Sepsis Real-Time Risk Prediction Model for Intensive Care Unit Patients Based on Machine Learning
Sepsis is a syndrome of organ dysfunction caused by the body's dysfunctional response to infection,with high morbidity and mortality.The traditional scoring system has low specificity.Based on the LightGBM machine learning framework,this study proposed a model for early prediction and risk assessment of sepsis to provide timely intervention for patients with potential risk of sepsis.In order to realize the model,a time series feature construction method based on LASSO feature selection and sliding window path reintegration and a time series clustering sampling method based on dynamic time regularization algorithm were proposed.We selected clinical information from 29 239 patients in the MIMIC-Ⅲ dataset and 40 336 patients in the PhysioNet/CinC 2019 challenge dataset to train and validate the model.The sensitivity,specificity,and area under the receiver operation characteristic curve(AUC)of the proposed model on the MIMIC-Ⅲ and PhysioNet/CinC 2019 independent test sets were 0.737 7,0.730 4,0.814 7 and 0.802 6,0.789 1,0.873 0,respectively.Compared with the state-of-the-art method EASP,the improvement of AUC was 3.62％and 2.83％respectively.In conclusion,the established model could predict the risk of sepsis in real time,reveal the important factors affecting the occurrence of sepsis,and provide a basis for timely intervention of people at risk of sepsis.
sepsis predictionmachine learningfeature constructiontime series sampling
NETL
NSTL
万方数据
