RESEARCH · RESEARCH · #586
LearnActCoder (Learn-Then-Act): role-aware error memory boosts CPT coding F1 on MIMIC-III
arXiv:2609.19721v1 introduces Learn‑Then‑Act, an inference‑time adaptation framework that converts a small labeled LEARN batch into a structured Mistake Knowledge Database (MistakeKDB) and routes lessons to a recall‑oriented Coder or precision‑oriented Judge. Instantiated as LearnActCoder with lookup grounding, the method raises CPT/HCPCS F1 by 5.9 percentage points on 150 matched MIMIC‑III notes, shifts ICD‑10 coding toward higher precision on a MIMIC‑IV cohort (with recall loss), and preserves a stable ICD operating point on 1,000 held‑out MIMIC‑III notes; ICD‑9/ICD‑10 F1 changes are mixed and overall CPT/HCPCS absolute performance remains low; evaluation is retrospective.
KEY POINTS
- arXiv:2609.19721v1 introduces Learn‑Then‑Act, an inference‑time adaptation framework that converts a small labeled LEARN batch into a structured Mistake Knowledge Database (MistakeKDB) and routes lessons to a recall‑oriented Coder or precision‑oriented Judge.
- Instantiated as LearnActCoder with lookup grounding, the method raises CPT/HCPCS F1 by 5.9 percentage points on 150 matched MIMIC‑III notes, shifts ICD‑10 coding toward higher precision on a MIMIC‑IV cohort (with recall loss), and preserves a stable ICD operating point on 1,000 held‑out MIMIC‑III notes; ICD‑9/ICD‑10 F1 changes are mixed and overall CPT/HCPCS absolute performance remains low; evaluation is retrospective.
- Demonstrates that a structured, role‑aware error memory applied at inference time can adapt clinical coding behavior and improve CPT coding metrics without model retraining or workflow changes.
WHY IT MATTERS
Demonstrates that a structured, role‑aware error memory applied at inference time can adapt clinical coding behavior and improve CPT coding metrics without model retraining or workflow changes.