Elamur Research⌕
☰
home / papers / 2026 / CPTCoder: A Reliable LLM System for Medical Procedure Code Prediction

CPTCoder: A Reliable LLM System for Medical Procedure Code Prediction

Paper · 2026 · Benlu Wang, Ziyao Shangguan, Kyle Tegtmeyer et al.

Summary

CPTCoder targets automatic assignment of Current Procedural Terminology (CPT) codes to clinical reports. This is a long-tailed, multi-label task in which codes are five-digit identifiers, so a single wrong digit names a different billable procedure. The authors fine-tune Llama-3.1-8B-Instruct with LoRA. They add "Code-as-Token," which places each of 365 CPT codes in the tokenizer as one atomic token, initialized from both the code string and its official description. Constrained decoding ensures every output is a valid code. Per-code confidence comes from reweighting an n-best list, and a repair step removes mutually exclusive code pairs. The system runs as a vLLM-backed web demo where a coder pastes a report, reviews ranked candidates and sees summed RVU values. The authors also build MIMIC-CPT, a PhysioNet benchmark whose test split is intentionally hard, and evaluate out-of-domain on Hospital-298. CPTCoder beats all baselines on micro-F1 in both settings and makes fewer digit-level near-miss errors. This supports framing CPT coding as controlled structured generation that assists human coders rather than replacing them.

Key points

  • MIMIC-CPT gold set: 38,673 cleaned report–code pairs from MIMIC-IV-Note, split into 30,938 train (350 codes) and 7,735 test (322 codes). A separate silver set of 412,297 weakly aligned pairs from MIMIC-III/IV is released but not used in any experiment.
  • Hardened test split: 88.14% of test reports have a label set never seen in training, and 38.54% contain a code with at most five training occurrences.
  • Hospital-298: 298 de-identified radiology and IR reports with labels from a billing expert, cross-checked by two radiology experts. Models are tested on it with no adaptation.
  • Main results: CPTCoder reaches 0.608 micro-F1 on the MIMIC hard test and 0.509 on Hospital-298, against 0.486 and 0.458 for plain LoRA-tuned LLaMA-3.1-8B. Zero-shot LLMs such as GPT-OSS-120B and MedGemma score close to zero. A TF-IDF SVM baseline slightly exceeds CPTCoder on MIMIC macro-F1.
  • Code-as-Token ablation: the near-miss false-positive rate drops from 26.16% to 17.05%. End-to-end inference usually takes under 10 seconds per report.

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

🎓Stanford UniversityStanford · United States434 works · 314,419 citations🎓Yale UniversityNew Haven · United States95 works · 12,765 citations

 Fields, topics and concepts

📚 Shelves 6The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 worksArts and Humanities · 16,039 worksLanguage and Linguistics · 15,891 worksHealth Professions · 69 worksHealth Information Management · 50 works

Cite

Cite as Benlu Wang, Ziyao Shangguan, Kyle Tegtmeyer, Zhenyu Zhang, Sophie Chheang and Arman Cohan. 2026. CPTCoder: A Reliable LLM System for Medical Procedure Code Prediction. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 3: System Demonstrations).

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.acl-demo.60

@inproceedings{wang2026cptcoder,
  title = {{CPTCoder: A Reliable LLM System for Medical Procedure Code Prediction}},
  author = {Benlu Wang, Ziyao Shangguan, Kyle Tegtmeyer, Zhenyu Zhang, Sophie Chheang and Arman Cohan},
  booktitle = {Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 3: System Demonstrations)},
  year = {2026},
  doi = {10.18653/v1/2026.acl-demo.60},
  url = {https://doi.org/10.18653/v1/2026.acl-demo.60}
}

