181 — Scalable and accurate deep learning for electronic health records

Much like yesterday’s paper, today’s paper is about predicting patient outcomes and medical events using a deep-learning approach on historical medical data.

A lot has changed in the nearly two years since Miotto et al was published, and of course the deep-learning world looks a bit different through the lens of the endless fields of GPU clusters at Google. But today’s paper differs in approach from yesterday’s in a few important ways.

First, this system performs no “site harmonization,” or reconciliation of patient records from different medical centers. This means that there is much less overhead when deploying this system, because no human needs to proofread the data first.

Second, the system makes no suppositions about the structure of the data. This is pretty classic Google style (which isn’t necessarily a bad thing!) but putting all of your features into a big Feature Soup and blasting it through a net is a capability generally only possible for teams with a lot of compute. On the other hand, this generally leads to smarter, more refined models that can run on limited resources, so I’m certainly not complaining.

The model was able to predict inpatient death and unplanned 30-day readmission events with extremely high accuracy, and was able to point to other high-probability medical events in certain cases. The researchers also demonstrate explainability by providing a list of the features most heavily used when generating a prediction.