LLMs for ICD Coding - What 3 Recent Papers Tell Us

Medical coding—assigning ICD-10 codes to clinical notes—is a $15B+ industry bottleneck. Human coders spend ~30 minutes per case, errors cause billing disputes and patient safety issues, and the 70,000-code label space makes automation notoriously hard. Three recent papers attempt to crack this with LLMs. Here’s what they reveal about where the field is heading.

---

1. MedCodER: A Generative AI Assistant for Medical Coding (arXiv 2409.15368)

Three-stage LLM pipeline (GPT-4): extraction → retrieval augmentation → re-ranking. Evaluated on extended ACI-BENCH (184 synthetic clinical notes, 360 ICD-10 codes). Claims SOTA over PLM-ICD and LLM Tree-Search.

---

Structural Contribution: Interpretability-First Design

Decomposing the task into explicit diagnosis extraction → evidence grounding → code ranking creates an auditable chain that human coders can verify. This contrasts sharply with end-to-end classifiers (PLM-ICD) that offer no justification. The supporting evidence annotation effort—despite its small scale—creates a template for future datasets.

---

2. Code Like Humans: A Multi-Agent Solution for Medical Coding (arXiv 2509.05378)

Four-agent pipeline mimicking NHS coding workflow: evidence extractor → index navigator → tabular validator → code reconciler. Evaluated on MDACE (302 real MIMIC-III charts with human-verified evidence spans).

---

Brilliant Insight: The Alphabetic Index Is the Missing Link

Human coders don’t start from memory—they start from the ICD alphabetic index, a resource mapping clinical terms to codes with synonyms, eponyms, and contextual cues. Prior ML approaches ignored this entirely, treating coding as pure memorization. CLH’s index navigator retrieves candidate codes from this authoritative source, enabling true open-set inference over the full 70K codes without retraining.

---

3. Toward Reliable Clinical Coding with Language Models: Verification and Lightweight Adaptation (arXiv 2510.07629)

Verification pipeline: generate codes → expand via ICD hierarchy (siblings, cousins, 1-hop/2-hop neighbors) → LLM-based selection. New double expert-annotated outpatient dataset (207 ACI-BENCH notes with ICD-10-CM). Tests Claude Haiku/Sonnet, GPT-4o, Llama, PLM-ICD. Claims +16 F1 for fine-tuned Haiku-3 with verification.

---

Practical Contribution: Verification as Model-Agnostic Post-Processing

The expansion → selection pipeline is elegantly simple: take predicted codes, retrieve structurally related candidates (siblings, cousins, neighbors from both tabular and index lists), then ask an LLM to pick the best match given the note. This works across model families (Claude, GPT, Llama, PLM-ICD) and requires no retraining. Notably, presenting descriptions only (not codes) yields highest verification accuracy—CoT reasoning hurts performance here.

---

High-Level Summary (My Own Perspective)

Paper	Architecture	Key Innovation	Data
MedCodER	Extract → Retrieve → Re-rank	Evidence grounding	ACI-BENCH
CLH	4-agent NHS workflow	Alphabetic index retrieval	MDACE
Verification	Generate → Expand → Verify	Hierarchy-aware refinement	ACI-BENCH

What They Agree On

Pipeline > End-to-End: All reject monolithic classifiers; decomposition enables interpretability and modularity.

The Uncomfortable Truth

Pipeline Fragility and Error Cascades: End-to-end differentiability was the revolution of the 2010s because it allowed the model to recover from upstream noise. Reverting to discrete, non-differentiable agent hand-offs reintroduces the brittleness we spent a decade eliminating.

Where to Go Next

1. Hybrid Retrieval
   • Sparse: TF-IDF/BM25
   • Dense: Contrastive/LLM-trained embeddings
   • Structural: Sibling/cousin/neighbor expansion
2. ICD-Aware Tokenizer
   Don’t use BPE/SentencePiece. Design tokenizer where each ICD code is a single token.
3. Dataset for Agentic LLM

---

References

1. Baksi, K. D. et al. MedCodER: A generative AI assistant for medical coding. in Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 3: Industry Track) 449–459 (2025).
2. Motzfeldt, A. et al. Code like humans: A multi-agent solution for medical coding. in Findings of the Association for Computational Linguistics: EMNLP 2025 22612–22627 (Association for Computational Linguistics, 2025).
3. Yuan, M., Shing, H.-C., Strong, M. & Shivade, C. Toward Reliable Clinical Coding with Language Models: Verification and Lightweight Adaptation. in Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing: Industry Track 173–184 (2025).

---