Representation Learning for Early Sepsis Prediction

Abstract

As part of the PhysioNet/Computing in Cardiology Challenge 2019, we propose a neural network called AECNet to early detect sepsis based on physiological data. AEC-Net consists of two main components: 1) an Auto Encoder for dimension reduction and feature extraction, and 2) a Fully Connected Neural Network (FCNN) taking the extracted features by the Auto Encoder as the input and generating prediction of sepsis as output. The losses of both the Auto Encoder and FCNN are minimized concurrently. This concurrent optimization helps AEC-Net to have a better generalization and the extracted features by Auto Encoder to be more relevant to the classification problem. Finally, we propose an ensemble method of AECNet, Random Forest and Gradient Boosting Decision Trees to achieve a better prediction. We train our proposed models using data from 40336 patients with 40 physiological features ranging from 8 to 336 hours. Our team Infolab USC evaluated Ensemble with the hidden full test set of the Physionet Challenge 2019, and achieved a Utility score of 0.284 and 24 place in the challenge.