Severely injured children achieve the best outcomes when treated at centers that provide specialized pediatric trauma care. Assessing the need for high-level trauma care is a complex classification problem that is affected by a very large number of potentially interacting factors (high-dimensional data), including age, mechanism of injury, known or suspected injuries and the physiological responses to injury. Despite this known complexity, approaches to pediatric trauma triage have been based on expert-derived rules, partly because of challenge of data acquisition in a prehospital setting. New approaches to data acquisition, however, are being rapidly introduced that allow access to increasing amounts of data at the injury scene and during transport, replacing the challenge of data capture with that of managing large numbers of explanatory variables. The overall goal of this project is to develop a triage system that increases the likelihood that injured children are treated at hospitals with the capability of optimizing outcome after injury. The purpose of this proposal is to develop more accurate methods for predicting the outcome and resource needs of injured children based on data available in prehospital and emergency department settings. We hypothesize that the relationship between observable prehospital and early hospital features (patient characteristics, physiologic status, anatomic sites of injury, mechanism of injury and prehospital treatments) and the need for and level of care required for injured children is highly complex, requiring approaches for modeling high-dimensional data to achieve accurate prediction. This hypothesis will be tested in two aims: 1. compare the impact of low- and high-dimensional data on the performance of models predicting time-dependent outcomes and resource utilization after pediatric injury;2. build high-dimensional multivariate probability models that predict outcomes after pediatric injury using data from individual injury datasets and integrated data from heterogeneous injury datasets. The hypothesis to be tested under Aim 1 is that prediction of time-dependent outcomes and resource utilization after pediatric injury will be improved by modeling high-dimensional data.
Aim 1 will be pursued using data obtained from two national trauma databases to develop and compare models based on low- and high-dimensional data.
This aim will require extending our innovative approach to high-dimensional regression analysis to handle time- dependent response variables and competing risks. The hypothesis to be tested under Aim 2 is that prediction of outcomes after pediatric injury will be improved using integrated data obtained from heterogeneous injury datasets.
Aim 2 will be pursued using a motor vehicle crash dataset and a trauma database to develop multivariate probability models based on data from each dataset and integrated data from both datasets.
This aim will require developing novel approaches for building Bayesian graphical models from distributed high- dimensional data. This proposal will bridge gaps in our understanding of the impact of domain complexity on the accuracy of prediction in prehospital and emergency department settings.
Severely injured children achieve the best outcomes when treated at hospitals that provide specialized pediatric trauma care. Determining the need for high-level pediatric trauma care is a complex classification problem that is influenced by a very large number of potentially interacting factors, including age, mechanism of injury, known or suspected injuries and the physiological responses to injury. In this proposal, novel statistical approaches that account for this complexity will be developed for more accurately predicting the need for high-level pediatric trauma care among injured children.
|Beck, Haley E; Mittal, Sushil; Madigan, David et al. (2015) Reassessing mechanism as a predictor of pediatric injury mortality. J Surg Res 199:641-6|
|Mittal, Sushil; Madigan, David; Burd, Randall S et al. (2014) High-dimensional, massive sample-size Cox proportional hazards regression for survival analysis. Biostatistics 15:207-21|
|Mittal, Sushil; Madigan, David; Cheng, Jerry Q et al. (2013) Large-scale parametric survival analysis. Stat Med :|
|Simpson, Shawn E; Madigan, David; Zorych, Ivan et al. (2013) Multiple self-controlled case series for large-scale longitudinal observational databases. Biometrics 69:893-902|