Early Detection of Sepsis Using Feature Selection, Feature Extraction, and Neural Network Classification
Erik Gilbertson, Khristian Jones, Abigail M Stroh, Bradley M. Whitaker · Computing in cardiology · 2019
Introduction: This work represents an entry to the 2019 PhysioNET Computing in Cardiology Challenge.Algorithm: Using the supplied biomedical data, we reduce the original 40 features to 10 principal components.One additional feature is generated from a quick Sequential Organ Failure Assessment (qSOFA) score.These 11 features are then fed into a deep neural network classifier implemented in Tensorflow.The features associated with each hour are analyzed independently.A sigmoid function is used for the activation functions, RMSprop for the optimizer functions, and categorical cross entropy for the loss function.We have found that this setup works best for our current method and leads to the highest accuracy with minimal loss.Results: By testing our algorithm on a subset of the given dataset we achieved a validation score of 0.038.The official score received from the competition (under the team name "Whitaker's Warriors") was 0.022 placing us 65th out of the 78 scored entries.Conclusions: Achieving a positive utility score of 0.022 shows our method of combining PCA with a quick SOFA score and classifying with a neural network is a promising approach.More work in the future could be done to increase the accuracy of the model by adding additional features to the input of the classifier and adjusting the parameters of the neural network.