Can machine learning be used to quantify blood loss based on hemodynamic data? Experimental study of controlled hemorrhagic shock in anesthetized piglets

Background Hemorrhagic compensatory physiological mechanisms may initially lead to an underestimation of the severity of a hemorrhage, for which quantification of blood lost volume is difficult. The objective of our study was to assess the suitability of machine learning (ML) to quantify blood depletion from critical care settings routinely available clinical and ultrasound data during the management of a hemorrhagic patient in a model of controlled hemorrhagic shock in anesthetized piglets. Methods We used data from anesthetized and ventilated piglets that underwent blood spoliation at 2 mL kg −1 min −1 until a mean arterial pressure (MAP) of 40 mmHg was achieved. The data recorded consisted of the hemodynamic and echocardiographic parameters routinely measured during the care of critically ill patients. We used these data to develop six machine learning models for blood loss quantification. Judgment criteria were the coefficient of determination R 2 , root mean squared error (RMSE) and mean absolute error (MAE), reported in mL between real depletion and that assessed by the ML models. Results A total of 84 piglets were included. For assessing blood depletion, R 2 was 0.62 ± 0.08, 0.61 ± 0.10, 0.60 ± 0.10, 0.60 ± 0.08, 0.58 ± 0.11, 0.57 ± 0.13, and 0.54 ± 0.10 for Support Vector Machine, XGBoost, Multicorrelation Models, Random Forest, K-Nearest Neighbors, Ensemble, and Multilayer Perceptron, respectively. The best-performing model (SVM) achieved an RMSE of 240 ± 27 mL and MAE of 186 ± 17 mL. Area under curve for detecting abundant blood loss (> 500 mL, equivalent to 15 mL kg −1 ) ranged from 0.856 to 0.895 across all classification models. Conclusion In this model of controlled hemorrhagic shock in piglets, ML seems to offer promising accuracy to quantify blood depletion using hemodynamic and echographic measures.