Time series forecasts and volatility measures as predictors of post-surgical death and kidney injury

Patients anesthetized during surgery can experience post-surgical adverse outcomes, such as kidney injury or death. In this study, we examine time series forecasts and volatility measures of perioperative physiologic data in an effort to predict these adverse outcomes. We build upon random forest models from a previous study and evaluate them based on their receiver operating characteristic (ROC) curves and their area under the curve (AUC) values. Additionally, we examine which additional variables are the most important to the predictive models. Our results indicate that volatility measures, especially those of blood oxygen saturation (SpO₂%), improve prediction of death. Pre-existing conditions were among the most important predictors for both outcomes.