When a child arrives at the emergency room after a head injury, every minute counts—so does the accuracy of the tools doctors use to decide whether to order a CT scan. At UC San Francisco, a team led by Dr. Jean Feng has created HACHI, an AI framework that’s redefining how clinical prediction tools are built, blending machine speed with human wisdom to deliver faster, more trustworthy results. The name—short for Human+Agent Co-design for Healthcare Instruments—honors Hachikō, the loyal dog, symbolizing the power of learning through consistent feedback and partnership.
Most AI tools in medicine are 'black boxes'—fast but opaque, making clinicians hesitant to rely on them. HACHI flips the script by designing AI to work with doctors, not just for data. Instead of delivering complex algorithms no one can interpret, HACHI uses large language models to scan vast medical records, surface potential risk factors, and then hands those findings to clinicians who validate, refine, and guide the model toward what truly matters in patient care. This human-in-the-loop approach ensures transparency, reduces bias, and builds trust—critical ingredients for tools meant to support life-and-death decisions.
In two rigorous tests, HACHI proved its worth. For pediatric traumatic brain injury, the team developed a five-factor model—based on symptoms like vomiting and loss of consciousness—that outperformed existing tools in predicting which children would be diagnosed with brain injuries. For adults undergoing surgery, HACHI improved the prediction of acute kidney injury by identifying not only known risks like pre-existing kidney disease but also overlooked signals buried in clinical notes. Most strikingly, these models were built in under eight hours over just three or four feedback cycles—a process that typically takes months.
The implications are profound. With HACHI, hospitals could rapidly develop and refine prediction tools tailored to their own patient populations, accelerating innovation without sacrificing reliability. The UCSF team is now preparing to test HACHI-generated models in live clinical environments and expand its use to conditions like sepsis and heart failure.
As AI reshapes medicine, HACHI offers a hopeful vision: not of machines replacing doctors, but of machines learning from them. The future of healthcare innovation may not come from algorithms alone, but from the quiet collaboration between a clinician’s insight and a machine’s memory—iterating, learning, and improving, one patient at a time.