Gazing Into the Crystal Ball or Looking Through the Rear View Mirror? Prediction of Neurologic Outcome in Survivors of Pediatric Critical Illness*
Although a novel risk prediction model using data collected in the first 4 hours of PICU admission to predict morbidity and mortality has been described recently (5), from a clinical perspective, we know that new adverse events or procedures that occur during the PICU stay must have a further influence on morbidity. It is surprising, therefore, that there is currently no tool to predict adverse outcome taking into account patient factors at admission as well as what happens on the PICU. In this issue of Critical Care Medicine, Gupta et al (6) have taken the first step toward filling this knowledge gap. They report the development and validation of a tool to predict favorable neurologic outcome in children admitted to a PICU. Using advanced statistical modeling and a large dataset (Virtual Pediatric Systems [VPS]) comprised of more than 160,000 patients from 90 U.S. hospitals over a 7-year period, they tested the ability of 20 risk factors to predict neurologic outcome at the end of PICU stay. Unfavorable neurologic outcome was defined as a decline in the Pediatric Cerebral Performance Category (PCPC) (7) score from PICU admission to PICU discharge by 2 or more points, which, in this cohort, occurred in 1.04% of patients. Risk factors known at admission (weight, PIM2 score, head/nonhead trauma), as well as those that became apparent only during the PICU stay (cardiac arrest, stroke, use of mechanical ventilation, and prolonged ICU stay and mechanical ventilation), were independently associated with unfavorable neurologic outcome. Conversely, the presence of a chromosomal anomaly, cardiac surgery, and utilization of nitric oxide were independently associated with favorable outcome. The area under the receiver operator characteristic curve for the final model was 0.90, indicating that the tool had excellent discrimination to distinguish patients who had favorable outcome from those who did not. Model calibration was also good: the observed number of children with favorable outcome were similar to the expected number of children with favorable outcome when assessed in quintiles of risk.
The key strength of this analysis is that it has provided a new risk prediction model for an outcome that is clinically important and one that has not been included in many other risk prediction models in pediatric intensive care. Another advantage is that it could potentially be used at various time points during the patient’s PICU course to sequentially update their risk of unfavorable neurologic outcome, especially after adverse events such as a cardiac arrest have occurred. The large sample size, and the large diversity of hospitals represented in the dataset, provide some reassurance that the results can be generalized to other settings, at least in the United States where the PICUs were based. In addition to developing the model, the authors validated the model by splitting the data randomly into 10 groups and using a different 1/10th of the data each time on 10 different occasions for validation.
Yet, some fundamental questions remain unanswered.