Objective: Sepsis is a medical condition, which can arise from the human immune systems’ response to infection. There is a high mortality rate among people with sepsis, especially among people left untreated or having late intervention. Thus, early detection of the disease is of paramount importance for successful treatment. The aim of this study is to develop a machine learning approach to predict the onset of sepsis.
Methods: We propose a deep neural network architecture to analyse features derived from time-windowed sensor measurements, such as heart rate and diastolic blood pressure, and static features, such as age and gender. The proposed neural network has a low-computational complexity, consisting of a few hidden layers, which enables our model to be run in workstation agnostic manner, with or without additional hardware acceleration, in real-time. Our model is causal in the sense that only current and previous information in time is used to predict the sepsis event. This ensures that the prediction of the event is not influenced by any future information.
Results: The proposed method is assessed on the PhysioNet/Computing in Cardiology Challenge 2019 dataset for early prediction of sepsis. We have evaluated our model on the training data using 8-fold cross-validation and the following metrics: PhysioNet utility function, AUC, accuracy, sensitivity, specificity, and precision. The mean (standard deviation) of the aforementioned metrics are: 0.329 (0.021) utility, 0.764 (0.015) AUC, 0.782 (0.012) accuracy, 0.621 (0.029) sensitivity, 0.785 (0.012) specificity, and 0.050 (0.002) precision. Future directions of our work will include neural network architecture optimization, and addressing the missing measurements problem by integrated probabilistic model utilizing time and/or intervariable correlations, in order to improve upon our current results.