Design And Implementation Of Real-warning Model For Spetic Shock Based On Deep Learning

Sepsis is life-threatening organ dysfunction caused by a dysregulated host response to infection.Septic shock currently refers to a state of acute circulatory failure associated with infection.Sepsis and septic shock are major critical care problems,affecting millions of people around the world each year,and one in four patients died.Early detection and treatment improves patient outcomes,and as such it poses an important challenge in medicine.Although early warning research was conducted in the past years,most of the researchers build the model with the static characteristics of the patients when they were admitted to the hospital.Few studies were conducted from the perspective of the patient's dynamic disease trajectory.In this thesis,we discuss the work for three parts:1)we use Medical Information Mart for Intensive Care III to construct septic shock cohortaccording to “The Third International Consensus Definitions for Sepsis and Septic Shock(Sepsis-3)” released in 2016,and extract streaming lab results,vitals,and medications.2)We develop real-time warning model on the septic shock dataset.The trajectories of thephysiological variables are modeled by recurrent neural network(RNN).We develop fournetwork,which are named Single-rate RNN,Multi-rate RNN,Multi-rate multi-goal RNNand Multi-rate multi-weighted-goal RNN,to solve problems such as irregularly spacedtime series,imprecise labels for the true time of septic shock onset.3)Models are trained on train set and test on test set.The test results show that the networkmodel developed in this paper can effectively predict septic shock before it occurs on thedataset.It shows that auc is 0.91 on the test set when it predicts at septic shock onset,andauc is 0.83 when it predicts at 6 hours before onset.The performance of neural model isbetter than baseline models such as logistic regression.