Sign up

convaiinnovations/medgemma-4b-ecginstruct-merged

🤗 Hugging Face sourceimage-text-to-textapache-2.04.3B params8.6 GBsafetensors✓ Checksums witnessedupdated today
No torrent yet →

MedGemma-4B ECGInstruct

Fine-tuned version of Google's MedGemma-4B-it model on the ECGInstruct dataset for automated ECG interpretation.

Model Description

This is a fully merged fine-tuned version of google/medgemma-4b-it trained on the PULSE-ECG/ECGInstruct dataset containing 1.15M ECG instruction-following examples. The LoRA adapter has been merged into the base model for easier deployment.

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%
Model Size: ~8.5 GB

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

Training Configuration:

  • Fine-tuning method: LoRA (r=32, alpha=64, dropout=0.05)
  • Learning rate: 2e-4 with cosine decay
  • Batch size: 128 (effective)
  • Optimizer: AdamW (fused)
  • Precision: bfloat16
  • Gradient checkpointing: Enabled

Training Metrics:

  • Final training loss: 0.9
  • Mean token accuracy: 90%
  • Entropy: 0.98
  • Total tokens processed: 103301284

Usage

Installation

pip install transformers pillow torch

Loading the Model

from transformers import AutoModelForImageTextToText, AutoProcessor
from PIL import Image

# Load model and processor
model_id = "convaiinnovations/medgemma-4b-ecginstruct-merged"
model = AutoModelForImageTextToText.from_pretrained(
    model_id,
    torch_dtype="auto",
    device_map="auto"
)
processor = AutoProcessor.from_pretrained(model_id)

Inference Example

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

# Prepare prompt using chat template
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 interpretation
outputs = model.generate(
    **inputs,
    max_new_tokens=512,
    do_sample=False
)

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

Example Prompts

# Detailed interpretation
"Analyze this ECG and provide a detailed interpretation."

# Specific abnormality detection
"What abnormalities are present in this ECG?"

# Diagnosis suggestion
"Based on this ECG, what is the most likely diagnosis?"

# Question answering
"Does this ECG show signs of atrial fibrillation?"

# Rate and rhythm
"What is the heart rate and rhythm in this ECG?"

Model Capabilities

This model can:

  • ✅ Interpret 12-lead ECG images
  • ✅ Identify cardiac abnormalities (arrhythmias, ischemia, hypertrophy, conduction blocks, etc.)
  • ✅ Generate detailed clinical reports
  • ✅ Answer specific questions about ECG findings
  • ✅ Provide diagnostic suggestions
  • ✅ Assess heart rate, rhythm, and axis
  • ✅ Detect ST-segment changes and T-wave abnormalities

Limitations

  • Trained primarily on adult ECG data
  • Performance may vary on pediatric ECGs
  • Should not replace professional medical diagnosis
  • Requires high-quality ECG images for optimal results
  • May struggle with very rare or unusual ECG patterns
  • Limited to English language outputs

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.

Important Notes:

  • This is an AI model and can make mistakes
  • ECG interpretation requires clinical context
  • Model outputs should be verified by trained clinicians
  • Not approved for clinical use or diagnostic purposes
  • Use responsibly and within appropriate medical oversight

Intended Use

Appropriate Uses:

  • Research in medical AI and computer vision
  • Educational demonstrations of ECG interpretation
  • Development of clinical decision support prototypes
  • Benchmarking ECG analysis algorithms

Inappropriate Uses:

  • Direct patient diagnosis without physician review
  • Replacement of trained medical professionals
  • Use in emergency or critical care settings without oversight
  • Commercial deployment without proper validation

Citation

If you use this model in your research, please cite:

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

Acknowledgments

  • Base Model: Google's MedGemma team for the foundation model
  • Dataset: PULSE-ECG team for the ECGInstruct dataset
  • Infrastructure: AIRAWAT AI Innovation Challenge by C-DAC (Centre for Development of Advanced Computing)
  • Frameworks: HuggingFace Transformers, PEFT, TRL, PyTorch

Related Resources

License

Apache 2.0 (following base model license)

Contact

For questions or issues, please open an issue on the model repository or contact the maintainers.