search

AI4HealthUOL / SSSD-ECG

Repository for the paper: 'Diffusion-based Conditional ECG Generation with Structured State Space Models'
MIT License
36 stars 4 forks source link

ECG Synthesis using MIMIC-IV ECGs #19

Open sanketahegde opened 5 days ago

sanketahegde commented 5 days ago

Hello @juanlopezcode ,

I have been trying to process the MIMIC-IV ECGs to use for training your SSSD-ECG model with the intention of generating ECGs. I also created multi-labels and ended up with a label length of 20. Currently, there are around 200,000 ECGs which I use for training which is quite large compared to PTB-XL dataset. At the end, I have my training set with data shape - (220000, 12, 5000) and label shape - (220000, 20). I already tried training once with these ECGs but the model did not learn properly and also started over-fitting after 6 epochs.

Now, my question is what modifications must I do to the currently available SSSD-ECG model setup to train and generate with MIMIC-IV ECGs?

Any suggestions are appreciated. Thank you!

juanlopezcode commented 4 days ago

Hi, I have never trained in the 5000 time step setting, perhaps is too long? I would try first at the 1000 (resample from 500hz to 100hz). also, not sure how the overfitting was measured or what it represents here. I would keep intact the network, and just train on 100Hz, it should already learn a lot because you have a lot of data.

sanketahegde commented 4 days ago

Thank you, I will try the training again with the resampled ECGs and get back if I still have some doubts.