From Documents to Spans: Span-Centric Learning for LLM-based Explainable Medical Coding

Xu Zhang; Wenxin Ma; Chenxu Wu; Rongsheng Wang; Zhiyang He; Xiaodong Tao; Kun Zhang; S. Kevin Zhou

cs.CLMar 16, 2026

From Documents to Spans: Span-Centric Learning for LLM-based Explainable Medical Coding

Authors: Xu Zhang, Wenxin Ma, Chenxu Wu, Rongsheng Wang, Zhiyang He, Xiaodong Tao, Kun Zhang, S. Kevin Zhou

Organizations: School of Biomedical Engineering, Division of Life Sciences and Medicine, USTC · MIRACLE Center, Suzhou Institute for Advance Research, USTC · iFlyHealth · Jiangsu Provincial Key Laboratory of Multimodal Digital Twin Technology ·

Abstract

Medical coding, which assigns standardized ICD codes to clinical documents, is a fundamental yet challenging task in modern healthcare. It requires extensive knowledge of a large code space and the ability to identify and interpret sparse supporting evidence in long, noisy clinical documents. Most existing methods rely on document-level supervision, which provides limited guidance on supporting evidence and insufficient training signals for rare codes. To address these limitations, we propose Span-Centric Learning (SCL), which separates coding knowledge acquisition from its application in full documents. SCL first learns coding knowledge from short spans, then learns to use this knowledge in full documents from a small set of annotated data. Experiments show that, our approach enables an 8B LLM to surpass state-of-the-art methods using only 248 complete documents for training. It also provides explicit supporting evidence for each predicted code, with strong alignment to human annotations. Further analysis shows that the model can learn knowledge of codes absent from documents through spans and apply that knowledge to full documents.

Figures & tables

Figure 1: Motivation. Human coders assign medical codes by identifying supporting evidence in clinical documents. Motivated by this process, our method first learns broad coding knowledge from span data, and then transfers learned knowledge to document-level explainable medical coding.

Figure 2: Overview of SCL framework. Span Data Construction (Sec 3.2 ) constructs, expands and refines a span library from public medical coding datasets. Span-Centric Staged Learning (Sec 3.3 ) first learns broad coding knowledge from the span library, then learns to apply knowledge to document-level medical coding from a small set of high-quality CoT data, with mixed-batch replay of a randomly sampled fraction of the span library. At inference, the model identifies evidence spans in unseen documents and assigns corresponding codes.

Full codes (complete ICD-10-CM code space, ${\sim}$ 70k codes)
Method	Venue	ACI			MDACE			Avg
Method	Venue	F1	P	R	F1	P	R	F1
PLM-ICD	NAACL 22 WS	21.0	12.5	65.2	54.8	52.4	57.5	37.9
GKI-ICD	ACL’25	42.6	31.8	64.4	54.5	53.5	55.5	48.6
MedCodER-GPT5	NAACL’25	62.0	57.0	68.0	40.0	37.0	42.0	51.0
CLH-GPT5	EMNLP’25	57.2	58.5	56.0	46.0	40.0	55.0	51.6

Table 1: Performance comparison with state-of-the-art methods. F1, P, and R denote micro-averaged F1, precision, and recall; Avg is the mean F1 over the two benchmarks. Bold : best in each setting; green : gain over the best baseline.

Method	Training data		ACI		MDACE		Time (hr)	Explainability
Method	Docs	Spans	Micro-F1	Macro-F1	Micro-F1	Macro-F1	Time (hr)	Explainability
SFT	120k	–	47.9	15.7	47.0	24.0	252	✗
SFT E	120k	–	53.6	29.6	55.7	30.2	264	✓
SCL (Ours)	248	200k	67.3 ∗ +19.4	41.0 +25.3	57.2 ∗ +10.2	32.4 +8.4	22	✓

Table 2: Full-document supervision vs. span-centric supervision. All models use Llama3-8B and are trained on a single H20 GPU. SFT and SFT E use about 120k training documents from MIMIC-IV, MDACE, and ACI, with the latter augmented with evidence generated by Llama-3.3-70B for document-level distillation. SCL uses 248 evidence-annotated documents and 200k spans. ∗ indicates

p<10^{-2}

vs. SFT (two-sided paired permutation tests, Holm-adjusted). Green : gain over SFT.

