Artificial intelligence has shown significant promise in healthcare, helping predict conditions ranging from sepsis and heart disease to cancer. Yet many AI-driven clinical prediction tools struggle to gain traction in everyday practice. A common challenge is that clinicians often find these systems difficult to interpret, limiting trust and adoption.
Researchers at the University of California, San Francisco (UCSF) believe they may have found a solution. Their newly developed framework, called HACHI, combines the analytical power of AI with the expertise of healthcare professionals to create prediction models that are both accurate and clinically meaningful. The approach, recently published in npj Digital Medicine, aims to bridge the gap between advanced machine learning and real-world clinical decision-making.
Combining AI and human expertise
The new framework, formally named Human+Agent Co-design for Healthcare Instruments (HACHI), is designed around a simple principle: allow artificial intelligence and clinicians to focus on what each does best. AI excels at processing vast quantities of electronic health records and identifying potential patterns linked to disease risk. Clinicians, meanwhile, provide contextual understanding, helping to identify meaningful findings while filtering out misleading associations and potential biases.
“The goal is to design AI agents to collaboratively work with clinicians and data scientists,” said lead author Jean Feng, associate professor of epidemiology and biostatistics at UCSF. “Together, they can build better tools than any group could alone.” Rather than producing complex “black-box” algorithms, HACHI focuses on creating transparent prediction models that clinicians can understand and trust. The framework uses AI to propose risk factors and clinical concepts, while human experts iteratively refine the selection process. The name HACHI is inspired by Hachikō, Japan’s famously loyal dog. According to the researchers, the framework reflects a similar principle of continuous learning through repeated feedback and refinement.
Testing in real clinical scenarios
To evaluate the approach, researchers applied HACHI to two important clinical challenges: predicting traumatic brain injury in children presenting to emergency departments after head trauma, and forecasting acute kidney injury in adults undergoing surgery. In the pediatric brain injury study, HACHI helped develop a concise five-factor prediction model based on key symptoms and clinical signs. By focusing on the most relevant indicators and eliminating distracting variables, the model outperformed conventional approaches in identifying children at risk of traumatic brain injury.
The framework also showed promising results in predicting acute kidney injury, a potentially serious complication following surgery. In this case, HACHI identified both established risk factors and previously overlooked clinical indicators, improving predictive performance across different patient populations and time periods. These findings suggest that combining AI-driven discovery with human clinical judgment can lead to more robust and generalizable models than relying on either approach alone.
Accelerating clinical model development
One of the most notable advantages of the HACHI framework is speed. Developing clinical prediction tools traditionally requires months of work involving data scientists, clinicians and extensive model validation. With HACHI, AI automatically generates and evaluates potential risk factors from clinical notes and patient records. Clinicians then review the findings and provide feedback.
Researchers found that after only three to four rounds of iteration, requiring less than eight hours of combined effort, teams were able to develop strong predictive models. This streamlined process could significantly reduce development timelines while maintaining clinical relevance. The UCSF team sees HACHI as part of a broader shift toward human-centered AI in healthcare. Rather than replacing clinical expertise, the framework positions AI as a collaborative partner that enhances decision-making.
Future research will focus on testing HACHI-generated models in real-world healthcare environments and expanding the framework to additional medical conditions. If successful, the approach could help accelerate the development of transparent, reliable and clinically trusted AI tools across multiple specialties. As healthcare systems increasingly adopt artificial intelligence, frameworks such as HACHI may offer a practical path forward. Thus combining machine efficiency with human judgment to deliver tools that clinicians are willing to use and patients can ultimately benefit from.
Turning ideas into AI tools
In an interview with ICT&health, Michał Nedoszytko, a cardiologist at Cliniques de l’Europe in Belgium and a long-time developer, recently talked about AI’s most immediate impact on healthcare. He believes that it may come from fixing the workflows that keep doctors away from patients. About four years ago, he became deeply involved in artificial intelligence. In invasive cardiology, his specialty, He started training neural networks to recognize specific changes in coronary angiography images. At the time, that was still relatively novel. In medicine, AI was largely confined to radiology and image analysis.
When ChatGPT arrived, he immediately saw its potential for a very different part of healthcare: the administrative layer. AI could finally help with the “gray zone” of clinical work, the tasks that are essential but take doctors away from patients. That was the starting point for his previsit project, a system designed to collect a patient’s medical history before the appointment. The idea was simple: streamline part of the history-taking process before the visit even begins. It turned out to work very well.
