| OHDSI Home | Forums | Wiki | Github |
I really enjoyed this paper published today by the data science folks at Mt Sinai:
Deep Patient: An Unsupervised Representation to Predict the Future of Patients from the Electronic Health Records (http://www.nature.com/articles/srep26094)
Maybe no shockers in here, but I really thought the clarity of the paper was exemplary and the approach solid and systematic. I’d be curious the thoughts of those on the patient prediction WG if you get the chance.
Best,
Jon
This is impressive! I can’t figure out how they handle the time-stamped
nature of the data in the deep learning. I’ll ask.
@David_Madigan: Did you ever get a response to the “time-stamped nature of the data” question? Our assessment based on their methodology is that they don’t handle temporal conditions. Same with the Georgia Tech paper.
@jon_duke
I developed deep learning model (recurrent neural network) based on CDM to hand time-stamped nature of data. You can see the poster here.
I’m trying to make this model as an OHDSI tool now with @Rijnbeek as a subtool for PLP package.
Hi @mgkahn and all,
I don’t believe the Deep Patient approach handles temporal conditions. The focus is on representation learning from all data available for a given patient record in order to predict codes from a vocabulary of common diagnosis codes.
It would be really nice to have an approach that tackles the temporality on the representation side. We had done some work on incorporating longitudinal data looking a prediction of CKD using Kalman filters and topic models for notes and we (@aperotte, Rajesh Ranganath, Dave Blei, and I) found that (no surprise!) it helps to incorporate longitudinal data in a temporally aware fashion (more at https://academic.oup.com/jamia/article/22/4/872/1746401). Last year, we (same team) have looked into incorporating time, but on the side of predictions: In our Deep Survival Analysis paper (
http://proceedings.mlr.press/v56/Ranganath16.pdf) we compute a survival curve through time for risk of a specific code (CHD onset), given a particular snapshop of a patient’s data in time (a month). Not exactly what you are looking for, I assume, but I think including temporality is a particularly exciting area of work, and deep learning approaches can help.
I asked Dudley but did not get a response. Looking at the paper again its
fairly clear that they are ignoring time.