University of Otago, Dunedin, New Zealand
More than half of the ICD codes in MIMIC appear fewer than ten times in training. Those rare codes are hard to learn, and clinical notes rarely describe them in the wording of the official ICD description: a note says "former smoker" where the code says "personal history of tobacco use".
CoLa-ICD gives each code a richer, clinically worded representation and ties it to the exact span of the note that supports it. It combines UMLS synonyms with LLM-generated clinical terms, shares signal between related codes through a co-occurrence graph, lets each term look for its own evidence in the note, and uses contrastive learning with hard negatives so that look-alike codes stay apart.
All results are averaged over 10 random seeds. Full comparisons with CAML, PLM-ICD, PLM-CA, MSMN, KEPTLongformer and CoRelation are in the paper.
UMLS synonyms plus LLM-generated clinical terms for every code
A PPMI code graph shares signal across related labels
Each term queries the note for its own evidence
Hard negatives keep look-alike codes apart
youwillbethebest/Cola-ICD
Official PyTorch implementation · Python 3.10
git clone https://github.com/youwillbethebest/Cola-ICD.git cd Cola-ICD conda create -n cola-icd python=3.10 conda activate cola-icd pip install -r requirements.txt
@inproceedings{cheng-etal-2026-cola-icd,
title = {{CoLa-ICD}: A Knowledge-Enhanced Framework for Long-Tail Automated Medical Coding},
author = {Cheng, Yihang and Liesaputra, Veronica and Trotman, Andrew},
booktitle = {Findings of the Association for Computational Linguistics: EMNLP 2026},
month = oct,
year = {2026},
address = {Budapest, Hungary},
publisher = {Association for Computational Linguistics}
}