AI & ML
AI-Assisted Hypothesis Generation for Hospital Quality Improvement
arXiv cs.HC0 views
arXiv:2604.20055v2 Announce Type: replace-cross Abstract: Hospital quality improvement (QI) programs generally proceed in two stages: (1) hypothesis generation, which identifies candidate factors that may improve care quality and efficiency, and (2) confirmatory analyses, which implement and evaluate QI initiatives targeting these factors. Although most research focuses on the confirmatory stage, hypothesis generation is equally important because it shapes where teams focus their effort. Currently, hypothesis generation depends on expert-driven, qualitative methods (e.g., Lean Healthcare tools) that can be anecdotal, hard to reproduce, time-consuming, and resource-intensive. Large language models (LLMs) could accelerate this process by analyzing clinical notes at scale, but existing AI tools can only scale up well-defined tasks, whereas hypothesis generation is a "fuzzy," exploratory, and iterative process. We propose treating not only LLM instructions but also the problem specification and validation process as tunable natural-language (hyper)parameters. Through a procedure we term "Human-AI Spec-Solution Co-Optimization," we iteratively elicit clinical experts' definitions and judgments while aligning the AI pipeline with their intended criteria. We retrospectively applied this framework at an urban safety-net hospital to length of stay and 30-day unplanned readmissions. The final pipelines screen patient charts for potential factors and filter for clinically relevant ones by self-rating, and these self-ratings achieved over 70% concordance with human experts (within one Likert point) on held-out encounters. When scaled up hospital-wide, the pipelines identified a superset of the candidate factors found in prior, primarily manual analyses. By providing auditable reasoning traces and quantifying factor prevalence, the pipelines support a more quantitative and reproducible way to prioritize factors for prospective QI interventions.
Source:https://arxiv.org/abs/2604.20055PreviousDoes AI Assistance Preserve or Collapse Disagreement? A Study of Pre-Annotations in Ambiguous Video LabelingNextA human-in-the-loop Bayesian optimization framework for constraint-aware bioprocess development
Related posts
Trying to get our Arms around AI efficiency
0Stack Overflow Blog
Why an Embedded Game Fails on Some Phones, and How to Debug It
0dev.to
Microsoft 365 vs Office 2024: The Break-Even Math
0dev.to
Master AI Evals: Build Reliable, Scalable Model Evaluation Pipelines
0dev.to
What to give someone who misses a place
0dev.to
Google is pushing OEMs against GrapheneOS and bootloader unlocking
0Hacker News