In-hospital mortality prediction for patients with congestive heart failure using long short-term memory and gated recurrent units
Loading...
Issue Date
2021-05
Authors
Silva Ferreira, Brenno William
Advisor
Sinha, Kaushik
Citation
Abstract
Mortality prediction is a determinant factor in Intensive Care Units (ICU). The high costs
associated with a patient’s stay in an ICU can place a burden on any health care system around the
world. The ability to accurately predict the risks of patients dying can provide important insight to
help hospitals manage resources efficiently and therefore reduce costs and help reduce the number
of fatalities caused by CHF. Another crucial factor is the data availability which has become more
prominent with the usage of electronic health records (EHR) by the majority of hospitals. This
makes it possible to extract patient’s health records during their stay in an ICU, allowing models
to use that information to make precise predictions. In this analysis, two versions of the same
dataset are used. One version contains measurements of various chart variables while the other
version contains measurements of chart and lab variables. Long short-term memory (LSTM) and
Gated Recurrent Unit (GRU) were used to predict if a patient will die after a gap window of at
least 4 hours, 8 hours and 12 hours after the last measurement, where the gap window is a time
interval that separates the time of the last measurement used for making a prediction and the time
of the actual event (death in hospital or discharge from hospital). For all classifiers, as the gap
window increases, the model performance decreases as well. Moreover, it is clearly noticeable that
the performance increases as the lab data is included into the training process of the predictive
model. The highest AUC score obtained was 88.53% using LSTM with a 4-hour prediction window
on chart and lab variables. The smallest AUC score was approximately 72.07% using LSTM
with a 12-hour prediction window on chart variables only. Sensitivity and specificity obtained for
each model and prediction windows is also described in detail in this thesis.
Table of Content
Description
Thesis (M.S.)-- Wichita State University, College of Engineering, Dept. of Electrical Engineering and Computer Science