References

23 references, 12 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. American Medical Association. 2026. find
  2. CPT 2026 Professional Edition. American Medical Association, Chicago, IL. find
  3. Centers for Medicare & Medicaid Services. PFS Relative Value Files. https://www.cms.gov/ medicare/payment/fee-schedules/physician/ pfs-relative-value-files. Accessed 2026-02-27. ↗
  4. Yonatan Geifman and Ran El-Yaniv. 2017. Selective classification for deep neural networks. In Advances in Neural Information Processing Systems 30 (NeurIPS 2017), pages 4878–4887. find
  5. Ary L. Goldberger, Luis A. N. Amaral, Leon Glass, Jeffrey M. Hausdorff, Plamen Ch. Ivanov, Roger G. Mark, Joseph E. Mietus, George B. Moody, Chung-Kang Peng, and H. Eugene Stanley. 2000. Physiobank, physiotoolkit, and physionet: Components of a new research resource for complex physiologic signals.… find
  6. Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, and 1 others. 2024. The llama 3 herd of models. Preprint, arXiv:2407.21783.
  7. Chuan Guo, Geoff Pleiss, Yu Sun, and Kilian Q. Weinberger. 2017. On calibration of modern neural networks. In Proceedings of the 34th International Conference on Machine Learning (ICML 2017), volume 70 of Proceedings of Machine Learning Research, pages 1321–1330. PMLR. find
  8. Chris Hokamp and Qun Liu. 2017. Lexically constrained decoding for sequence generation using grid beam search. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1535– 1546, Vancouver, Canada. Association for Computational Lingui…
  9. Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2022. Lora: Low-rank adaptation of large language models. In The Tenth International Conference on Learning Representations (ICLR 2022). OpenReview.net.
  10. Alistair E. W. 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. Scientific Data, 3:160035.
  11. Shankar Kumar and William Byrne. 2004. Minimum Bayes-risk decoding for statistical machine translation. In Proceedings of the Human Language Technology Conference of the North American Chapter of the Association for Computational Linguistics: HLT-NAACL 2004, pages 169–176, Boston, Massachusetts, USA…
  12. Woosuk Kwon, Zhuohan Li, Siyuan Zhuang, Ying Sheng, Lianmin Zheng, Cody Hao Yu, Joseph E. Gonzalez, Hao Zhang, and Ion Stoica. 2023. Efficient memory management for large language model serving with pagedattention. Preprint, arXiv:2309.06180.
  13. James Mullenbach, Sarah Wiegreffe, Jon Duke, Jimeng Sun, and Jacob Eisenstein. 2018. Explainable prediction of medical codes from clinical text. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1…
  14. OpenAI. 2025. gpt-oss-120b & gpt-oss-20b model card. Preprint, arXiv:2508.10925.
  15. Matt Post and David Vilar. 2018. Fast lexically constrained decoding with dynamic beam allocation for neural machine translation. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers),…
  16. Gerard Salton and Christopher Buckley. 1988. Term-weighting approaches in automatic text retrieval. Information Processing & Management, 24(5):513– 523. find
  17. Alistair Johnson, Lucas Bulgarelli, Tom Pollard, Steven Horng, Leo Anthony Celi, and Roger Mark. 2023a. MIMIC-IV. PhysioNet. Version 2.2. find
  18. Andrew Sellergren, Chufan Gao, Fereshteh Mahvar, Timo Kohlberger, Fayaz Jamil, Madeleine Traverse, Alberto Tono, Bashir Sadjad, Lin Yang, and 1 others. 2026. Medgemma 1.5 technical report. arXiv preprint arXiv:2604.05081. — [arXiv:2604.05081](https://arxiv.org/abs/2604.05081) find
  19. Alistair Johnson, Tom Pollard, Steven Horng, Leo Anthony Celi, and Roger Mark. 2023b. MIMIC-IV-Note: Deidentified free-text clinical notes. PhysioNet. Version 2.2. find
  20. Andrew Sellergren, Sahar Kazemzadeh, Tiam Jaroensri, Atilla Kiraly, Madeleine Traverse, Timo Kohlberger, and 1 others. 2025. Medgemma technical report. Preprint, arXiv:2507.05201. find
  21. Alistair E. W. Johnson, Lucas Bulgarelli, Lu Shen, and 1 others. 2023c. Mimic-iv, a freely accessible electronic health record dataset. Scientific Data, 10(1):1.
  22. Benjamin Warner, Antoine Chaffin, Benjamin Clavié, Orion Weller, Oskar Hallström, Said Taghadouini, Alexis Gallagher, Raja Biswas, Faisal Ladhak, Tom Aarsen, Griffin Thomas Adams, Jeremy Howard, and Iacopo Poli. 2025. Smarter, better, faster, longer: A modern bidirectional encoder for fast, memory e…
  23. An Yang and 1 others. 2025. Qwen3 technical report. Preprint, arXiv:2505.09388.

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/wang-2026-cptcoder-reliable-llm-system-medical/