This repository is to host models and satasets for Heart Transplant rejection prediction using longitudanal ECG data. There are 3 main types of Models that we are experimenting with:
Model 1 is just a simple 1d CNN to try and predict transplant rejection from a single ECg sample
Model 2 employs the Model 1 architecture as a twin feature extraction layer for 2 longitudanal ECG samples for a single patient These samples are passed onto a feed forward network alognside normalized time information, and Biopsy results for sample 1.
One version of this model just concatenates the features extracted from the CNN layer with the time difference and biopsy label.
The second version of this model concatenates (sample2-sample1)/dt, sample 1 and the biopsy results for sample 1.
Not implemented yet Goal of model 3 is to use a transformer to learn if past time points of ECGs can lead to a better prediction of rejection at the current time point.