Modeling multivariate time-series variables in healthcare using deep learning methods
Abstract/Contents
- Abstract
- Well-being of patients or operation of healthcare systems can be evaluated by monitoring certain health metrics or workflow parameters. This monitoring is done by periodically measuring those metrics and evaluating their variations and changes over time. The resulting measurements create a set of data points along the temporal dimension, which are called time series data. Decisions to improve patient outcome or healthcare operation often depend on accurately forecasting future trends in those time series variables. Thus, time series analysis and prediction is of very high importance in healthcare provisioning. Variables in healthcare are often interrelated, and future values of a given variable depend on past observation and trends of that variable as well as other related variables. Thus, in many cases, time series prediction tasks in healthcare are multivariate. In this dissertation, we develop a multivariate time series predictive model that can simultaneously learn from multiple time series variables over a long temporal window. We implement this model as a convolutional neural network, which we call PatientFlowNet. We use this model to predict patient flow in hospital emergency departments. Specifically, we predict the rates at which patients arrive, are treated, and are discharged from the hospital. We benchmark our model against the state-of-the-art methods in patient flow prediction using data from emergency departments in three different hospitals. We then reuse our model for predicting vital signs for patients under anesthesia, namely heart rate as well as systolic and diastolic blood pressure. We observe that our model has superior prediction accuracy in the case of patient flow in hospital emergency departments where short-term and long-term trends are present but hidden in the data and variables are significantly interrelated. However, where long-term trends are not present, as is the case with patient vital signs under anesthesia, the performance of our model is similar to the best baseline method. The convolutional design of PatientFlowNet allows us to extract dependencies between input and output variables by examining the values of convolutional filters in the first layer of the network. We provide visual and interpretable representations of learned dependencies between time series variables in each study.
Description
Type of resource | text |
---|---|
Form | electronic resource; remote; computer; online resource |
Extent | 1 online resource. |
Place | California |
Place | [Stanford, California] |
Publisher | [Stanford University] |
Copyright date | 2021; ©2021 |
Publication date | 2021; 2021 |
Issuance | monographic |
Language | English |
Creators/Contributors
Author | Sharafat, Ali Reza |
---|---|
Degree supervisor | Bayati, Mohsen |
Thesis advisor | Bayati, Mohsen |
Thesis advisor | Weissman, Tsachy |
Thesis advisor | Widrow, Bernard, 1929- |
Degree committee member | Weissman, Tsachy |
Degree committee member | Widrow, Bernard, 1929- |
Associated with | Stanford University, Department of Electrical Engineering |
Subjects
Genre | Theses |
---|---|
Genre | Text |
Bibliographic information
Statement of responsibility | Ali Reza Sharafat. |
---|---|
Note | Submitted to the Department of Electrical Engineering. |
Thesis | Thesis Ph.D. Stanford University 2021. |
Location | https://purl.stanford.edu/fv290sn7548 |
Access conditions
- Copyright
- © 2021 by Ali Reza Sharafat
Also listed in
Loading usage metrics...