---
title: "Interpretable ICD Code Classification with Faithful Sentence"
description: "The authors design a classifier whose decision for each code depends, by construction, only on a few sentences. 19 references."
authors: ["Yichen Wang", "Lian Hong", "Masato Mizogaki", "Shunnosuke Umeda", "Toshimune Kenmotsu", "Akihiro Tamura", "Daniel Andrade"]
year: 2026
venue: "BioNLP 2026"
source: "https://aclanthology.org/2026.bionlp-1.54/"
doi: "10.18653/v1/2026.bionlp-1.54"
licence: "CC BY 4.0 (ACL Anthology)"
licence_url: "https://creativecommons.org/licenses/by/4.0/"
summary_licence: "Elamur Cosmic License"
canonical: "https://research.elamur.ai/free/en/papers/wang-2026-interpretable-icd-code-classification-faithful/"
---
[Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/papers/wang-2026-interpretable-icd-code-classification-faithful/)

[home](https://md.research.elamur.ai/free/en/) / [papers](https://md.research.elamur.ai/free/en/papers/) / [2026](https://md.research.elamur.ai/free/en/papers/2026/page/33/) / Interpretable ICD Code Classification with Faithful Sentence Extraction

# Interpretable ICD Code Classification with Faithful Sentence Extraction

**Authors:** [Yichen Wang](https://md.research.elamur.ai/free/en/authors/yichen-wang/), Lian Hong, Masato Mizogaki, Shunnosuke Umeda, Toshimune Kenmotsu, [Akihiro Tamura](https://md.research.elamur.ai/free/en/authors/akihiro-tamura/) and Daniel Andrade · **Year:** 2026 · **Venue:** [BioNLP 2026](https://md.research.elamur.ai/free/en/venues/bionlp/) ([BioNLP 2026](https://md.research.elamur.ai/free/en/venues/bionlp/2026/)) · **Publisher:** [Association for Computational Linguistics](https://md.research.elamur.ai/free/en/publishers/association-for-computational-linguistics/) · **Licence:** [CC BY 4.0 (ACL Anthology)](https://creativecommons.org/licenses/by/4.0/)

**Links:** [Publisher's page](https://aclanthology.org/2026.bionlp-1.54/) · [PDF](https://aclanthology.org/2026.bionlp-1.54.pdf) · [DOI 10.18653/v1/2026.bionlp-1.54](https://doi.org/10.18653/v1/2026.bionlp-1.54) · [OpenAlex](https://openalex.org/W7166813676)

**Citations:** 0 citations (OpenAlex, 2026-10-14)

## Summary
Transformer models such as PLM-CA lead automatic ICD coding of discharge summaries, but their attention weights are not reliable explanations. Nothing ensures that text outside the highlighted parts is irrelevant to a prediction. The authors design a classifier whose decision for each code depends, by construction, only on a few sentences. Each sentence is embedded independently, either by an LLM or as a bag-of-words vector. A linear layer then scores each sentence per code, and the top-k sentences are mean-pooled and rescored to give that code's logit. Because selection never looks at the rest of the document, removing unselected sentences provably cannot change the output. The encoder and classifier are trained jointly without any rationale annotations, first on all sentences and then with top-k selection switched on. On MIMIC-III, keeping 10 sentences per code drops more than 90% of most records. With this setting, LLM-SeparateTopK beats every extraction baseline and even full-document LLM pooling, though it still trails full-text PLM-CA, especially on F1. The result is short, checkable evidence for each assigned code.

## Key points
- **Data and evaluation:** MIMIC-III discharge summaries with the MIMIC-III clean splits of Edin et al. (2023), segmented with NLTK (median of 141 sentences per document). Models are evaluated on 3,681 ICD codes.
- **Variants:** LLM-SeparateTopK picks k sentences per code, while LLM-GlobalTopK picks one shared set of k sentences. BoW-SeparateTopK and BoW-GlobalTopK use n-gram features (1–3-grams appearing in at least 3 sentences) instead of an LLM.
- **Main result (k = 10):** LLM-SeparateTopK reaches micro/macro AUC-ROC of 98.2/94.2, micro F1 of 51.7 and MAP of 55.5. Full-document PLM-CA reaches 98.9/95.7, 59.8 and 65.0. On macro F1 the gap is larger: 18.3 vs 29.0.
- **Baselines beaten:** the proposed model outperforms IG-GlobalTopK (Integrated Gradients sentence selection plus retrained PLM-CA; macro AUC 86.1) and head/tail extraction. It also beats LLM-All (macro AUC 85.0), suggesting that sentence selection filters out noise.
- **Limitations and future work:** sentences are scored without context, which faithfulness requires but may miss short, context-dependent sentences. The authors plan to compare the extracted rationales with human evidence in MDACE.

*Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.*

## Institutions

- [Doshisha University](https://md.research.elamur.ai/free/en/universities/doshisha-university/) — University, Kyoto, Japan · 3 works · 149 citations
- [Hiroshima University](https://md.research.elamur.ai/free/en/universities/hiroshima-university/) — University, Hiroshima, Japan · 2 works

📚 Shelves (4): [🖥️ Computer Science](https://md.research.elamur.ai/free/en/shelves/computer-science/) · 16,667 works · [🤖 Artificial Intelligence](https://md.research.elamur.ai/free/en/shelves/computer-science/artificial-intelligence/) · 16,194 works · [🏥 Health Professions](https://md.research.elamur.ai/free/en/shelves/health-professions/) · 69 works · [🏥 Health Information Management](https://md.research.elamur.ai/free/en/shelves/health-professions/health-information-management/) · 50 works

## Cite

Cite as: Yichen Wang, Lian Hong, Masato Mizogaki, Shunnosuke Umeda, Toshimune Kenmotsu, Akihiro Tamura and Daniel Andrade. 2026. Interpretable ICD Code Classification with Faithful Sentence Extraction. BioNLP 2026.

The paper itself, not this page. BibTeX from the library's record of the work.

```bibtex
@inproceedings{wang2026interpretable,
  title = {{Interpretable ICD Code Classification with Faithful Sentence Extraction}},
  author = {Yichen Wang, Lian Hong, Masato Mizogaki, Shunnosuke Umeda, Toshimune Kenmotsu, Akihiro Tamura and Daniel Andrade},
  booktitle = {BioNLP 2026},
  year = {2026},
  doi = {10.18653/v1/2026.bionlp-1.54},
  url = {https://doi.org/10.18653/v1/2026.bionlp-1.54}
}
```

- [← previous Interlinear Glosses as a Multilingual Pivot for Machine Translation: An Updated Study on Turkish with Restricted Resources](https://md.research.elamur.ai/free/en/papers/ozer-2026-interlinear-glosses-as-multilingual-pivot/)
- [next → Interpretable Safety Alignment via SAE-Constructed Low-Rank Subspace Adaptation](https://md.research.elamur.ai/free/en/papers/wang-2026-interpretable-safety-alignment-sae-constructed/)
- [↑ back to papers of 2026](https://md.research.elamur.ai/free/en/papers/2026/page/33/)

