Researchers at the University of Iowa have developed two complementary artificial intelligence approaches that address a central challenge in modern biomedical data science: data that is incomplete and data that is misleading. The studies show that effectively addressing these challenges can enhance the robustness of AI systems.
“These cutting-edge AI techniques will provide biomedical researchers and practitioners with tools for improving patient care and decision-making,” said Kishlay Jha, assistant professor of electrical and computer engineering and senior author on both papers.
The first study, called DiffRisk, focuses on healthcare data. Electronic health records are often incomplete, and most AI systems attempt to fill in missing values before making predictions. DiffRisk takes a different approach, treating missing information as potentially informative. The researchers found that patterns in what clinicians choose not to measure can provide valuable clues about a patient's condition.
By learning from both observed and missing information, DiffRisk improved predictions of outcomes such as patient deterioration, hospital mortality, and length of stay, outperforming existing methods, especially when records were incomplete. Shailesh Dahal, an electrical and computer engineering graduate student, is the first author of this DiffRisk work.
The second study, SPUR-XMC, focuses on the scientific literature rather than patient care. Instead of analyzing missing medical measurements, it helps AI identify rare diseases, genes, and other specialized biomedical concepts within research papers. The system addresses "spurious correlations," where AI learns misleading shortcuts instead of genuine scientific relationships. By integrating large language models with expert-curated medical knowledge, SPUR-XMC improved the ability to correctly classify research articles, particularly those involving uncommon concepts that are often overlooked by conventional systems. Ratri Mukherjee, an electrical and computer engineering graduate student, is the first author of this SPUR-XMC work.
Although the applications differ, the central message of both studies is similar: AI performs better when it moves beyond obvious patterns and pays attention to information that is rare, absent, or easily ignored.
DiffRisk improves clinical prediction by recognizing that missing medical data can carry meaning, while SPUR-XMC improves scientific document analysis by helping AI distinguish truly relevant concepts from misleading associations. Together, these advances suggest a new generation of AI systems that are not only more accurate, but also better equipped to handle the complexity and uncertainty of real-world biomedical data.
Their potential benefits range from earlier identification of patients at risk to faster discovery and organization of scientific knowledge, ultimately helping clinicians and researchers make better-informed decisions.