Abstract
Sepsis is one of the leading causes of death in hospitals across the world, and it has attracted more and more attention in increasingly aging countries. Every year 5.4 million people worldwide die of sepsis. With the development of social life, predicting sepsis has become more and more important. The main sign of sepsis is multiple organ failure. In 2016, advances in medicine and technology helped redefine the disease standards for this disease. This thesis uses the Sepsis-3 standard to study adult patients. Infected patients with a sequential organ failure assessment (SOFA) score higher than 2 are marked as sepsis patients.
Nowadays, with the advancing at a rapid pace of data mining and artificial intelligence(AI), people’s research on the problem of sepsis prediction has become more and more in-depth. This thesis mainly focuses on the prediction of the probability of septicemia among patients in the intensive care unit(ICU). We have developed a deep temporal convolutional network to predict sepsis. At the same time, a machine learning model (decision tree) and a deep learning LSTM model have been developed as the test benchmark model. MIMIC-III is the source database for model development,validation and testing. Our goal is to use 12-hours observational health data to predict whether sepsis will occur in following 6 hours. Our innovation is to mark sepsis with the time of onset instead of the ICD-9 code. The project first used Postgres to extract relevant data from MIMIC-III, and performed data preprocessing, and then established one machine learning model for sepsis prediction, Decision tree and two deep learning models TCN and LSTM, Decision tree and LSTM model as a benchmark model to verify the performance of the TCN model. The three models are optimized separately. The decision tree uses GridSearchCV to automatically adjust the parameters max_depth, and finally the best max_depth is selected as 5. LSTM and TCN a6re optimized by setting epochs, the best model is the model with the highest verification accuracy for 20 iterations. Evaluation metrics (Accuracy, Precision, Recall, F1-score, and AUC-ROC) will be used to measure the performance of the model. When predicting sepsis 6 hours before onset on the new reality label, the area under the ROC curve of our proposed TCN model is 0.944, the accuracy is 0.893. The results show that, compared with machine learning methods and LSTM, time convolutional networks converge faster and have better performance. The model is robust and high-precision, and may be used as a tool for hospital sepsis prediction.
