Improving Mortality Prediction using Text Embeddings

In high intensity low resource settings (e.g. mass casualty events) it may not always be possible to triage patients using the resources available. How can we improve on mortality prediction systems without using sensory data that are often available only in ICUs to help hospital system triage patients? In this project we explore the utility and performance of predicting mortality using text embeddings generated from doctor's note from the MIMIC-3 dataset, a publicaly available clinical database, and quantify its effects when the text embedding is used in conjuction with existing clinical features.