Early Sepsis Diagnosis with Recurrent Neural Networks

Dariusz Kucharski, Dominik Rzepka, Mateusz Pabian
Comarch S.A.


In the stage of data preprocessing the completion of missing values in the input time series was applied, using three strategies. For the case of absence of some data points in the given feature, their values were obtained using linear interpolation. If the feature value was provided only for single time point, this value was assumed also for the other time instants. For the records with entirely missing feature values, its values were set to zeros. Additionally, all the input values were supplemented with additional Boolean attribute, indicating whether their value was originally present, or completed artificially. To handle the class unbalance in the dataset, all records with at least one appearance of sepsis were selected along with the same number of randomly chosen records of healthy patients. The dataset was then divided into training, validation and test sets. For the data points classification, recurrent neural networks were used. The model consisted of two LSTM layers with 512 hidden units each, and a dense projection layer with 256 neurons. The output layer was a single neuron with sigmoid activation function, which returned a probability of sepsis in a current time point. To optimize the decision threshold for the neural network output we minimized f1-score using logistic regression to find a compromise between accuracy and utility function value. Finally 0.63 AUROC and 0.22 AUPRC score for neural network output and 64% accuracy and 0.41 points for utility functions for decision threshold were achieved.