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

2026-06-11

AI Summary

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.

Abstract

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.

Problem

• ICD coding
• large language models
• post-training
• medical coding
• generative models

Innovation

• post-training
• generative ICD coding
• reinforcement learning
• PHI curriculum
• supervised fine-tuning