Specific Aims / Hypothesis: Pediatric severe sepsis, an overwhelming systemic inflammatory response to infection, is a leading cause of mortality in pediatric intensive care units (PICUs) worldwide. Despite decades of sepsis research, the main therapies available for the management of sepsis essentially remain goal-directed fluid resuscitation, timely antibiotic administration, and supportive care for organ dysfunction. Additional adjunctive therapies including blood product transfusions, plasmapheresis, and extracorporeal membranous oxygenation (ECMO)are often administered empirically in patients with sepsis and multiple organ dysfunction syndrome (MODS). While these are important adjuncts in the management of critically ill patients, the efficacy of such therapies and the conditions under which their use may be warranted remain unknown and controversial. 

Multi-institution randomized trials require large numbers of patients, have high costs and take long periods of time. Retrospective analysis of multi-institution databases of pediatric critical care patients offer a valuable tool to assist researchers in describing resource utilization and comparing intervention effectiveness across the population. However a critical gap exists in the lack of a comprehensive multi-institution dataset which: (a) accurately identifies patients with severe sepsis and septic shock, (b) includes comprehensive measurements and results sufficient to stratify patients and detect MODS, and (c) provides longitudinal resource utilization data. 

To address this problem, we will join two large, multi-center databases of ICU care and leverage the advantages of both datasets to address the above knowledge gap. Linkage will be accomplished using probabilistic algorithms which match similarly to human pattern recognition and do not rely on common identifiers. The objectives of this study are to measure variations in resource utilization across centers, determine the association of adjuvant therapies with mortality, and detect the presence of MODS in children with severe sepsis or septic shock through the creation of a linked multicenter dataset. Our overall hypothesis is that substantial variations in resource utilization exist in the management of severe sepsis and septic shock after adjusting for patient-level covariates. Identification of these variations will allow for novel investigations of the comparative effectiveness of such therapies. Bringing together informaticians and critical care epidemiology researchers in a new collaboration, we will complete these specific aims:

Aim 1 will create a combined, multicenter dataset through linkage of the VPS and PEDSnet datasets and describe the characteristics of this dataset with respect to patients, diagnoses and outcomes.This dataset will include six years of data at the eight academic pediatrics institutions in the PEDSnet dataset. We have created and applied this probabilistic linkage model in our single-center and validated the linkage accuracy matching by patient identifiers.We hypothesize that we will create a multicenter cohort of patients with an estimated cohort size of 3,000 PICU patients with severe sepsis and septic shock. 

Aim 2 will measure and compare resource utilization of key interventions and adjunctive therapies across centers in patients with severe sepsis and septic shock, including: vasoactive medications, invasive and non-invasive ventilation, inhaled nitric oxide, blood product transfusions, plasmapheresis, hemodialysis and ECMO. We will determine the associations of these interventions with outcomes of mortality and ICU-free days in models adjusted for age, sex and illness severity. We hypothesize that the use of these intensive therapies is associated with increased mortality in children with severe sepsis and septic shock, and their use will vary across centers. 

Exploratory Aim 3 will evaluate the capability to detect MODS in patients with severe sepsis and septic shock on day 7 and day 28 using standard diagnostic criteria. We will develop and validate a supervised classification algorithm to identify the presence of MODS in our multicenter cohort at day 7 and 28 after onset of severe sepsis and septic shock. We hypothesize that the richness of our combined dataset will detect the presence of MODS with high sensitivity and specificity using a manually reviewed single center dataset as a gold-standard. 

Summary: Completion of these aims will provide a powerful, multicenter dataset for critical care research, will demonstrate compelling preliminary data regarding the impact of unproven adjunctive therapies for sepsis on hospital mortality, and will establish automated identification of MODS in children with severe sepsis and septic shock. Our findings will provide generalizable epidemiology regarding experimental therapies in pediatric severe sepsis and septic shock which will inform the design of future multicenter intervention trials. Finally, the informatics approach utilized in this collaborative proposal will be adapted to other critically ill cohorts in the PICU, allowing for rapid, detailed assessment of the effectiveness of therapies administered in the setting of diverse critical illnesses using existing clinical datasets.