180 — Deep Patient: An Unsupervised Representation to Predict the Future of Patients from the Electronic Health Records

This paper uses around 700,000 patient records from a Mount Sinai electronic health record (EHR) data warehouse in a neural network model that aims to predict patient outcomes based upon their histories and the outcomes of similar patients in the dataset.

This method of informing diagnosis doesn’t replace physicians; it instead helps to find difficult-to-see relationships that a human observer could easily miss.

The researchers designed an encoder that maps a patient into a latent space so that it can be understood by a neural network. Conventional embedding methods, they argued, were inadequate to describe a patient because they were often either sparse or redundant, and rarely embedded nicely.

The results of using their new embedding strategy? The researchers were able to correctly predict patient outcomes for patients with common conditions like certain types of cancer, diabetes, and some varieties of heart failure.

The system is limited, however, by the available ICD codes loaded into the system, and cannot predict outcomes for conditions it has only rarely seen. But the authors note that these limitations are manageable, and can be mitigated by exposing the system to a larger featureset and dataset size.