Why readmission prediction result lower than GCT's paper

NYUMedML / GNN_for_EHR

Code for "Graph Neural Network on Electronic Health Records for Predicting Alzheimer’s Disease"

GNU General Public License v3.0

251 stars 59 forks source link

Why readmission prediction result lower than GCT's paper #8

Open LeslieHoloway opened 3 years ago

LeslieHoloway commented 3 years ago

In your paper, the result of the readmission task based on eICU dataset is 0.3986. GCT's AUPRC is 0.3794.

But in GCT paper (Learning the Graphical Structure of Electronic Health Records with Graph Convolutional Transformer), the result of the readmission task based on eICU dataset is 0.5313.

What makes such a difference? I ran both codes from your repo and theirs. I found out that the results were close to yours.

jackzhu727 commented 3 years ago

I met the same problem when I was doing the experiment referring to the repo you mentioned, so I only reported the result from the code. I also notice "shallow" method in the table above has a closer result in terms of scale, but I don't really know the reason for the gap.