DeepICD-R1-7B: Clinical Reasoning for ICD-10-CM Diagnosis
DeepICD-R1-7B is a specialized 7.6 billion parameter language model developed by DATEXIS, built upon the Qwen2.5-7B-Instruct architecture. Its core function is to predict a single ICD-10-CM diagnosis code from clinical admission notes, a critical task in healthcare. What sets this model apart is its unique training methodology, the DeepICD-R1 framework, which frames diagnosis prediction as a reasoning problem.
Key Capabilities & Differentiators
- Interpretable Reasoning: Unlike many LLMs, DeepICD-R1-7B generates a structured reasoning trace (within
<think> tags) alongside the predicted ICD-10-CM code (within <diagnosis> tags), offering transparency into its decision-making process. - Reinforcement Learning Optimization: Training involves a combination of supervised reasoning traces and reinforcement learning using Group Relative Policy Optimization (GRPO), enhanced by hierarchical reward signals. These rewards are aligned with the ICD code structure, providing partial credit for correct high-level categories.
- Clinical Domain Specialization: Fine-tuned on de-identified clinical admission notes paired with ICD-10-CM codes from datasets like MIMIC-IV, making it highly proficient in clinical NLP and healthcare reasoning.
- Hierarchical Evaluation: Performance is measured using macro-averaged F1 scores at various ICD hierarchy levels (Chapter, Category, Full code), allowing for nuanced assessment of diagnostic accuracy.
Intended Use Cases
This model is primarily intended for research purposes in areas such as:
- Clinical reasoning research
- ICD-10-CM coding prediction research
- Reinforcement learning for language models
- Generation of reasoning traces from clinical text
Important Note: This model is strictly for research and must not be used for medical diagnosis, clinical decision support, patient triage, or automated medical coding without expert supervision due to inherent limitations and ethical considerations.