ECG-Vision LV©

Automated Diagnosis of LV Systolic Dysfunction from Electrocardiographic (ECG) Images.

For DEMONSTRATION only and not intended for clinical use. The approach for the development of this tool is available on Circulation.

If you are interested in joining us as a validation partner, please contact us.

For other recent developments in Automated ECG Diagnosis, please check out our study Nature Communications 2022, and the tool ECG Dx©

INSTRUCTIONS

  • Watch a brief video to learn appropriate use. The steps, expected outputs, and preprocessing considerations are also outlined below.

  • Upload a scanned image of a 12-lead ECG.

  • The tool is designed to identify ECG images with patterns suggestive of reduced LV systolic function (defined as an LVEF < 40%), meriting further echocardiographic evaluation.

  • The tool will report the predicted probability of LV Systolic Dysfunction.

  • It is meant only for 12-lead ECGs.

  • Technical specifications of uploaded images:

    • ECG image should be upright, i.e., in landscape mode, either as a digital or scanned image of ECG. Users should manually crop the image to the graph area and remove all metadata while cropping, including patient identifiers. This is the ideal input for the application.

    • A photograph of an ECG, if used, should be as similar to a scanned image as possible and obtained in the same orientation against a light background.

    • Preprocessing of the image includes rotation corrections to align with a horizontal image in landscape mode. A cropping function will automatically detect and crop out the remaining metadata if an image with metadata is uploaded. Finally, the image is converted to grayscale, and brightness and contrast are evaluated, and out-of-distribution ECGs for brightness or contrast are flagged without generating predictions. For ECGS that are processed, brightness and contrast are rescaled, and predictions are generated.

    • The uploaded image (Left) and a preprocessed output image (Right) should be reviewed to ensure the entire ECG area was appropriately represented and the preprocessing did not remove any portions of the ECG. Please confirm these before reviewing the generated predictions. The “Normalize” check box is selected as default, though if predictions on the original image are needed, please uncheck the box. We suggested keeping the box checked.

Copyright: Cardiovascular Data Science (CarDS) Lab, Yale School of Medicine, 2022