AI Framework Helps Clinicians Create Trustworthy Risk Prediction Tools

Artificial intelligence (AI) is increasingly used to estimate risks for conditions such as sepsis, heart disease, and cancer, yet many models remain difficult for clinicians to interpret or trust. This limits adoption at the bedside and can slow decision-making in emergency and perioperative settings. Hospitals need transparent tools that reflect clinical judgment and reduce bias. To help address this challenge, researchers have developed a framework that combines AI with clinician oversight to create usable prediction models.

The Human+Agent Co-design for Healthcare Instruments (HACHI) framework from the University of California, San Francisco (UCSF) pairs AI agents with clinicians and data scientists. The system analyzes large volumes of electronic medical records to surface candidate predictors. Clinicians then review these suggestions to identify bias, correct errors, and select variables that make sense in practice.

HACHI is designed to build simple, transparent clinical prediction models rather than opaque “black boxes.” AI first scans clinical notes to test potential risk factors and clinical concepts. Clinicians provide iterative feedback in successive rounds, refining the model until it aligns with real-world reasoning and workflow.

In evaluations, HACHI outperformed commonly used approaches on two care challenges. For pediatric head trauma, the framework produced a five-factor model of signs and symptoms that more accurately predicted whether a child presenting to the emergency department would later receive a traumatic brain injury diagnosis. 

For acute kidney injury—defined as a sudden decline in kidney function—in adults undergoing surgery, HACHI identified established and previously overlooked risk factors and maintained improved performance across different time periods.

Model development progressed rapidly. After just three to four feedback cycles requiring less than eight hours, teams produced strong models, potentially compressing a process that often takes months. Published in npj Digital Medicine on June 6, 2026, the work will next be tested in real-world clinical settings, with plans to extend HACHI-generated models to additional conditions.

“The goal is to design AI agents to collaboratively work with clinicians and data scientists. Together, they can build better tools than any group could alone,” said Jean Feng, Ph.D., associate professor of epidemiology and biostatistics at UCSF.

Related Links
UCSF