Mortality prediction of rats in acute hemorrhagic shock using various machine learning techniques

Abstract

Casualty triage at initial contact on scene by first responders is crucial to save hemorrhagic patents'' lives in civilian trauma or on the battlefield. However, there are few studies to demonstrate important physiological variables related to mortality and the most suitable machine learning techniques for casualty triage among various techniques. This study was conducted to suggest a mortality predicting model for casualty triage on scene using various machine learning techniques through a rat model in acute hemorrhage. Furthermore, this study determined important physiological variables for mortality prediction to effectively minimize impression time for first responders.Thirty-six anesthetized rats were randomized into three groups according to volume of controlled blood loss. Uncontrolled hemorrhage was induced simultaneously by combination of controlled blood loss and tail amputation in all rats. This study measured heart rate (HR), systolic and diastolic blood pressures (SBP and DBP), mean arterial pressure (MAP), pulse pressure (PPR), respiration rate (RR), temperature (TEMP), lactate concentration (LC), peripheral perfusion (PP), shock index (SI, SI=HR/SBP), and the new hemorrhage-induced severity index (NI, NI=LC/PP) as candidates for input variable of each machine learning techniques. All variables were analyzed for changes between pre- and post-hemorrhage to investigate the effects of hemorrhage on mortality.The training data set was used to construct models based on popular machine learning techniques including logistic regression (LR), artificial neural networks (ANN), random forest (RF), and support vector machines (SVM). To select important variables for mortality prediction in a rat model, variables selection was performed with algorithm of consistency subset evaluation using 10-fold cross validation. The testing data set was used to assess the performance of the optimized models consisting of the selected variables for predicting mortality using sensitivity, specificity, accuracy and area under curve (AUC) of the receiver operating characteristic (ROC).For the LR model, sensitivity, specificity, accuracy, and AUC were 0.678, 1.000, 0.833, and 0.833, respectively. For the ANN model, sensitivity, specificity, accuracy, and AUC were 0.833, 1.000, 0.917, and 0.917, respectively. For the RF model, sensitivity, specificity, accuracy, and AUC were 0.833, 1.000, 0.917, and 0.903, respectively. For the SVM model, sensitivity, specificity, accuracy, and AUC were 1.000, 0.833, 0.917, and 0.972, respectively. The SVM model showed better AUC than that of the LR, ANN, and RF models for mortality prediction. The important variables selected by SVM were LC and NI.In conclusion, the SVM model with selected variables, LC and NI, was superior to the LR, ANN, and RF models in predicting mortality resulting from acute hemorrhagic shock in a rat model. These machine learning techniques may be very helpful to first responders to accurately make causality triage decisions and rapidly perform proper treatments for hemorrhagic patients in civilian trauma or on the battlefield in the future.