JupyterCon in New York 2018

About this course

hi everyone as you've just heard I'm Kristin Capdevila I'm a data scientist at healthcare startup called prognosis and I'm here to tell you about what we

do and how AWS stage maker helps us do it so this is our moon shot and as you can see we're an ambitious company with a very ambitious moonshot disease eradication more immediately as you can

see after step one we're trying to learn to predict disease early and this is a problem obviously that has immense social and commercial impact as can be seen by a list of partners and investors

here so we sell products into these three verticals life sciences diagnostics and payers here you can see some of our our bigger clients and investors and so the question is how are

we doing this well we're combining three things one our teams of data scientists and engineers – an experienced group of clinical experts and three our unique data asset we have the world's largest

clinical laboratory data set over 17 billion de-identified records describing the patient journeys of 200 million people here in the United States so here's an overview of our predictive

modeling pipeline I imagine this looks familiar to a lot of you out there we ingest data from disparate sources on the left and impose our schema on it then we use a

combination of domain expertise and machine learning to harmonize it by cleaning up things like different naming conventions different units of measure and to organize and tag relegating sorry

relevant information for use downstream extracting data from text fields is a good example and for the rest of this short talk I'll be focusing on this section in the bright pink box which

runs on AWS sage maker so even after cleaning and harmonizing lab data it can be hard to work with it's irregular time series data different patients have different numbers of tests performed at

irregular intervals it's also high-dimensional we have tagging logic for over 7,000 different types of tests and a lot of those test records could have multiple

ICD diagnostic codes of which there are roughly 68,000 and there's various other things attached and you end up in a very very large space and it's sparse some of these tests are quite rare so you aren't

going to see them very often and a lot of you are probably also familiar with this or some variation on it if you're a data scientist the stuff that you're interested in is in orange but the vast

majority of the initial man-hours and code are involved in all the blue stuff down below propping up those notebooks and the model code and like anything else all those pieces have to be

maintained so sage maker for us has let us focus on the stuff that we care about and offload development and management of all the other pieces to Amazon we can prototype our models in a sage maker

notebook and then very easily deploy that to a cluster of gpu-accelerated house for large scale training and once we've produced our model artifact we can deploy it as an endpoint in Sage Maker

as well and we're ready to make predictions if however we do need to reach further down the stack we can provide our own docker images or Fork Amazon since they're on github and our

own build of a framework and run those on their GPU docker hosts and use their deployment tools so that was a really big selling point for us that we're not necessarily locked into any particular

tooling we can replace as much as we need when we need to and this has let us focus on this this big picture idea we've able to put our limited resources and time to what we really care about

which is predicting disease that's the end of my talk we are higher ringg please send resumes I'll be walking around so if you have any questions come up in and talk to me

thanks everyone you