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

text

<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

python

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:

bibtex

@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}
}