Intensive Care

Three members of the organizing committee met in March 2012 to define the methodology, to select the subtopics for study, and to identify the experts in the field. Experts were invited based on their record of publications in PARDS in the past 5 years and their participation in clinical research studies in pediatric critical care. The final list of 27 experts, representing 21 academic institutions and eight countries, constituted the PALICC expert group (Appendix 1). Of note, only one expert declined to participate due to personal reasons; two experts initially agreed to participate but were subsequently unable for personal reasons.
The first PALICC meeting took place in Chicago, IL, on October 2, 2012, in conjunction with the fall meeting of the PALISI Network. At this meeting, we discussed and agreed upon conference subtopics, the project timeline, and the consensus methodology (Fig. 1). Experts were also assigned to each of the nine subtopics. The modified Delphi approach previously employed by the French Society of Pediatric Intensive Care (13 ) was chosen as the methodology to achieve consensus. This approach was necessary because of the limited data and low level of available evidence, as well as the high variability in clinical practice in PARDS. A detailed description of this methodology is available in the supplement published in Pediatric Critical Care Medicine (14 (link)).
Between the first and second meeting, each group of experts undertook a comprehensive, standardized literature review. Upon completion, each group drafted their recommendations along with detailed arguments to support them. The second meeting occurred in Montreal, QC, Canada, on April 18–19, 2013. At this 2-day meeting, the recommendations were discussed and the wording of each agreed upon by the majority of experts. Possible omissions for any of the nine topics were also discussed. After the second meeting, recommendations with their respective arguments (long texts) were distributed to each expert for electronic scoring by the Research ANd Development/University of California Los Angeles (RAND/UCLA) appropriateness method (15 ). Experts with a disclosed conflict of interest were excluded from voting on areas where any real or perceived conflict was identified. After the initial scoring, all recommendations were consolidated by the organizing committee.
Agreement was determined by voting using the RAND/UCLA scale (scores range from 1 to 9), with each expert having an equal vote but with the highest and lowest scores discarded after each vote. “Strong” agreement required that all experts rank the recommendation 7 or higher. “Weak” agreement meant that at least one more expert ranked the recommendation below 7, but the median vote was at least 7. Those with strong agreement were considered complete, and those with weak agreement were revised based on comments by the experts. These revised recommendations were then distributed for a second round of electronic voting. After this voting, some reworded recommendations obtained a strong agreement. For the remaining recommendations with a weak agreement after the second round, the percentage of experts who rated 7 or above was calculated and is reported after each weak recommendation. With this method of calculation, a strong agreement corresponded to a percentage of agreement more than 95% (no more than one expert rated below 7 on the RAND/UCLA scale).
The third and final meeting took place on October 9, 2013, in Paris, France. Each group presented their final recommendations, and a third round of voting was conducted for several specific but unresolved recommendations related to the definitions. The organizers believed it was vital to achieve strong agreement regarding definitions, and this was accomplished after much dialog and debate. Additionally, each group of experts presented their consensus regarding key areas of controversy and future research.

This investigation was performed in the CPCCRN of the Eunice Kennedy Shriver National Institute of Child Health and Human Development (8 (link)). Detailed methods for the TOPICC data collection have been previously described (6 (link)). There were seven sites, and one was composed of two institutions. In brief, patients from newborn to less than 18 years were randomly selected and stratified by hospital from December 4, 2011, to April 7, 2013. Patients from both general/medical and cardiac/cardiovascular PICUs were included. Moribund patients (vital signs incompatible with life for the first 2 hr after PICU admission) were excluded. Only the first PICU admission during hospitalization was included. The protocol was approved by all participating institutional review boards. Other analyses using this database have been published (6 (link), 7 (link), 9 (link), 10 (link)).
Data included descriptive and demographic information (Table 1). Interventions included both surgery and interventional catheterization. Cardiac arrest included closed chest massage within 24 hours before hospitalization or after hospital admission but before PICU admission. Admission source was classified as emergency department, inpatient unit, postintervention unit, or admission from another institution. Diagnosis was classified by the system of primary dysfunction based on the reason for PICU admission; cardiovascular conditions were classified as congenital or acquired.
The primary outcome in this analysis was hospital survival versus death.
Physiologic status was measured using the PRISM physiologic variables (5 (link)) with a shortened time interval (2 hr before PICU admission to 4 hr after admission for laboratory data and the first 4 hr of PICU care for other physiologic variables). For model building, the PRISM components were separated into cardiovascular (heart rate, systolic blood pressure, and temperature), neurologic (pupillary reactivity and mental status), respiratory (arterial P_o2, pH, P_co2, and total bicarbonate), chemical (glucose, potassium, blood urea nitrogen, and creatinine), and hematologic (WBC count, platelet count, prothrombin, and partial thromboplastin time) components, and the total PRISM was also separated into neurologic and non-neurologic categories.
The time interval for assessing PRISM data was modified for cardiac patients under 91 days old because some institutions admit infants to the PICU before a cardiac intervention to “optimize” the clinical status but not for intensive care; in these cases, the postintervention period more accurately reflects intensive care. However, in other infants for whom the cardiac intervention is delayed after PICU admission or the intervention is a therapy required because of failed medical management of the acute condition, the routine PRISM data collection time interval is an appropriate reflection of critical illness. Therefore, we identified infants for whom it would be more appropriate to use data from the 4 hours after the cardiac intervention (postintervention time interval) and those for whom using the admission time interval was more appropriate. We operationalized this decision on the conditions likely to present within the first 90 days, the time period when the vast majority of these conditions present (Table 2).
Statistical analyses used SAS 9.4 (SAS Institute Inc., Cary, NC) for descriptive statistics, model development, and fit assessment and R 3.1.1 (The R Foundation for Statistical Computing, Vienna, Austria; http://www.wu.ac.at/statmath) for evaluation of predictive ability. Patient characteristics were descriptively compared and evaluated across sites using the Kruskal-Wallis test for continuous variables and the Pearson chi-square test for categorical variables. The statistical analysis was under the direction of R.H.
The dataset was randomly divided into a derivation set (75%) for model building and a validation set (25%) stratified by the study site. Univariate mortality odds ratios were computed, and variables with a significance level of less than 0.1 were considered candidate predictors for the final model. As was the case for the previously published trichotomous (death, survival with significant new morbidity, and intact survival) model construction, a nonautomated (examined by biostatistician and clinician at each step) backward stepwise selection approach was used to select factors. Multicategorical factors (e.g., diagnostic categories) had factors combined when appropriate per statistical and clinical criteria. Clinician input was included (and paramount) in this process to ensure that the model fit was relevant and consistent with clinical information. Construction of a clinically relevant, sufficiently predictive model using predictors readily available to the clinician took precedence over inclusion based solely on statistical significance. We were cognizant of the existing trichotomous outcome model and attempted, when statistically justified, to create a compatible two-outcome model that could aid in a smooth transition to using the three-outcome approach.
Final candidate models were evaluated based on 2D receiver operating characteristic (ROC) curves (discrimination) and the Hosmer-Lemeshow goodness of fit (calibration). For the entire dataset, goodness of fit with respect to key subgroups was assessed by examining SMRs for descriptive and diagnostic categories not used in the final model. Only categories with at least 10 outcomes in observed and expected cells were used.