Introduction: Fluid boluses are commonly administered to children with suspected septic shock, but response is variable. Inability to predict response to fluid may lead to boluses being given without clinical improvement and ultimately fluid overload. Previous attempts to predict fluid response in children have frequently involved special devices and have not always included critically ill children. We hypothesize that a model using routinely collected variables in the pediatric intensive care unit (PICU) will be able to predict increase in mean arterial pressure (MAP) with fluid bolus in children with hypotension and suspected infection. Methods: This was a single-center retrospective cohort study of intravenous fluid boluses ≥200ml or 10ml/kg given to children < 18 years old with a MAP < first percentile for age within the first 72 hours of PICU admission between 2012-2022. Fluid responsiveness was defined as increase in MAP to ≥ fifth percentile for age by average 20 minutes post-bolus or next documented MAP. Logistic regression, ElasticNet, and XGBoosted Trees models were trained on 70% of the data and evaluated on the remaining 30%. Results: Of 655 boluses given to 278 patients, 233 (36%) were considered fluid responsive. Boluses were normal saline (50%), lactated ringers (29%), packed red blood cells (18%), and albumin (3%). Median bolus volume was 10.5ml/kg (IQR 9.8,19.3) and patient age 9.9 years (IQR 3.8,14.2). The best performing model was XGBoosted Trees—using nine nested decision trees with most frequently used features of: age, bolus volume per weight, weight, bicarbonate, platelets, ALC, pre-bolus MAP, hemoglobin, temperature, CO2, glucose, creatinine, ANC, chloride, and BUN. Bootstrapped AUROC was 0.77 (95%CI 0.72,0.81) and AUPRC 0.84 (95%CI 0.80,0.88), with test set sensitivity of 0.90, and positive predictive value of 0.77. 404 boluses (62%) were correctly predicted not to respond; 48 (7%) were predicted not to respond but did. Conclusions: A model using routinely collected data can reasonably predict fluid responsiveness for hypotensive children with suspected infection. Further research is needed to evaluate how high frequency physiologic and time-series data may change model performance and adapt such a model to personalize fluid resuscitation.
Walker et al. (Sun,) studied this question.