samwell/ncd-gemma4-e4b-lora

NCD Risk Assessment Model (Gemma 4 E4B Fine-tuned)

A fine-tuned Gemma 4 E4B model for predicting Non-Communicable Disease (NCD) risk - specifically Type 2 Diabetes and Hypertension - from patient clinical data.

Model Description

This model was fine-tuned on 49,214 synthetic patient records to provide clinical decision support for NCD screening in resource-limited settings, particularly designed for deployment in Ghana and similar healthcare contexts.

Attribute	Value
Base Model	google/gemma-4-E4B-it
Fine-tuning Method	QLoRA (4-bit) with Unsloth
LoRA Rank	32
Training Data	39,371 examples
Final Loss	0.1842
Training Time	100 minutes (H200 GPU)

Intended Use

Primary Use Case: Clinical Decision Support (CDS) for NCD risk screening

Target Users:

• Healthcare workers in primary care settings
• Community health workers conducting NCD screenings
• EHR systems (e.g., OpenMRS/HopeOS) for automated risk assessment

Input: Patient demographics, vitals, and lab values Output: Structured risk assessment with clinical reasoning

Model Files

File	Format	Size	Use Case
adapter_model.safetensors	LoRA	~340MB	Fine-tuning, merging
ncd-gemma4-q4_k_m.gguf	GGUF	~2.5GB	Local inference (llama.cpp, Ollama)

How to Use

With Transformers + PEFT (LoRA)

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

# Load base model
base_model = AutoModelForCausalLM.from_pretrained("google/gemma-4-E4B-it")
tokenizer = AutoTokenizer.from_pretrained("google/gemma-4-E4B-it")

# Load LoRA adapter
model = PeftModel.from_pretrained(base_model, "samwell/ncd-gemma4-e4b-lora")

# Inference
prompt = """Based on the following patient record, assess the risk of Type 2 diabetes and hypertension.

Patient: 55yo Male
Vitals: BP 152/94 mmHg, BMI 31.2
Labs: Glucose 126 mg/dL, HbA1c 6.4%
Active conditions: Prediabetes"""

inputs = tokenizer(prompt, return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=512)
print(tokenizer.decode(outputs[0]))

With llama.cpp (GGUF)

# Download GGUF
huggingface-cli download samwell/ncd-gemma4-e4b-lora ncd-gemma4-q4_k_m.gguf

# Run inference
./llama-cli -m ncd-gemma4-q4_k_m.gguf -p "Patient: 45yo Female, BP 140/90..."

With Ollama

# Create model
echo 'FROM ./ncd-gemma4-q4_k_m.gguf' > Modelfile
ollama create ncd-cds -f Modelfile

# Run
ollama run ncd-cds "Patient: 45yo Female, BP 140/90..."

Example Output

Input:

Patient: 55yo Male
Vitals: BP 152/94 mmHg, BMI 31.2
Labs: Glucose 126 mg/dL, HbA1c 6.4%
Active conditions: Prediabetes

Output:

## Risk Assessment

**Type 2 Diabetes: MODERATE**
- HbA1c 6.4% (prediabetes range 5.7-6.4%)
- BMI 31.2 (obese)
- Age 55 years (≥45 increases risk)

**Hypertension: HIGH**
- BP 152/94 mmHg (Stage 2 Hypertension)
- Systolic ≥140 or Diastolic ≥90 confirms diagnosis

## Recommendations

1. **Lifestyle modifications**: Weight loss, DASH diet, sodium reduction
2. **Blood pressure monitoring**: Home BP monitoring recommended
3. **Lab follow-up**: Repeat HbA1c in 3 months
4. **Consider referral**: If BP remains elevated, initiate antihypertensive therapy

Training Details

Dataset

• Source: Synthea synthetic patient generator
• Size: 49,214 examples (39,371 train / 4,921 val / 4,922 test)
• Format: Instruction-tuning (instruction/input/output)
• Dataset: samwell/synthea-ncd-instructions

Training Configuration

# Model
MODEL_NAME = "google/gemma-4-E4B-it"
MAX_SEQ_LENGTH = 2048
LOAD_IN_4BIT = True  # QLoRA

# LoRA
LORA_R = 32
LORA_ALPHA = 32
TARGET_MODULES = ["q_proj", "k_proj", "v_proj", "o_proj",
                  "gate_proj", "up_proj", "down_proj"]

# Training
BATCH_SIZE = 8
GRADIENT_ACCUMULATION = 2  # Effective batch = 16
LEARNING_RATE = 2e-4
NUM_EPOCHS = 3

Training Curve

• Initial loss: 1.71
• Final loss: 0.1842
• Training time: 100 minutes on NVIDIA H200 (80GB)

Limitations

1. Synthetic data only: Trained on Synthea-generated data, not real patient records
2. Limited NCDs: Currently only assesses diabetes and hypertension
3. Not a diagnostic tool: Intended for screening support, not clinical diagnosis
4. Requires clinical validation: Must be validated by healthcare professionals before clinical use

Ethical Considerations

• Not FDA/CE approved for clinical diagnosis
• Should be used as decision support, not replacement for clinical judgment
• Predictions should be reviewed by qualified healthcare providers
• Model may reflect biases in training data

Citation

@misc{ncd-gemma4-2026,
  author = {HopeOS Team},
  title = {NCD Risk Assessment Model: Fine-tuned Gemma 4 for Diabetes and Hypertension Prediction},
  year = {2026},
  publisher = {HuggingFace},
  url = {https://huggingface.co/samwell/ncd-gemma4-e4b-lora}
}

Related Resources

• Dataset: samwell/synthea-ncd-instructions
• Base Model: google/gemma-4-E4B-it
• Training Library: Unsloth

License

This model is released under the Gemma license.