Aims Sepsis is a challenging problem for public health around the world. Timely detection, tracking and forecasting its early onset are crucial to support clinical decision-making. Our objective is to construct a new sepsis predictor that learn from health data and forecast whether a patient is at risk of developing the infection within the subsequent hours.
Methods The proposed deep architecture combined a Recurrent Neural Networks (RNNs) with fully connected neural network. Missing data was addressed with a two-steps imputation approach. First, we applied a backward-interpolation algorithm. Then a masking layer was added at the top of the network to exclude all remaining missing values from network calculations. Synthetic Minority Over-sampling Technique (SMOTE) was applied to handle the class imbalance in the data. Records from the PhysioNet/Challenge 2019 dataset were partitioned into three groups: sepsis, nonsepsis-sepsis and non-sepsis. The network was cross-validated on 90% and tested on 10% of the dataset, keeping the same distribution of records for each patient group.
Results The proposed algorithm achieved a 17.38 of Utility function on the hold-out dataset.