Rule-based (lexical matching)
Metric	Doc	Doc & Span
Exact Match Rate	99.3	98.9
Hit Rate (IoU $>$ 0)	67.1	74.5 +7.4
Hit Rate (IoU $>$ 0.5)	47.0	54.2 +7.1
LLM-based (semantic matching)
Recall	64.6	73.9 +9.2

Table 3: Evaluation of model-predicted evidence on MDACE. We additionally include LLM-based metrics, since there may be multiple semantically equivalent evidence spans for one code in a document. All scores are percentages; bold : better; green : gain over Doc.

Figure 3: Learning new code knowledge from spans on ACI. (a) We divide ACI test-set codes into two groups: codes covered by both the 248 training documents and the span library, and codes covered by the span library but absent from all 248 training documents. (b) Span-level supervision substantially improves coding performance for codes absent from these documents, without requiring additional training documents.

Span library	Naturalness	Plausibility
Initial spans	98.71	77.38
+ Span Synthesis	96.50	66.51
+ Iterative Refinement	100.00	100.00

Table 4: Span quality across iterative refinement. Naturalness and plausibility are pass rates (%) under LLM-based assessment.

One-stage
Strategy	Micro-F1	Time (hr)
Direct mixing	52.8	50
Separate long/short batches	42.9	20
Two-stage
w/o replay	43.8	22
w/ replay	57.2	22

Table 5: Ablation of training strategies. We compare different strategies for mixing long documents and short spans. Our two-stage strategy achieves the best performance while substantially shortening training time.

Documents only
Variant	Micro-F1	Macro-F1
w/o evidence CoT	36.8	13.9
w/ evidence CoT	43.9	17.9
Documents & spans
Initial spans	48.9	22.7
+ Span Synthesis	53.7	27.9

Table 6: Incremental ablation of span-centric supervision. Starting from initial spans, we progressively add span synthesis and iterative refinement. Span synthesis uses no rejected spans or rejection reasons; this feedback is introduced with iterative refinement.

Appendix figures & tables6 assets

Supplementary material from the paper’s appendix.

Appendix

Llama-3.2
Size	Micro-F1	Macro-F1	P	R
1B	40.8	16.0	48.6	35.2
3B	53.2	26.0	60.6	47.4
Qwen2.5
0.5B	31.8	10.6	44.2	24.9
1.5B	50.6	24.1	57.8	45.0

Appendix

Table 7: Scaling law of SCL under restricted code evaluation. Results on MDACE with Llama and Qwen backbones of different sizes. P and R denote micro-averaged precision and recall.

Figure 4: Paradigm comparison on an MDACE test case. Supervised methods predict codes without evidence. Explainable methods provide evidence but with lower coding accuracy. SCL recovers both the evidence spans and the ICD codes that match the expert annotation.

Note	CC: Right arm pain. HPI: Ms. Martinez is a 43-year-old female who presents today for an evaluation of right arm pain after she hitting her arm on a pole while playing tennis . She had mild swelling at the time. There is pain with motion. She denies any numbness or tingling in her hand. She has a past medical history of allergies and takes Flonase. Exam: Examination of the left wrist shows tenderness over the distal radius and pain with flexion of the wrist and thumb. Results: X-rays of the left wrist show no obvious signs of acute fracture or other abnormality. Impression: Right wrist sprain . Plan: At this point, I discussed the diagnosis and treatment options with the patient. I have recommended ice and anti-inflammatory medications. I am prescribing Motrin, 800 mg to take 3 times a day with food. She will follow up with me in 1 week or as needed.
Prediction	Right wrist sprain	S63.501A	Unspecified sprain of right wrist, initial encounter
	hitting her arm on a pole	W22.8XXA	Striking against or struck by other objects, initial encounter
	playing tennis	Y93.73	Activity, racquet and hand sports
Gold		S63.501A	Unspecified sprain of right wrist, initial encounter

Appendix

Table 8: Failure case from the ACI test set. The note documents a right wrist sprain after the patient hit her arm on a pole while playing tennis. SCL codes the injury, the mechanism, and the activity, while the gold annotation contains only the injury code.

