Artificial intelligence can help predict a patient’s risk for conditions such as sepsis, heart disease, and cancer. But many of these tools fall short in real-life clinical practice because they are difficult for doctors to interpret and trust.
Researchers at UC San Francisco have developed a new way to use AI to build clinical prediction tools that combines the speed of artificial intelligence with the judgment of human experts.
The approach, described in a paper published on June 6, 2026 in the Nature Portfolio journal npj Digital Medicine, offers a new model for developing these tools that pairs AI’s ability to rapidly analyze medical records with the expertise of clinicians who can identify bias, spot errors, and ensure the results make sense in practice.
The framework, called HACHI, short for Human+Agent Co-design for Healthcare Instruments, divides the work between AI and clinicians, allowing each to do what they do best. HACHI uses AI to sift through large volumes of medical records in search of predictive clues, while human experts help determine which findings are meaningful enough to include in the prediction model.
“The goal is to design AI agents to collaboratively work with clinicians and data scientists,” said the paper’s lead author, Jean Feng, PhD, associate professor of epidemiology and biostatistics at UCSF. “Together, they can build better tools than any group can do could alone.”
Rather than building complex black-box systems, HACHI uses AI to identify the risk factors and clinical concepts most likely to improve a simple, transparent prediction model. Named after Hachikō, Japan’s famously loyal dog, the framework reflects the power of iterative learning. Just as dogs learn through repeated training and feedback, clinicians continually guide and refine the AI’s work to improve the resulting prediction models.
The researchers found that HACHI outperformed commonly used approaches tested in the study in two real-world clinical challenges: predicting traumatic brain injury in children after head trauma and predicting acute kidney injury in adults undergoing surgery.
To study traumatic brain injuries, the team used HACHI to develop a five-factor model of signs and symptoms to predict whether a child presenting to the emergency department after head trauma would ultimately be diagnosed with a traumatic brain injury. By focusing on the most meaningful risk factors and eliminating misleading signals, the model predicted traumatic brain injuries more accurately than existing methods.
For acute kidney injury, or a sudden decline in kidney function, the HACHI method identified both established and previously overlooked risk factors and improved performance across different time periods.
To build these models, AI identifies and tests potential risk factors in clinical notes, while clinicians review the results and suggest improvements. After just three or four rounds of feedback, or less than eight hours, teams developed strong models, potentially shortening a process that often takes months.
The researchers plan to test HACHI-generated models in real-world clinical settings and expand the framework to other medical conditions. They believe the approach could help accelerate the development of practical, reliable prediction models across health care.
Authors: Additional UCSF authors included Avni Kothari, MS; Andrew Bishara, MD; Lucas Zier, MD; Patrick Vossler, PhD; and Aaron Kornblith, MD.
Funding: This work was supported by the Patient-Centered Outcomes Research Institute (PCORI) (ME-2022C1-25619), the National Institute of General Medical Sciences of the National Institutes of Health (K23GM151611), the Eunice Kennedy Shriver National Institute of Child Health and Human Development of the National Institutes of Health (K23HD110716), and the Zuckerberg Priscilla Chan Quality Improvement Fund through the San Francisco General Foundation.
Disclosures: The authors reported no competing interests.