Using Machine Learning to Diagnose Bacterial Sepsis in the Critically Ill Patients

摘要

Sepsis is a life-threatening organ dysfunction caused by a dysregu-lated host response to infection. Early antibiotic therapy to patients with sepsis is necessary. Every hour of therapy delay could reduce the survival chance of patients with severe sepsis by 7.6%. Certain biomarkers like blood routine and C-reactive protein (CRP) are not sufficient to diagnose bacterial sepsis, and their sensitivity and specificity are relatively low. Procalcitonin (PCT) is the best diagnostic biomarker for sepsis so far, but is still not effective when sepsis occurs with some complications. Machine learning techniques were thus proposed to support diagnosis in this paper. A backpropagation artificial neural network (ANN) classifier, a support vector machine (SVM) classifier and a random forest (RF) classifier were trained and tested using the electronic health record (EHR) data of 185 critically ill patients. The area under curve (AUC), accuracy, sensitivity, and specificity of the ANN, SVM, and RF classifiers were (0.931, 90.8%, 90.2%, 91.6%), (0.940, 88.6%, 92.2%, 84.3%) and (0.953, 89.2%, 88.2%, 90.4%) respectively, which outperformed PCT where the corresponding values were (0.896, 0.716, 0.952, 0.822). In conclusion, the ANN and SVM classifiers explored have better diagnostic value on bacterial sepsis than any single biomarkers involve in this study.