## References

19 references, 4 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

1. Ilyas Aden, Christopher HT Child, and Constantino Carlos Reyes-Aldasoro. 2024. International classification of diseases prediction from mimiic-iii clinical text using pre-trained clinicalbert and nlp deep learning models achieving state of the art. Big Data and Cognitive Computing, 8(5):47. [find](https://scholar.google.com/scholar?q=Ilyas+Aden%2C+Christopher+HT+Child%2C+and+Constantino+Carlos+Reyes%2DAldasoro%2E+2024%2E+International+classification+of+diseases+prediction+from+mimiic%2Diii+clinical+text+using+pre%2Dtrained+clinicalbert+and+nlp+)
2. Hua Cheng, Rana Jafari, April Russell, Russell Klopfer, Edmond Lu, Benjamin Striner, and Matthew Gormley. 2023. MDACE: MIMIC documents annotated with code evidence. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 7534–7550, To… [find](https://scholar.google.com/scholar?q=Hua+Cheng%2C+Rana+Jafari%2C+April+Russell%2C+Russell+Klopfer%2C+Edmond+Lu%2C+Benjamin+Striner%2C+and+Matthew+Gormley%2E+2023%2E+MDACE%3A+MIMIC+documents+annotated+with+code+evidence%2E+In+Proceedings+of+the+61st+Annual+M)
3. Hang Dong, Matúš Falis, William Whiteley, Beatrice Alex, Joshua Matterson, Shaoxiong Ji, Jiaoyan Chen, and Honghan Wu. 2022. Automated clinical coding: what, why, and where we are? NPJ digital medicine, 5(1):159. [find](https://scholar.google.com/scholar?q=Hang+Dong%2C+Matu%CC%81s%CC%8C+Falis%2C+William+Whiteley%2C+Beatrice+Alex%2C+Joshua+Matterson%2C+Shaoxiong+Ji%2C+Jiaoyan+Chen%2C+and+Honghan+Wu%2E+2022%2E+Automated+clinical+coding%3A+what%2C+why%2C+and+where+we+are%3F+NPJ+digital+medic)
4. James C. Douglas, Yidong Gan, Ben Hachey, and Jonathan K. Kummerfeld. 2025. Less is more: Explainable and efficient ICD code prediction with clinical entities. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 30835–30847, Vienn… [find](https://scholar.google.com/scholar?q=James+C%2E+Douglas%2C+Yidong+Gan%2C+Ben+Hachey%2C+and+Jonathan+K%2E+Kummerfeld%2E+2025%2E+Less+is+more%3A+Explainable+and+efficient+ICD+code+prediction+with+clinical+entities%2E+In+Proceedings+of+the+63rd+Annual+Meetin)
5. Joakim Edin, Alexander Junge, Jakob D Havtorn, Lasse Borgholt, Maria Maistro, Tuukka Ruotsalo, and Lars Maaløe. 2023. Automated medical coding on mimic-iii and mimic-iv: a critical review and replicability study. In Proceedings of the 46th international ACM SIGIR conference on research and developme… [find](https://scholar.google.com/scholar?q=Joakim+Edin%2C+Alexander+Junge%2C+Jakob+D+Havtorn%2C+Lasse+Borgholt%2C+Maria+Maistro%2C+Tuukka+Ruotsalo%2C+and+Lars+Maal%C3%B8e%2E+2023%2E+Automated+medical+coding+on+mimic%2Diii+and+mimic%2Div%3A+a+critical+review+and+replicab)
6. Joakim Edin, Maria Maistro, Lars Maaløe, Lasse Borgholt, Jakob Drachmann Havtorn, and Tuukka Ruotsalo. 2024. An unsupervised approach to achieve supervised-level explainability in healthcare records. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 486… [find](https://scholar.google.com/scholar?q=Joakim+Edin%2C+Maria+Maistro%2C+Lars+Maal%C3%B8e%2C+Lasse+Borgholt%2C+Jakob+Drachmann+Havtorn%2C+and+Tuukka+Ruotsalo%2E+2024%2E+An+unsupervised+approach+to+achieve+supervised%2Dlevel+explainability+in+healthcare+records%2E+)
7. Upol Ehsan, Pradyumna Tambwekar, Larry Chan, Brent Harrison, and Mark O Riedl. 2019. Automated rationale generation: a technique for explainable ai and its effects on human perceptions. In Proceedings of the 24th international conference on intelligent user interfaces, pages 263–274. [find](https://scholar.google.com/scholar?q=Upol+Ehsan%2C+Pradyumna+Tambwekar%2C+Larry+Chan%2C+Brent+Harrison%2C+and+Mark+O+Riedl%2E+2019%2E+Automated+rationale+generation%3A+a+technique+for+explainable+ai+and+its+effects+on+human+perceptions%2E+In+Proceedings)
8. Shi Feng, Eric Wallace, Alvin Grissom II, Mohit Iyyer, Pedro Rodriguez, and Jordan Boyd-Graber. 2018. Pathologies of neural models make interpretations difficult. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 3719–3728. [find](https://scholar.google.com/scholar?q=Shi+Feng%2C+Eric+Wallace%2C+Alvin+Grissom+II%2C+Mohit+Iyyer%2C+Pedro+Rodriguez%2C+and+Jordan+Boyd%2DGraber%2E+2018%2E+Pathologies+of+neural+models+make+interpretations+difficult%2E+In+Proceedings+of+the+2018+Conference)
9. Elize Herrewijnen, Dong Nguyen, Jelte Mense, Floris Bex, and 1 others. 2021. Machine-annotated rationales: faithfully explaining text classification. In Proceedings for the Explainable Agency in AI Workshop at the 35th AAAI Conference on Artificial Intelligence (Washington DC: AAAI Press), pages 11–… [find](https://scholar.google.com/scholar?q=Elize+Herrewijnen%2C+Dong+Nguyen%2C+Jelte+Mense%2C+Floris+Bex%2C+and+1+others%2E+2021%2E+Machine%2Dannotated+rationales%3A+faithfully+explaining+text+classification%2E+In+Proceedings+for+the+Explainable+Agency+in+AI+Wo)
