Overview

MedGemma-4B ECGInstruct is a 4.3 billion parameter vision-language model developed by ConvAI Innovations, fine-tuned from Google's MedGemma-4B-it. It is specifically designed for automated interpretation of 12-lead ECG images, leveraging the extensive PULSE-ECG/ECGInstruct dataset which contains 1.15 million ECG image-text pairs. The model was trained for 72 hours on 8x NVIDIA A100 GPUs, achieving a final token accuracy of 86.83%.

Key Capabilities

• Interprets 12-lead ECG images: Analyzes visual ECG data to extract clinical insights.
• Identifies cardiac abnormalities: Detects arrhythmias, ischemia, hypertrophy, and conduction blocks.
• Generates detailed clinical reports: Provides comprehensive interpretations based on ECG findings.
• Answers specific questions: Responds to queries about ECG findings, diagnosis suggestions, heart rate, and rhythm.

Intended Use

This model is suitable for:

• Research in medical AI and computer vision: A valuable tool for academic and developmental purposes.
• Educational demonstrations: Useful for teaching and learning ECG interpretation.
• Development of clinical decision support prototypes: Can serve as a component in early-stage medical software development.

Limitations

• Primarily trained on adult ECG data; performance on pediatric ECGs may vary.
• NOT validated for clinical use and should not replace professional medical diagnosis. It is for research and educational purposes only.