its core, SCAI contains 66 million medical facts represented as semantic triples—basic statements such as “Penicillin treats pneumococcal pneumonia.” These are linked into large semantic networks that allow the tool to draw logical inferences. By combining this structured knowledge with techniques like knowledge graphs and retrieval-augmented generation, SCAI can uncover hidden patterns, check external sources before responding, and reduce the tendency to “make up” answers that plagues other AI models.

In order to limit bias, another problem that plagues AI medical tools, UB researchers didn’t use data like clinical notes, which can contain biases. They’re also continuing research to refine the software, by detecting and eliminating bias in the recommendations SCAI provides. This involves developing and deploying a risk mitigation strategy for each type of bias.   

In early results, the researchers have been able to reverse racial bias by inserting specific instructions into the large language model (LLM) prompt. Developing a set of bias-associated risk mitigation strategies has the potential to improve the safety of LLM usage in clinical settings.

Despite its power, Elkin stresses that SCAI is not designed to replace physicians; it’s meant to serve them.

“SCAI can have a conversation with you,” he said. “It can add to your decision-making and thinking based on its own reasoning.”

The bottom line, Elkin said, is collaboration. “Artificial intelligence isn’t going to replace doctors—but a doctor who uses AI may replace a doctor who does not.”