Scalable and accurate deep learning for electronic health records


I saw the reception that this got on Twitter, but I’m still curious about the data model they are using.

For background, I’m a physician who took the 2017 versions of both courses and have been doing some deep learning with EHR data. I’ve gotten pretty good results with everything that I learned in this course: ~92% accuracy in predicting hospital length of stay less than 24 hours vs greater than 24 hours.

However, most of my data is specifically selected and I’d love to implement this model to include more available information which might be relevant. I basically copied the data model from a paper in one of the later course 2 lessons about PICU mortality:

From this paper, can we tell exactly how they formatted the data to feed into the model? Is it basically the same as the PICU paper? It seems for the RNN model they essentially did the same thing, but kept the different data types separate and then later concatenated them and then fed them into a new RNN? How does anything about FHIR fit in to any of this?

Thanks for your suggestions.

1 Like