Reproduce “Human-AI Co-design for Clinical Prediction Models” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Jean Feng, Avni Kothari, Patrick Vossler, Andrew Bishara, Lucas Zier, Newton Addo, Aaron Kornblith, Yan Shuo Tan and 1 more
Researchers from the University of California, San Francisco (UCSF), National University of Singapore, and Microsoft Research developed HACHI, an iterative human-AI co-design framework, to accelerate the development of interpretable clinical prediction models (CPMs). The framework demonstrated improved performance and generalizability for traumatic brain injury (TBI) and acute kidney injury (AKI) prediction tasks, with HACHI's TBI model achieving an AUC of 0.91 compared to PECARN's 0.75.
Key insights
- Iterative human oversight is critical for identifying and correcting issues like data leakage, spurious correlations, and misaligned coefficients, which often arise in purely AI-driven model development.
- AI agents can efficiently surface novel, clinically relevant concepts from unstructured text that are often overlooked by traditional methods or human-only brainstorming.
- The co-design approach effectively bridges the 'gulf of envisionment' by allowing humans to dynamically refine objectives and adjust model specifications based on iterative AI outputs, leading to more robust and clinically credible models.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.