---
library_name: transformers
tags:
  - medical
  - icd10
  - text-generation
  - lora
  - healthcare
---

# ICD-10 Code Predictor

A fine-tuned language model that predicts ICD-10 diagnosis codes from clinical text descriptions.

## Model Details

### Model Description

This model takes a plain English description of a patient's symptoms or condition and outputs the corresponding ICD-10 diagnosis code. It is built on Meta's Llama 3.2 3B base model, fine-tuned using LoRA (Low-Rank Adaptation) with the Unsloth library for efficient training.

- **Developed by:** StudioIlios
- **Model type:** Causal Language Model (LoRA fine-tuned)
- **Language(s):** English (clinical/medical text)
- **Base Model:** meta-llama/Llama-3.2-3B
- **Fine-tuning method:** LoRA via Unsloth
- **License:** [More Information Needed]

## Uses

### Direct Use

Input a clinical description of a patient's condition and the model will return the predicted ICD-10 code.

**Example prompt:**
```
Patient has diabetes mellitus with high blood sugar. What is the ICD10 code?
```

**Example output:**
```
The ICD10 code for Diabetes mellitus is E11.9
```

### Downstream Use

- Medical billing automation
- Insurance claim processing
- EHR (Electronic Health Record) systems
- Healthcare apps requiring automatic diagnosis code suggestion

### Out-of-Scope Use

- This model should **not** be used as a substitute for professional medical diagnosis
- Not suitable for rare or highly complex conditions without human verification
- Not intended for real-time critical care decisions

## Bias, Risks, and Limitations

- Model predictions should always be verified by a qualified medical coder or physician
- May not accurately predict codes for uncommon or highly specific conditions
- Performance depends on how clearly the condition is described in the input

### Recommendations

Always have a medical professional review the predicted ICD-10 codes before using them for billing or insurance purposes.

## How to Get Started with the Model

```python
from transformers import AutoTokenizer, AutoModelForCausalLM
from peft import PeftModel

base_model = AutoModelForCausalLM.from_pretrained("meta-llama/Llama-3.2-3B")
tokenizer = AutoTokenizer.from_pretrained("StudioIlios/icd10-model")
model = PeftModel.from_pretrained(base_model, "StudioIlios/icd10-model")

prompt = """
Patient has diabetes mellitus with high blood sugar.
What is the ICD10 code?
"""

inputs = tokenizer(prompt, return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=100, temperature=0.7, do_sample=True)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
```

## Training Details

### Training Data

Fine-tuned on medical clinical text paired with ICD-10 diagnosis codes.

### Training Procedure

#### Training Hyperparameters

- **Training regime:** LoRA fine-tuning with bf16 mixed precision
- **Library:** Unsloth
- **Base model:** Llama 3.2 3B

## Evaluation

### Results

The model correctly predicts common ICD-10 codes from plain English clinical descriptions.

**Sample tested:**

| Input | Predicted Code |
|---|---|
| Diabetes mellitus with high blood sugar | E11.9 |

## Technical Specifications

### Model Architecture

- Base: Llama 3.2 3B (causal language model)
- Adapter: LoRA (Low-Rank Adaptation)
- Files: `adapter_config.json`, `adapter_model.safetensors`

### Hardware Used for Training

- GPU: NVIDIA Tesla T4 (Google Colab)