medical coding knowledge

Existing automatic ICD coding systems lack interpretability for clinical long-text, multi-label scenarios, and current evaluation methods fail to validate evidence–code consistency. Method: Leveraging the MDACE dataset, we propose a novel evidence matching metric that quantifies semantic overlap between model-extracted textual evidence and ground-truth code descriptions, integrating text matching and evidence alignment techniques to systematically assess the effectiveness and bias of mainstream interpretability methods in evidence extraction. Contribution/Results: Our evaluation identifies successful and failure cases, revealing that while current methods capture partial ground-truth evidence, overall evidence–code consistency remains limited. The proposed metric significantly enhances objectivity and clinical plausibility in interpretability assessment. This work establishes a reproducible, clinically grounded evaluation framework—supporting the development, diagnostic traceability, and trustworthy deployment of interpretable medical coding systems.

This study addresses three critical challenges in ICD-10-CM coding automation: suboptimal large language model (LLM) selection, insufficient clinical context modeling, and training data redundancy. To tackle these, we propose a modular medical coding framework. First, we introduce an LLM-as-Judge evaluation paradigm integrated with Plackett–Luce ranking for principled, task-specific LLM selection. Second, we design a semantic-similarity-driven redundancy-aware sampling strategy to mitigate data duplication. Third, we develop a section-aware contextual modeling approach coupled with context-sensitive prompting to enhance clinical paragraph understanding and coding consistency. Evaluated on two real-world institutional datasets, our selected open-source LLM—after fine-tuning—significantly outperforms mainstream baselines in both in-domain and out-of-domain settings (F1 gains of 3.2–5.8 percentage points). Further incorporating clinical paragraph enhancement yields additional performance improvements.

This study addresses the inefficiency and systematic undercoding of secondary diagnoses in manual medical coding by developing a multimodal language model trained on 5.8 million electronic health records from 1.8 million patients in eastern Denmark—a population-scale cohort encompassing nearly all medical specialties. The model integrates clinical notes, medication records, and laboratory data to predict ICD-10 codes. Evaluated on a hold-out set of 270,000 patients, it achieves a micro-averaged F1 score of 71.8% and a top-10 recall of 95.5%. It also identified thousands of cases with missed secondary diagnoses, 76–86% of which were confirmed as valid upon manual review. The approach can automate approximately 50% of coding tasks, offering a scalable tool for epidemiological and multimorbidity research.

Accurate ICD-10 coding of hospital discharge summaries is clinically critical yet error-prone due to its hierarchical, fine-grained nature. Method: We evaluated 11 large language models (LLMs), including Gemini 2.5 Pro, on high-frequency ICD-10 codes using a standardized coder-style prompting template; clinical entities were pre-extracted via cTAKES to enhance input fidelity. Hierarchical classification was performed to reflect real-world coding constraints. Contribution/Results: We introduced and systematically compared LLMs with structured reasoning capabilities against non-reasoning counterparts. All models achieved ≤57% macro-F1, with performance inversely correlated with code specificity. Reasoning-capable models consistently outperformed non-reasoning ones, with Gemini 2.5 Pro achieving the highest F1. These findings indicate that current LLMs can serve as efficiency-enhancing decision-support tools in clinical coding but remain insufficient to replace human coders due to persistent accuracy limitations in complex hierarchical classification.

Automated ICD coding in clinical practice lacks verifiable textual evidence, undermining the reliability of human-AI collaboration. Method: This paper reformulates ICD coding as an interpretable entity linking task—enabling explicit alignment between diagnostic codes and supporting text spans in clinical notes for the first time. We propose a few-shot, explainable coding framework that integrates parameter-efficient fine-tuning with constrained decoding, leveraging large language models (LLMs) for evidence-aware, end-to-end reasoning. Contribution/Results: Our approach maintains high coding accuracy while significantly improving the precision and verifiability of code attribution. In few-shot settings, it outperforms conventional classification baselines. By generating traceable, auditable textual evidence for each assigned code, the method enhances transparency and supports regulatory compliance, advancing the deployment of trustworthy medical AI systems.

This work addresses the scarcity of real-world clinical text due to privacy constraints, which hinders the development of clinical AI systems. The authors propose a modular synthetic data generation pipeline that integrates structured patient modeling, semi-structured clinical course simulation, and large language model (LLM)-driven generation of unstructured clinical notes. This approach ensures longitudinal consistency and clinical plausibility while enabling diverse writing styles. Innovatively, the framework incorporates an LLM-based validation and refinement mechanism to enhance the fidelity and realism of the synthetic data. The study releases a benchmark dataset comprising 70 virtual patients, each with 20–50 clinical notes spanning their entire hospitalization, offering a high-quality, scalable resource for developing and evaluating clinical AI tools.

Existing studies on automatic ICD coding with large language models (LLMs) rely solely on prompting-based evaluation, substantially underestimating their true potential. This work establishes a unified evaluation protocol to systematically compare post-training strategies—including prompting, supervised fine-tuning (SFT), and reinforcement learning via GRPO—and presents the first application of GRPO to generative LLMs for ICD coding. Furthermore, we introduce the PHI Diagnostic Curriculum strategy to specifically address under-coding issues and improve macro-recall. Experimental results demonstrate that SFT yields the primary performance gains, GRPO further refines the predicted code sets, and the PHI strategy effectively enhances macro-level recall, collectively advancing the state of the art in LLM-based clinical coding.

This work addresses the limitation of existing electronic health record (EHR) foundation models that treat ICD diagnosis codes as flat tokens, thereby neglecting their intrinsic clinical hierarchical structure and underutilizing semantic information. To overcome this, the study introduces the ICD-10-CM hierarchy as an inductive bias into EHR modeling for the first time, proposing a hierarchy-enhanced Transformer and a hierarchy-aware graph neural network that jointly leverage diagnosis co-occurrence patterns and multi-granular ICD codes. The model is pretrained on MIMIC-IV and evaluated via frozen probing on eICU for cross-dataset generalization. Results demonstrate consistent and significant improvements over flat-code baselines in both in-domain and cross-domain settings, enhancing downstream prediction performance and yielding more semantically coherent embedding spaces, thus validating the broad utility of hierarchical modeling across diverse tasks and architectures.

This work addresses the longstanding reliance on manual medical coding, which is inefficient and error-prone, and overcomes key limitations of existing automated approaches—namely poor generalizability to new coding systems and lack of interpretability. The authors propose an agent-based reasoning framework that emulates human expert decision-making by dynamically retrieving official coding guidelines and integrating them with clinical text understanding, enabling adaptation to any coding system without retraining. This approach achieves zero-shot cross-system transfer—the first of its kind—and provides traceable justifications linking predicted codes to supporting evidence in the source documents. Evaluated across five real-world and publicly available datasets spanning multiple countries and clinical specialties, the method demonstrates state-of-the-art performance and strong practical deployability.

This study addresses a critical limitation in current clinical AI systems, which typically treat electronic health records (EHRs) as static inputs and lack capabilities for longitudinal modeling and interpretable reasoning over patient histories. The work proposes the first EHR-integrated AI framework specifically designed for longitudinal clinical reasoning, combining clinical natural language processing, cross-visit trajectory modeling, and qualitative physician interviews to systematically evaluate existing systems’ performance in temporal dynamics, information integration, and missing-data inference. Findings reveal that prevailing approaches predominantly rely on visit-level or aggregated representations, omit explicit longitudinal mechanisms, and prioritize predictive accuracy over clinical interpretability. By shifting the focus from prediction-centric paradigms toward time-sensitive, explainable decision support, this framework advances the development of clinically meaningful AI tools.