Distilling Expert-level Planning for Real-world Diabetes Prescribing

Yein Park, Sihyeon Park, Taeyun Roh, Mujeen Sung, Jaehoon Yun, Sooyeon Lee, Sanghoon Lee, Jaewoo Kang

Keywords: Real-world Clinical AI, Trustworthy Medical AI, Diabetes Prescribing, Medical Expert Planning, EMR Benchmark, Reimbursement-Aware Evaluation

TL;DR: We present a LLM-based clinical agent for real-world diabetes prescribing that learns expert treatment planning, uses dynamic medical search, and is evaluated on an EMR-derived reimbursement-aware benchmark reflecting real clinical constraints.

Abstract:

Real-world prescribing is constrained not only by clinical appropriateness but also by reimbursement policy, prior treatment history, and patient preference, which remain under-modeled in exam-style medical QA training and evaluation.
We present a practical framework for expert-usable clinical agents for diabetes prescribing that combines expert planning distillation-distilling an expert prescribing workflow from patient assessment and safety checks to candidate regimen selection and reimbursement verification-with dynamic search over guidelines, reimbursement criteria, drug information, and up-to-date references.
To evaluate real-world prescribability, we construct an EMR-derived diabetes prescribing benchmark from 70 de-identified real-world patient cases and evaluate on 140 test instances with a clinician-aligned, reimbursement-aware rubric.
On this benchmark, our expert planning distillation and dynamic search improve reimbursement-aware prescribing quality over general and biomedical baselines and approach the performance of frontier proprietary models.

Submission Number: 33