Convolutional and Recurrent Neural Networks for Early Detection of Sepsis Using Hourly Physiological Data from Patients in Intensive Care Unit

PhysioNet Challenge 2019 aims to develop novel solutions for early detection the clinical onset of sepsis. We used training date provided and develop machine learning classifiers to predict clinical sepsis 6 - 12 hours ahead of the clinical onset.Neural networks possess abilities to uncover insights from complex datasets. We have trained an ensemble classifier with a convolutional neural network (CNN) and a recurrent neural network (RNN) for early detection of sepsis. The classifiers were implemented in in Python using Keras with Tensorflow as back-end. Both networks were combined using bagging to achieve better performance.The database appeared imbalanced, and the class (positive) with small number of data entries was oversampled proportionally before training. 90% of the augmented and oversampled data were used for training, with 10% for testing. Our team "Leicester Fox" achieved accuracy of 90.6%, AUROC of 0.745 and AUPRC of 0.079 in the test set A. The entry for the official phase of the PhysioNet/CinC 2019 competition received a normalized utility score of 0.237 on full data. Machine learning and neural networks approaches showed potential application for better prediction of sepsis, using real-world database with random missing data and imbalanced classes.