Predictive Modeling On Its Head: A Pipeline That Finds Cancer, Asthma and Hemophilia

What do asthma and hemophilia have in common? Is that 25 year old with heart failure real or a data error?

These questions were some of the many challenges the Oscar data team had to overcome to create accurate health profiles for our members. The ACA (Affordable Care Act) includes in its regulations a provision that gives insurers a financial reward for knowing exactly which conditions our members have.

Unfortunately, doctors and hospitals don't have this motivation and they often forget to document what the health status of their patients is. We need to infer which of our members are being treated for conditions that have not been documented in our data.  

Tackling this problem involved building a pipeline that starts with claims data - the data we get from healthcare providers in order to pay them; and builds from there a robust ecosystem that can add signals from different sources. Prescription data, self-reported health statuses, prior authorization requests, and hospital procedures are all inputs into a large scale ML model that tells us what conditions our members are likely to have. 

This talk will cover how the data science team at Oscar solved this problem and drove real business value through this work, covering four main lessons we learned along the way.