convaiinnovations/medgemma-4b-ecginstruct

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: ConvAI Innovations
Base Model: google/medgemma-4b-it
Training Infrastructure: AIRAWAT (C-DAC) - 8x NVIDIA A100 40GB GPUs
Training Duration: 72 hours (3 days)
Final Token Accuracy: 86.83%
Final Training Loss: 0.6188
GPU-Hours: 576
Model Size: ~8.5 GB

Training Details

Training Data

• Dataset: PULSE-ECG/ECGInstruct (1.15M samples)
• Samples: 1,156,110 ECG image-text pairs
• Image Sources: MIMIC-IV-ECG (~800K), PTB-XL (22K), CODE-15% (346K), ChapmanShaoxing
• Task: Vision-language instruction following for ECG interpretation
• Demographics: Age range 0-95 years, 52% male / 48% female
• Disease Classes: 5 superclasses (NORM, MI, STTC, CD, HYP), 24 subclasses

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)
• Target modules: all-linear (including vision encoder)
• Learning rate: 1.2e-5 with cosine decay
• Batch size: 192 effective (4 per GPU × 8 GPUs × 6 gradient accumulation)
• Optimizer: AdamW (fused)
• Precision: bfloat16
• Gradient checkpointing: Enabled
• Max sequence length: 2048 tokens
• Max new tokens: 512

Training Metrics:

• Final training loss: 0.6188
• Mean token accuracy: 86.83%
• Training throughput: ~9.67 samples/sec
• Total tokens processed: 103M+

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"
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

Performance

Training Metrics:

Metric	Value
Token Accuracy	86.83%
Final Loss	0.6188
Training Time	72 hours
GPU-Hours	576

Inference Metrics (A100 GPU):

Metric	Value
TTFT (Time to First Token)	~150ms
ISL (Input Sequence Length)	2048 tokens
OSL (Output Sequence Length)	512 tokens
End-to-End Latency	2-3 seconds
Throughput	~45 tokens/sec

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 for RESEARCH AND EDUCATIONAL PURPOSES ONLY.

• ❌ NOT validated for clinical use
• ❌ NOT FDA/CE approved
• ❌ NOT a substitute for professional medical diagnosis
• ❌ Should NOT be used for patient care decisions

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
• Has not been tested on external clinical datasets

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

• LoRA Adapter Version: convaiinnovations/medgemma-4b-ecginstruct-lora
• Base Model: google/medgemma-4b-it
• Training Dataset: PULSE-ECG/ECGInstruct

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.