OHDSI Home | Forums | Wiki | Github

Looking for collaborators to validate stacked autoencoder from OMOP-CDM


(Michael Rosenberg) #1

I posted something similar a while back, but we just completed an unsupervised analysis using 5-layer stacked autoencoder on 26k OMOP-CDM features (diagnosis codes and medications) in 2.3M patient EHR, and are looking for collaborators to validate/test replication accuracy. Happy to share our weights and model specs (Python code). Just let me know if interested.


(Tim Bunnell) #2

Very much interested in this work. We’ve been looking into applying a similar approach on the Nemours pediatric EHR data.


(Ifeanyi Agwulonu) #3

I am interested. I’d like to know how I can help - ie if I meet your criteria


(Michael Rosenberg) #4

Awesome, message me your contact information and I’ll send a file with the list of the OMOP concepts (i.e., features) we used in our model to see if there’s enough overlap to see about sending our weights.


(Andrew Williams) #5

I’d like to investigate whether we can participate also.
Contact me at: awilliams15 at tuftsmedicalcenter dot org.


(Christian Reich) #6

@michael.a.rosenberg:

The easiest would probably be if you posted the encoder, and then announced your work as a network study. And then we can help getting you participants no problem.


(Michael Rosenberg) #7

The manuscript where we describe the model is in submission, so might need to wait until publication before posting the model weights anywhere. Here’s the preprint: https://www.biorxiv.org/content/10.1101/520866v1

I’m happy to post the features we used as CSV file. If there’s a mechanism for this please let me know.


t