Potential Idea

[chriszhou0916]

@abbaphant @andrewucla @arnoldtao

I found this paper that did real-time mortality prediction from MIMIC III data. They look at a patient's chart info from the past several hours to predict whether the patient dies in the ICU. I think we can do something similar, maybe not predicting overall mortality but mortality within the next X hours in real time. This way we can tell the physician to prioritize patient who have the highest risk of dying immediately. Real time prediction is important because patient status can be very unpredictable. A good model should weigh recent statuses more.

Features we could look at include: (most are prob in chartevents table)

• patient related info like age, diagnosis, recent procedures
• heart rate, blood pressure, respiratory rate, oxygen, glucose... lab results
• medications

What do you guys think?

[andrewucla]

Good idea. Let us discuss further tomorrow.

[andrewucla]

Also I suggest using Tableau to do preliminary data visualization and exploration (although I don't know if we are required to stick with python). Overall Tableau is one of the best enterprise level data visualization software, and we can have free access to it for the first year.

[abbaphant]

Cool idea, I like the implications for triage. I'll take a closer look at the paper before class tomorrow. I agree Tableau is powerful for visualization, but does it play well with jupyter notebooks? For reproducibility sake, we'll want to be able to show the code used to generate graphs.

[chriszhou0916]

I found the code repository for the paper above. I feel like we shouldn't ask a question that requires a lot of data wrangling. Ideally, we want to get to a trainable model matrix ASAP. Everyone agree?