Deep Patient Representation of Clinical Notes Via Multi-Task Learning For Mortality Prediction

Authors

Yuqi Si, University of Texas Health Science Center at Houston, School of Health Information Sciences, Houston TX, USA
Kirk Roberts

Publication Date

1-1-2019

Journal

AMIA Summits on Translational Science Proceedings

Abstract

We propose a deep learning-based multi-task learning (MTL) architecture focusing on patient mortality predictions from clinical notes. The MTL framework enables the model to learn a patient representation that generalizes to a variety of clinical prediction tasks. Moreover, we demonstrate how MTL enables small but consistent gains on a single classification task (e.g., in-hospital mortality prediction) simply by incorporating related tasks (e.g., 30-day and 1-year mortality prediction) into the MTL framework. To accomplish this, we utilize a multi-level Convolutional Neural Network (CNN) associated with a MTL loss component. The model is evaluated with 3, 5, and 20 tasks and is consistently able to produce a higher-performing model than a single-task learning (STL) classifier. We further discuss the effect of the multi-task model on other clinical outcomes of interest, including being able to produce high-quality representations that can be utilized to great effect by simpler models. Overall, this study demonstrates the efficiency and generalizability of MTL across tasks that STL fails to leverage.

PMID

31259035

PMCID

PMC6568068

PubMedCentral® Posted Date

May 2019

PubMedCentral® Full Text Version

Post-Print