Researchers at the University of California, San Francisco (UCSF) have developed HACHI, an AI framework that enables clinicians to construct transparent, interpretable clinical prediction models through human-AI collaboration. Published in npj Digital Medicine in 2026, the HACHI methodology addresses significant limitations in current medical AI, particularly the difficulty healthcare providers face in trusting complex, opaque algorithms. The framework, which stands for Human+Agent Co-design for Healthcare Instruments, utilizes large language models combined with a continuous human-in-the-loop workflow to blend rapid data analysis with expert clinical judgment. Under the guidance of Jean Feng, Ph.D., associate professor of epidemiology and biostatistics at UCSF, the system assigns roles based on the distinct capabilities of AI and human experts. AI agents process vast volumes of medical records to surface potential risk factors and predictive patterns. Clinicians subsequently review these outputs to eliminate bias, correct errors, and verify that findings align with medical reality. This iterative process, named after Hachikō to symbolize learning through feedback, aims to generate simple, explainable prediction models rather than black-box systems. "We aim to design AI agents to collaboratively work with clinicians and data scientists," Feng noted. "Together, they can build better tools than any group could alone." Validation studies demonstrated HACHI's effectiveness in real-world clinical challenges. In a pediatric assessment for traumatic brain injury (TBI), the framework produced a five-factor model based on signs and symptoms to predict TBI outcomes following head trauma. By focusing on high-value risk factors and removing misleading signals, the model outperformed existing prediction methods. In a separate study on acute kidney injury (AKI) in adults undergoing surgery, HACHI identified both known and previously unrecognized risk factors, improving predictive accuracy across multiple time intervals. The framework also offers substantial gains in development speed. While constructing robust prediction models traditionally consumes months of effort, the UCSF team reported that effective models could be created in fewer than eight hours through just three or four rounds of clinician feedback. Researchers intend to evaluate HACHI-generated models within actual clinical settings and expand the approach to address a broader range of medical conditions. This development marks a significant step toward accelerating the creation of reliable, practical AI instruments that clinicians can confidently integrate into patient care.