Two Stage Deep Neural Networks for Early Prediction of Sepsis

Naoki Nonaka and Jun Seita
RIKEN


Abstract

Sepsis is a body’s response to infection, causing tissue damage, organ failure, or death. Early detection of sepsis leads not only improving patient survival but also reducing medical costs. Therefore, in this study, we aim to predict the onset of sepsis from the values measured in Intensive Care Unit (ICU).

In ICU, each variable is measured at different frequency or at irregular interval. Therefore, the data obtained in ICU contains a large number of missing values causing poor prediction accuracy. To address this issue, we propose two stage model. The first stage is a data imputation stage using deep generative models, and the second stage is prediction stage using gated recurrent unit (GRU). The deep generative model is a method of modeling the data generation process by deep neural network, and is used in tasks such as image noise removal and image inpaiting. The proposed method uses two deep generative models, VRNN and VAE. VRNN is used to model relationships in time series direction, and VAE is used to model relationships between variables at each time step. Then VAE and VRNN are combined to predict a missing values in the data. The missing value-supplemented data is then given to the GRU to predict onset of sepsis. In addition, class weight was introduced to penalize misclassification of positive examples because the number of positive examples is overwhelmingly small compared to negative examples in the published data set,

The published dataset was split into train, valid and test set. Our model was trained using 32,268 train samples and validated using 4,034 samples. Five independent experiments were conducted using 4,034 test samples, and as a result of evaluation, the utility score achieved 0.323.