DeepICD-R1-zero-32B

Model Summary

DeepICD-R1-zero-32B is a clinical reasoning model designed for ICD-10-CM diagnosis outcome prediction from admission notes.
It follows the DeepICD-R1 framework, which treats diagnosis prediction as a reasoning task optimized with reinforcement learning and structured reward signals.

This checkpoint corresponds to a “R1-Zero” style model, meaning it was trained primarily through reinforcement learning without a supervised fine-tuning (SFT) initialization, allowing reasoning behaviors to emerge directly from reward optimization.

The approach is inspired by reasoning-focused training pipelines where reinforcement learning alone can induce structured reasoning behaviors and self-verification in large language models.

---

Model Details

• Model name: DeepICD-R1-zero-32B
• Organization: DATEXIS
• Model size: ~32B parameters
• Task: Single ICD-10-CM diagnosis prediction from clinical text
• Training paradigm: Reinforcement learning (GRPO-style)
• Framework: VERL reinforcement learning trainer
• Domain: Clinical NLP / medical reasoning

Related Research

This model follows the DeepICD-R1 framework introduced in:

DeepICD-R1: Medical Reasoning through Hierarchical Rewards and Unsupervised Distillation

The paper proposes a system for diagnosis prediction that combines:

• structured reasoning traces
• hierarchical reward signals aligned with ICD code structure
• reinforcement learning for reasoning optimization

---

Intended Use

This model is intended for research purposes, including:

• clinical reasoning experiments
• ICD-10-CM code prediction research
• reinforcement learning for language models
• reasoning trace generation
• structured prediction from clinical notes

Out-of-Scope Use

This model must not be used for:

• medical diagnosis
• clinical decision making
• patient triage
• automated medical coding without expert supervision
• billing or compliance workflows

---

Training Methodology

R1-Zero Training Paradigm

The model follows a Zero-stage reasoning training approach, where reinforcement learning is applied directly to a base language model without prior supervised instruction tuning.

This method encourages models to discover reasoning strategies autonomously during training, allowing behaviors such as:

• chain-of-thought reasoning
• self-verification
• iterative reasoning refinement

to emerge naturally from the reward signal.

However, purely RL-trained models may also exhibit issues such as:

• repetitive reasoning patterns
• readability problems
• mixed language outputs

---

Training Data

The training task uses clinical admission notes paired with ICD-10-CM diagnoses, derived from de-identified electronic health record datasets such as MIMIC-IV.

Task formulation:

• Input: admission note describing a patient case
• Output: reasoning trace and predicted ICD-10-CM code

The model learns to infer diagnostic outcomes based on the textual description of the patient presentation.

---

Output Format

The model is trained to produce structured outputs separating reasoning from the final diagnosis.

Example

<think>
The patient presents with ...
Symptoms and history suggest ...
...
</think>

<diagnosis>
M5116
</diagnosis>

The reasoning trace allows the model to explain how the diagnosis is derived from the clinical note.

---

Evaluation

Evaluation follows the methodology described in the DeepICD-R1 paper.

Performance is typically measured using macro-averaged F1 scores at multiple levels of the ICD hierarchy.

Level	Description
Chapter	Broad ICD category
Category	First three digits
Full code	Complete ICD-10 code

Hierarchical evaluation allows partial credit when the model predicts the correct high-level diagnostic category even if the full code is incorrect.

---

Limitations

Models following the DeepICD-R1 framework share several limitations.

Dataset limitations

• Training data consists primarily of English clinical notes
• Distribution reflects hospital-specific patient populations
• ICD labels are highly imbalanced, affecting rare diagnoses

Model limitations

• Reasoning traces may appear convincing while being incorrect
• Predictions may fail for rare or long-tail diagnoses
• Models may demonstrate premature diagnostic closure
• Reinforcement learning signals are only proxies for expert feedback

---

Ethical Considerations

This model is trained on de-identified clinical data and intended strictly for research.

Potential risks include:

• propagation of dataset biases
• overconfidence in generated reasoning
• misuse in clinical decision making

Appropriate safeguards include:

• expert oversight
• dataset bias evaluation
• fairness audits
• controlled deployment environments

---

Hardware and Training Setup

Typical training configuration for models in this family includes:

• GPUs: multi-GPU training (4–8 GPUs)
• Precision: bfloat16
• Rollout engine: vLLM
• Training framework: VERL PPO/GRPO trainer
• Sampling: multiple rollouts per prompt

---

Usage

Transformers Example

from transformers import AutoTokenizer, AutoModelForCausalLM

model_id = "DATEXIS/DeepICD-R1-zero-32B"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    device_map="auto",
    torch_dtype="auto"
)

prompt = """
You are a clinical reasoning model.

Given the following admission note,
produce reasoning in <think> tags
and a final ICD-10 diagnosis in <diagnosis> tags.

[ADMISSION NOTE]
"""

inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

outputs = model.generate(
    **inputs,
    max_new_tokens=512,
)

print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Recommended Inference Practices

• Use prompts consistent with the training format.
• Validate predicted ICD-10 codes against official code formats.
• Always review predictions with medical experts.
• Avoid exposing reasoning traces in safety-critical settings without verification.

---

Citation

If you use this model, please cite:

@inproceedings{roehr2026deepicdr1,
  title={DeepICD-R1: Medical Reasoning through Hierarchical Rewards and Unsupervised Distillation},
  author={R{\"o}hr, Tom and Steffek, Thomas and Teucher, Roman and Bressem, Keno and others},
  booktitle={Proceedings of LREC-COLING},
  year={2026}
}