Low-Resource Automatic ICD Coding for Chinese Electronic Medical Records: An Interpretable Multi-View Linear Framework with Audited Clinical Rules
- 1Zhejiang Provincial Center for Medical Science, Technology and Education Development
Abstract
ICD coding is the backbone of medical-record structuring, insurance payment, and clinical research. Todays best automatic coders are deep models, but they assume tens of thousands of labeled records, GPU servers, and machine-learning engineers. Many hospitals, especially small ones, have none of these. We study Chinese electronic medical records (EMRs) under a hard budget: 2,200 labeled records, CPU only, and fully interpretable decisions. Our method is lightweight and follows the way human coders work. First, codebook knowledge injection marks standard disease and operation names in the text as useful clues. Second, multi-view field weighting gives more weight to the clinical sections that matter most. Third, character n-gram TF-IDF features feed linear classifiers that score each candidate code. Fourth, a small set of clinical rules acts as a final safety check. Our main methodological finding is about this last step: naively applied rules do more harm than good (overall F1 drops from 0.8070 to 0.7995, and main-operation accuracy from 0.9727 to 0.9432). We therefore audit each rule on held-out data by counting how often it fixes a wrong prediction or breaks a correct one, and we remove rules that do more harm than good. This turns rule post-processing into a safety check that never hurts. On the CCL2026-Eval Task 6 dataset, our method reaches macro-F1 of 0.899 (main diagnosis) and 0.966 (main operation) in 5-fold cross-validation. A frequency-stratified analysis shows a clear long-tail limit: codes seen fewer than three times score zero F1. In a head-to-head test, a fine-tuned Chinese pretrained encoder scores lower than our method on all four tasks at this training size, while needing a GPU. Ablation tests show that the gains come mainly from field-weighted text. The framework trains in about 40 minutes on an ordinary CPU, and every prediction can be broken down into interpretable feature contributions.
Related
- Analytic Interoperability Framework for High-Fidelity Representation of Perioperative Electronic Health Record Data in the OMOP Common Data Model Using Clinical Representation and Analytic Concept Reconstruction
- From Clinical Free Text to Auditable Concepts: An Agentic Framework for Interpretable Prediction
- Interpretable deep learning framework for secure clinical data analytics
- Scalable Causal-Interpretable Machine Learning for Cancer Prescreening Using Electronic Health Records
- Hepatitis B and Sharps Safety Before Nursing Clinical Placement: A Narrative Review and Proposed Framework for Resource-Constrained Hospitals
BibTeX
@misc{xu2026lowresource,
title = {{Low-Resource Automatic ICD Coding for Chinese Electronic Medical Records: An Interpretable Multi-View Linear Framework with Audited Clinical Rules}},
author = {Y. Xu and J. Yan},
howpublished = {medRxiv},
year = {2026},
month = oct,
doi = {10.64898/2026.10.04.26364688},
url = {https://www.medrxiv.org/content/10.64898/2026.10.04.26364688v1},
note = {CC BY-NC-ND},
}
RIS
TY - UNPB
TI - Low-Resource Automatic ICD Coding for Chinese Electronic Medical Records: An Interpretable Multi-View Linear Framework with Audited Clinical Rules
AU - Y. Xu
AU - J. Yan
PY - 2026
DA - 2026/10/04
PB - medRxiv
DO - 10.64898/2026.10.04.26364688
UR - https://www.medrxiv.org/content/10.64898/2026.10.04.26364688v1
AB - ICD coding is the backbone of medical-record structuring, insurance payment, and clinical research. Todays best automatic coders are deep models, but they assume tens of thousands of labeled records, GPU servers, and machine-learning engineers. Many hospitals, especially small ones, have none of these. We study Chinese electronic medical records (EMRs) under a hard budget: 2,200 labeled records, CPU only, and fully interpretable decisions. Our method is lightweight and follows the way human coders work. First, codebook knowledge injection marks standard disease and operation names in the text as useful clues. Second, multi-view field weighting gives more weight to the clinical sections that matter most. Third, character n-gram TF-IDF features feed linear classifiers that score each candidate code. Fourth, a small set of clinical rules acts as a final safety check. Our main methodological finding is about this last step: naively applied rules do more harm than good (overall F1 drops from 0.8070 to 0.7995, and main-operation accuracy from 0.9727 to 0.9432). We therefore audit each rule on held-out data by counting how often it fixes a wrong prediction or breaks a correct one, and we remove rules that do more harm than good. This turns rule post-processing into a safety check that never hurts. On the CCL2026-Eval Task 6 dataset, our method reaches macro-F1 of 0.899 (main diagnosis) and 0.966 (main operation) in 5-fold cross-validation. A frequency-stratified analysis shows a clear long-tail limit: codes seen fewer than three times score zero F1. In a head-to-head test, a fine-tuned Chinese pretrained encoder scores lower than our method on all four tasks at this training size, while needing a GPU. Ablation tests show that the gains come mainly from field-weighted text. The framework trains in about 40 minutes on an ordinary CPU, and every prediction can be broken down into interpretable feature contributions.
ER -