This is a LoRA (Low-Rank Adaptation) fine-tuned version of google/medgemma-4b-it trained on the PULSE-ECG/ECGInstruct dataset containing 1.15M ECG instruction-following examples.

Developed by: convaiinnovations
Base Model: google/medgemma-4b-it
Training Infrastructure: AIRAWAT (C-DAC) - 8x NVIDIA A100 40GB GPUs
Training Duration: ~40 hours (1 epoch)
Final Token Accuracy: 88.23%

Training Details

Training Data

Dataset: PULSE-ECG/ECGInstruct
Samples: 1,154,110 training examples
Image Sources: MIMIC-IV-ECG, PTB-XL, CODE-15%, and other ECG datasets
Task: Vision-language instruction following for ECG interpretation

Training Procedure

Hardware:

8x NVIDIA A100 40GB GPUs (AIRAWAT supercomputer)
Distributed training with PyTorch DDP

Hyperparameters:

LoRA rank (r): 32
LoRA alpha: 64
LoRA dropout: 0.05
Learning rate: 2e-4
Batch size: 128 (effective)
Optimizer: AdamW (fused)
LR scheduler: Cosine with warmup
Precision: bfloat16
Gradient checkpointing: Enabled

Training Metrics:

Final training loss: 10.997
Mean token accuracy: 88.23%
Entropy: 1.796
Total tokens processed: 253,325,537

Usage

Installation

pip install transformers peft pillow torch

Loading the Model

python

from transformers import AutoModelForImageTextToText, AutoProcessor
from peft import PeftModel
from PIL import Image

# Load base model
base_model_id = "google/medgemma-4b-it"
model = AutoModelForImageTextToText.from_pretrained(
    base_model_id,
    torch_dtype="auto",
    device_map="auto"
)

# Load LoRA adapter
model = PeftModel.from_pretrained(
    model,
    "convaiinnovations/medgemma-4b-ecginstruct-lora"
)

# Load processor
processor = AutoProcessor.from_pretrained(base_model_id)

Inference Example

python

# Load ECG image
image = Image.open("ecg_image.png").convert("RGB")

# Prepare prompt
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image"},
            {"type": "text", "text": "Analyze this ECG and provide a detailed interpretation."}
        ]
    }
]

# Process input
text = processor.apply_chat_template(messages, add_generation_prompt=True, tokenize=False)
inputs = processor(text=[text], images=[[image]], return_tensors="pt", padding=True)
inputs = {k: v.to(model.device) for k, v in inputs.items()}

# Generate
outputs = model.generate(
    **inputs,
    max_new_tokens=512,
    do_sample=False,
    temperature=None,
    top_p=None
)

# Decode
response = processor.decode(outputs[0], skip_special_tokens=True)
print(response)

Model Capabilities

This model can:

✅ Interpret 12-lead ECG images
✅ Identify cardiac abnormalities (arrhythmias, ischemia, hypertrophy, etc.)
✅ Generate detailed clinical reports
✅ Answer questions about ECG findings
✅ Provide diagnostic suggestions

Limitations

Trained primarily on adult ECG data
Should not replace professional medical diagnosis
Performance may vary on ECG formats not seen during training
Requires high-quality ECG images for optimal results

Ethical Considerations

⚠️ Medical Disclaimer: This model is intended for research and educational purposes only. It should not be used as a substitute for professional medical advice, diagnosis, or treatment. Always consult qualified healthcare professionals for medical decisions.

Citation

If you use this model, please cite:

bibtex

@misc{medgemma-ecginstruct-lora,
  author = {convaiinnovations},
  title = {MedGemma-4B ECGInstruct LoRA},
  year = {2025},
  publisher = {HuggingFace},
  howpublished = {\url{https://huggingface.co/convaiinnovations/medgemma-4b-ecginstruct-lora}}
}

Acknowledgments

Base Model: Google's MedGemma team
Dataset: PULSE-ECG/ECGInstruct
Infrastructure: AIRAWAT AI Innovation Challenge (C-DAC)
Training Framework: HuggingFace Transformers, PEFT, TRL

License

Apache 2.0 (following base model license)