Run model and evaluation code questions

[Ján Pavlus]

Dear challenge team,

I have one bug (maybe) and two questions

1. Image generation is not generating the #Image label into the header file, which is used to load by run_model.py to find all images of the recording. I think this is the bug.

Now the two questions:

2. Load image function loads more image files for one record. From the image generator code, I expect that it is for signals longer than 10 seconds. The generator as the standard ECG plotter jumps over 10 seconds and prints out only the first 2.5 seconds for each lead (and a full 10 seconds for selected ones). The evaluation script loads the target signal, our estimated signal, and trim or pad the estimated signal to the target one. Then computes the SNR between them.
What will the target signal look like? Would it be only the 2.5 seconds? Because, for the 10-second signal, the estimated one will be padded by zeros, which gives a minimal SNR score as 3/4 of the signal will be padded by the zeros.

3. The connected question to the second one is what to do with the full-mode leads on the image. Should it be a matrix of 12 leads where one or more of the leads will have a length of 10-second samples and other leads padded by zero, or something else?

[PhysioNet Challenge]

Hi Ján,

Thanks for the good questions.

The add_image_filenames.py script adds the image filenames to the WFDB header files:
https://github.com/physionetchallenges/python-example-2024/blob/main/add_image_filenames.py

Please see the data processing scripts in this README to prepare data from the different data sources consistently (mostly adding and formatting information consistently in the header files, except for the part addressed by your next questions):
https://github.com/physionetchallenges/python-example-2024/

As you noticed, by default, the ECG image generation code plots an image with the first 2.5 seconds of the signal for each lead and the full 10 seconds of the signal for selected leads. We are in the process of updating the image generation code to update the WFDB files to replace unplotted segments of the signal with NaNs, which the evaluation code will ignore, i.e., we will only evaluate the plotted segments of the ECGs. I think that this addresses your second and third questions, but please follow-up if not.

Best,
Matt
(On behalf of the Challenge team.)

Please post questions and comments in the forum. However, if your question reveals information about your entry, then please email info at physionetchallenge.org. We may post parts of our reply publicly if we feel that all Challengers should benefit from it. We will not answer emails about the Challenge to any other address. This email is maintained by a group. Please do not email us individually.