Characterizing drivers of change in intraoperative cerebral saturation using supervised machine learning.

Abstract

Regional cerebral oxygen saturation (rSO₂) is used to monitor cerebral perfusion with emerging evidence that optimization of rSO₂ may improve neurological and non-neurological outcomes. To manipulate rSO₂ an understanding of the variables that drive its behavior is necessary, and this can be accomplished using supervised machine learning. This study aimed to establish a hierarchy by which various hemodynamic and ventilatory variables contribute to intraoperative changes in rSO₂. A post-hoc analysis 146 patients undergoing high risk surgery. rSO₂ was partitioned into segments with a change of at least 3% points over 5 min. Features from hemodynamic and ventilatory variables were used to train a machine learning classification algorithm (XGBoost) for prediction of association with either up or down-sloping rSO₂. The classifier was optimized and validated using five-fold cross validation. Feature importance was quantified based on information gain and permutation feature importance. The optimized classifier demonstrated a mean accuracy of 77.1% (SD 8.0%) and a mean area-under-ROC-curve of 0.86 (SD 0.06). The most important features based on information gain were the slope of the associated ETCO₂ signal, the slope of the SPO₂ signal, and the mean of the MAP signal. CO₂ is a significant mediator of changes in rSO₂ in an intraoperative setting, through its established effects on cerebral blood flow. This study furthers our overall understanding of the complex physiologic process that governs cerebral oxygenation by quantifying the hierarchy by which rSO₂ is affected. Clinical Trial Number NCT01838733 (ClinicalTrials.gov).