10. Chao-Wei Huang, Shang-Chi Tsai, and Yun-Nung Chen. 2022. PLM-ICD: Automatic ICD coding with pretrained language models. In Proceedings of the 4th Clinical Natural Language Processing Workshop, pages 10–20, Seattle, WA. Association for Computational Linguistics. [find](https://scholar.google.com/scholar?q=Chao%2DWei+Huang%2C+Shang%2DChi+Tsai%2C+and+Yun%2DNung+Chen%2E+2022%2E+PLM%2DICD%3A+Automatic+ICD+coding+with+pretrained+language+models%2E+In+Proceedings+of+the+4th+Clinical+Natural+Language+Processing+Workshop%2C+pages+1)
11. Alistair EW Johnson, Tom J Pollard, Lu Shen, Li-wei H Lehman, Mengling Feng, Mohammad Ghassemi, Benjamin Moody, Peter Szolovits, Leo Anthony Celi, and Roger G Mark. 2016. [Mimic-iii, a freely accessible critical care database](https://md.research.elamur.ai/free/en/papers/johnson-2016-mimic-iii-freely-accessible-critical/). Scientific data, 3(1):1–9.
12. Patrick Lewis, Myle Ott, Jingfei Du, and Veselin Stoyanov. 2020. [Pretrained language models for biomedical and clinical tasks: Understanding and extending the state-of-the-art](https://md.research.elamur.ai/free/en/papers/lewis-2020-pretrained-language-models-biomedical-clinical/). In Proceedings of the 3rd Clinical Natural Language Processing Workshop, pages 146–157, Online. Association for Computation…
13. Mukund Sundararajan, Ankur Taly, and Qiqi Yan. 2017. Axiomatic attribution for deep networks. In Proceedings of the 34th International Conference on Machine Learning, volume 70 of Proceedings of Machine Learning Research, pages 3319–3328. [find](https://scholar.google.com/scholar?q=Mukund+Sundararajan%2C+Ankur+Taly%2C+and+Qiqi+Yan%2E+2017%2E+Axiomatic+attribution+for+deep+networks%2E+In+Proceedings+of+the+34th+International+Conference+on+Machine+Learning%2C+volume+70+of+Proceedings+of+Machi)
14. Tong Zhou, Pengfei Cao, Yubo Chen, Kang Liu, Jun Zhao, Kun Niu, Weifeng Chong, and Shengping Liu. 2021. Automatic icd coding via interactive shared representation networks with self-distillation mechanism. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and… [find](https://scholar.google.com/scholar?q=Tong+Zhou%2C+Pengfei+Cao%2C+Yubo+Chen%2C+Kang+Liu%2C+Jun+Zhao%2C+Kun+Niu%2C+Weifeng+Chong%2C+and+Shengping+Liu%2E+2021%2E+Automatic+icd+coding+via+interactive+shared+representation+networks+with+self%2Ddistillation+mecha)
15. A First, each document is split into sentences using the sent tokenize function 3 of the NLTK library (Bird et al., 2009). Afterwards, the following post-processing steps are used to reduce conflation and improve comprehensiveness: 1. If there are two or more line breaks, split the sentence before a… [find](https://scholar.google.com/scholar?q=A+First%2C+each+document+is+split+into+sentences+using+the+sent+tokenize+function+3+of+the+NLTK+library+%28Bird+et+al%2E%2C+2009%29%2E+Afterwards%2C+the+following+post%2Dprocessing+steps+are+used+to+reduce+conflation)
16. Qing Lyu, Marianna Apidianaki, and Chris Callison-Burch. 2024. [Towards faithful model explanation in NLP: A survey](https://md.research.elamur.ai/free/en/papers/lyu-2024-faithful-model-explanation-nlp-survey/). Computational Linguistics, 50(2):657–723. 2. If there are two or more spaces, split the sentence before and after the spaces. 3. Replace the line break characters with spaces.
17. Shotaro MISAWA, Taiki FURUKAWA, Shintaro OYAMA, Ryuji KANO, Hirokazu YARIMIZU, Tomoki TANIGUCHI, Kohei ONODA, Kikue SATO, and Yoshimune SHIRATORI. 2023. Sentence extraction using outcome prediction model trained from clinical data. Proceedings of the Annual Conference of JSAI, JSAI2023:3Xin404– 3Xin… [find](https://scholar.google.com/scholar?q=Shotaro+MISAWA%2C+Taiki+FURUKAWA%2C+Shintaro+OYAMA%2C+Ryuji+KANO%2C+Hirokazu+YARIMIZU%2C+Tomoki+TANIGUCHI%2C+Kohei+ONODA%2C+Kikue+SATO%2C+and+Yoshimune+SHIRATORI%2E+2023%2E+Sentence+extraction+using+outcome+prediction+mo)
18. James Mullenbach, Sarah Wiegreffe, Jon Duke, Jimeng Sun, and Jacob Eisenstein. 2018. [Explainable prediction of medical codes from clinical text](https://md.research.elamur.ai/free/en/papers/mullenbach-2018-explainable-prediction-medical-codes-clinical/). In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1…
19. Ajay Madhavan Ravichandran, Julianna Grune, Nils Feldhus, Aljoscha Burchardt, Roland Roller, Sentence Splitting 4. Remove spaces at the beginning and end of the sentence. 5. If each sentence ends with a ”:”, join that sentence to the next sentence with a space. The resulting median number of sentenc… [find](https://scholar.google.com/scholar?q=Ajay+Madhavan+Ravichandran%2C+Julianna+Grune%2C+Nils+Feldhus%2C+Aljoscha+Burchardt%2C+Roland+Roller%2C+Sentence+Splitting+4%2E+Remove+spaces+at+the+beginning+and+end+of+the+sentence%2E+5%2E+If+each+sentence+ends+with)

*Updated 1 November 2026*