Machine learning methods to improve bedside fluid responsiveness prediction in severe sepsis or septic shock: an observational study.

Abstract

Passive leg raising (PLR) predicts fluid responsiveness in critical illness, although restrictions in mobilising patients often preclude this haemodynamic challenge being used. We investigated whether machine learning applied on transthoracic echocardiography (TTE) data might be used as a tool for predicting fluid responsiveness in critically ill patients. We studied, 100 critically ill patients (mean age: 62 yr [standard deviation: 14]) with severe sepsis or septic shock prospectively over 24 months. Transthoracic echocardiography measurements were performed at baseline, after PLR, and before and after a standardised fluid challenge in learning and test populations (n=50 patients each). A 15% increase in stroke volume defined fluid responsiveness. The machine learning methods used were classification and regression tree (CART), partial least-squares regression (PLS), neural network (NNET), and linear discriminant analysis (LDA). Each method was applied offline to determine whether fluid responsiveness may be predicted from left and right cardiac ventricular physiological changes detected by cardiac ultrasound. Predictive values for fluid responsiveness were compared by receiver operating characteristics (area under the curve [AUC]; mean [95% confidence intervals]). In the learning sample, the AUC values were PLR 0.76 (0.62-0.89), CART 0.83 (0.73-0.94), PLS 0.97 (0.93-1), NNET 0.93 (0.85-1), and LDA 0.90 (0.81-0.98). In the test sample, the AUC values were PLR 0.77 (0.64-0.91), CART 0.68 (0.54-0.81), PLS 0.83 (0.71-0.96), NNET 0.83 (0.71-0.94), and LDA 0.85 (0.74-0.96) respectively. The PLS model identified inferior vena cava collapsibility, velocity-time integral, S-wave, E/Ea ratio, and E-wave as key echocardiographic parameters. Machine learning generated several models for predicting fluid responsiveness that were comparable with the haemodynamic response to PLR.