search

greenelab / deep-review

A collaboratively written review paper on deep learning, genomics, and precision medicine
https://greenelab.github.io/deep-review/
Other
1.25k stars 270 forks source link

RETAIN: Interpretable Predictive Model in Healthcare using Reverse Time Attention Mechanism #189

Open blengerich opened 7 years ago

blengerich commented 7 years ago

https://arxiv.org/pdf/1608.05745v3.pdf

Accuracy and interpretation are two goals of any successful predictive models. Most existing works have to suffer the tradeoff between the two by either picking complex black box models such as recurrent neural networks (RNN) or relying on less accurate traditional models with better interpretation such as logistic regression. To address this dilemma, we present REverse Time AttentIoN model (RETAIN) for analyzing Electronic Health Records (EHR) data that achieves high accuracy while remaining clinically interpretable. RETAIN is a two-level neural attention model that can find influential past visits and significant clinical variables within those visits (e.g,. key diagnoses). RETAIN mimics physician practice by attending the EHR data in a reverse time order so that more recent clinical visits will likely get higher attention. Experiments on a large real EHR dataset of 14 million visits from 263K patients over 8 years confirmed the comparable predictive accuracy and computational scalability to the state-of-the-art methods such as RNN. Finally, we demonstrate the clinical interpretation with concrete examples from RETAIN.

Would this paper be appropriate for either the EHR subsection of the "categorize" section or the Clinical Decision Making subsection of the "treat" section? It doesn't directly match patients with treatments, but it does prioritize clinical interpretability

agitter commented 7 years ago

We haven't fully scoped the clinical decision making subsection so I'm not sure what exactly will go there. This is potentially a good fit. Are you interested in helping write about this, or other, topics?

Tagging @AvantiShri and @akundaje for relevance to the interpretation Discussion section as well.

blengerich commented 7 years ago

Yes, I would be interested in helping to write about this. I can start a draft of the "Categorize" subsection of the treat section if you would like. If someone is already working on this, I'd also be willing to help write elsewhere.

agitter commented 7 years ago

That would be excellent, no one is working on that to my knowledge. I'll update #188.