Authors: Kenan Li*, University of Southern California, Katherine Sward, University of Utah, Sandrah Eckel, University of Southern California
Topics: Medical and Health Geography
Keywords: Asthma Symptoms, Long Short Term Memory neural network, Convolutional Autoencoder, Deep Learning, mHealth
Session Type: Paper
Start / End Time: 9:55 AM / 11:35 AM
Room: Wilson A, Marriott, Mezzanine Level
Presentation File: No File Uploaded
Our goal is to develop and apply a method using deep learning for automatic identification of the key exposures related to asthma outcomes. These methods would be an alternative, data-driven approach in contrast to the typical “brute force” feature engineering approaches wherein numerous summary statistics are calculated a priori from multivariate exposure time-series and offered as inputs to machine learning models. We collected data on 11 patients participating in this panel study, which contains rich temporal information on minute-level exposures (e.g., residential indoor and outdoor PM, temperature, and relative humidity) and daily and weekly questionnaires with questions on asthma control and medication use. First, we trained a Convolutional Auto-Encoder (CAE) to compress the hourly average multivariate exposure time-series. We will explore this latent-space representation to discover patterns in the exposure data, and potentially identify new relevant transformations of the exposures. Second, we built a Long Short Term Memory (LSTM) neural network to predict daily asthma outcomes allowing for long- and short-term effects of exposures (CAE latent representations) and previous outcomes. Through approaches like those outlined above, we aim to develop methods for using CAE and LSTM in personalized medicine and environmental epidemiology.