Developing Robust Chronic Critical Illness Risk Models

点击翻译按钮获取中文摘要

 DESCRIPTION (provided by applicant): Chronic critical illness (CCI) leads to extended stays in intensive care units, reduces quality of life, adds nearly $20 billion annually in health-relate costs, and is a precursor to other conditions such as persistent inflammation, immunosuppression, and catabolism syndrome (PICS). At any one time, more than 100,000 patients suffer from CCI in the United States alone. CCI for surgical trauma patients is defined as an intensive care unit stay greater than or equal to 14 days with evidence of ongoing organ dysfunction. The availability of existing clinical data characterizing CCI provides the opportunity to apply advanced statistical methods to develop robust patient-level CCI risk models for trauma and sepsis research. More effective risk models are invaluable to practitioners, administrators, and policy makers, and lead to better decisions resulting in increased patient quality of life and reduced health care costs. Despite widespread use of patient-level prediction models for clinical events (e.g., mortality to compute severity of illness scores) patient-level CCI risk models are not currently available. Moreover, recent advances in statistics that can be applied to robust risk modeling of underlying pathologies are not easily accessible to health care researchers. Thus, utilizing improved statistical methods for developing a CCI risk model that can reveal the etiology for the disease, not merely predict onset, would significantly help scientists understand, assess, prevent, and treat CCI. This Phase I study investigates the feasibility of applying a Best Approximating Model (BAM) method to develop improved risk models for CCI on a NIGMS-sponsored dataset for a population of severely injured blunt trauma patients. The BAM method is a systematic model development approach that combines robust estimation, specification analyses, stochastic/exhaustive model search, and model validation within the single model selection/validation framework of a generalized additive model. A BAM is designed to handle common problems encountered in developing predictive risk models including possible model misspecification, missing values, and overfitting; as well as multicollinearity, small sample size bias, and Type I error inflation due to multiple model comparisons. In this Phase I study, a robust CCI risk model will be developed using an in-house BAM method, followed by a series of simulation studies designed to evaluate its performance. The simulation studies will also characterize the advantages of the BAM strategy for developing a robust CCI risk model over conventional statistical methods such as stepwise regression. Feasibility study results will provide the preliminary research needed for more advanced Phase II CCI risk model development, evaluation, and dissemination that, in turn, will establish the essential foundation for Phase III commercialization of an advanced prognostic technology.