Model output	Source text
Macro s ytosis	Macro c ytosis
Astha	Asth m a
emia	an emia
f e ll from a ladder	f a ll from ladder
ventricular fibrillation	vtach/ vfib arrest

Appendix

Table 9: Evidence rewriting errors that prevent source traceability. Model outputs and corresponding source text are reproduced verbatim; red marks the rewritten parts.

Category	Count	Unique codes
Initial assignments	1,765,371	16,156
Supported	1,380,071	14,275
No evidence found	363,639	1,635

Appendix

Table 10: Evidence availability in MIMIC-IV extraction. Counts refer to code assignments. No evidence found denotes assignments for which the extraction pipeline identifies no supporting span in the available discharge summary. Some extraction outputs could not be parsed and are excluded from both outcome categories, so their counts do not sum to the initial total.

Criterion	Matched/total	Agreement (%)
Naturalness	4,532 / 4,645	97.57
Plausibility	4,497 / 4,645	96.81

Appendix

Table 11: Agreement between LLM and human evaluations. All 4,645 spans passed both LLM-based criteria. Matched/total reports the number also judged by human evaluators to satisfy each criterion over the total number evaluated.

Automated Medical Coding

Automated medical coding uses clinical records and coding rules to assign standardized diagnosis and procedure codes for billing, care, and research.

See all 23 papers

Can Post-Training Turn LLMs into Good Medical Coders? An Empirical Study of Generative ICD Coding
Jun 11, 2026Ziqing Wang, Weihao Li, Shijie Chen +2Supervised Fine-Tuning Automated Medical Coding
Northwestern University
CoLa-ICD: A Knowledge-Enhanced Framework for Long-Tail Automated Medical Coding
Aug 31, 2026Yihang Cheng, Veronica Liesaputra, Andrew TrotmanLong-Tailed Classification Multi-Label Classification
University of Otago · Dunedin, New Zealand
Enhancing LLM Medical Coding with Structured External Knowledge
Apr 9, 2026Yidong Gan, David D. Nguyen, Yang Lin +4Automated Medical Coding Medical LLMs
Oracle Health and AI · The University of Sydney · CSIRO Data61

Jun 11, 2026J/K move · Enter open · S save

Can Post-Training Turn LLMs into Good Medical Coders? An Empirical Study of Generative ICD Coding

Ziqing Wang, Weihao Li, Shijie Chen, Yuan Luo+1

Northwestern University

Supervised Fine-Tuning Automated Medical Coding RL for Generative Models LLM Post-Training Medical LLMs

Open paper AskPDF arXiv

Automated International Classification of Diseases (ICD) coding is a core medical-coding task for billing, epidemiology, and clinical decision support. Generative large language models (LLMs) are often reported as weak medical coders, but this finding mainly comes from inference-time settings such as prompting, retrieval, reranking, or tool use, leaving the role of task-specific post-training underexplored. We present a controlled empirical study of post-training for generative ICD coding, comparing discriminative baselines with LLM coders across prompting, supervised fine-tuning, and reinforcement learning under a common protocol and metric set. To our knowledge, this is the first study to evaluate RL-based post-training for generative LLM coders in ICD coding. We further introduce PHI, a diagnostic curriculum that extends GRPO to refine missed-code cases. Our results show that prompting-only evaluation substantially underestimates the potential of LLMs for ICD coding. SFT provides the main capability jump, GRPO further improves code-set prediction beyond SFT, and PHI provides targeted gains on macro-level performance. These findings suggest that the main bottleneck is not the generative formulation alone, but how the model is adapted and optimized for full-taxonomy recall. We release our code, data splits, and checkpoints at https://github.com/AlexandreWANG915/LLM4ICD.

Abstract

Automated Medical Coding

Explore similar work

Can Post-Training Turn LLMs into Good Medical Coders? An Empirical Study of Generative ICD Coding

CoLa-ICD: A Knowledge-Enhanced Framework for Long-Tail Automated Medical Coding

Enhancing LLM Medical Coding with Structured External Knowledge