Sepsis-induced coagulopathy has poor prognosis; however, there is no established tool for predicting it. We aimed to create predictive models for coagulopathy progression using machine-learning techniques to evaluate predictive accuracies of machine-learning and conventional techniques. A post-hoc subgroup analysis was conducted based on the Japan Septic Disseminated Intravascular Coagulation retrospective study. We used the International Society on Thrombosis and Haemostasis disseminated intravascular coagulation (DIC) score to calculate the ΔDIC score as ((DIC score on Day 3) - (DIC score on Day 1)). The primary outcome was to determine whether the predictive accuracy of ΔDIC was more than 0. The secondary outcome was the actual predictive accuracy of ΔDIC (predicted ΔDIC-real ΔDIC). We used the machine-learning methods, such as random forests (RF), support vector machines (SVM), and neural networks (NN); their predictive accuracies were compared with those of conventional methods. In total, 1017 patients were included. Regarding DIC progression, predictive accuracy of the multiple linear regression, RF, SVM, and NN models was 63.7%, 67.0%, 64.4%, and 59.8%, respectively. The difference between predicted ΔDIC and real ΔDIC was 2.05, 1.54, 2.24, and 1.77 for the multiple linear regression, RF, SVM, and NN models, respectively. RF had the highest predictive accuracy.
Keywords: algorithms; artificial intelligence; disseminated intravascular coagulation; machine learning; sepsis.