BACKGROUND:
One of the most important tasks that cardiovascular clinicians perform is risk stratification, as that enables appropriate targeting of aggressive treatments to patients that are most likely to benefit from them. Contemporary risk stratification strategies include clinical scoring systems along with performance of noninvasive tests. Although these approaches are commonly used, clinicians still find themselves needing to incorporate multiple pieces of clinical information into a cohesive global risk assessment. The concept of utilizing data from large observational data sets to develop complex risk scores and to encourage their use in routine practice is therefore gradually evolving and gaining acceptance. The Logical Analysis of Data (LAD) is a potentially useful approach for systematically analyzing large databases for the purpose of developing and validating clinically useful risk prediction schemes. Unlike standard regression techniques, LAD does not primarily focus on individual risk factors and two-way interactions between them. Rather, LAD is designed to identify complex patterns of findings, or syndromes, that predict outcomes. This method has been applied to problems in economics, seismology and oil exploration, but not to medicine.
DESIGN NARRATIVE:
The study has three specific aims: 1). to apply LAD to develop and validate risk prediction instruments among patients undergoing different types of cardiac surgery. 2. to compare the predictive value of LAD predictive instruments with predictive instruments developed using standard statistical methods, including multiple time-phase parametric modeling. 3. to develop predictive instruments using relative risk forests, a new Monte Carlo method for estimating risk values in large survival data settings with large numbers of correlated variables. Relative risk forests are an adaptation of random forests introduced by Breiman. When possible these methods will be compared to LAD. Internal estimates for the generalization error, a measure of how well the method will generalize to other data settings, will be computed and will be used in the development of the predictive instrument. Relative risk forests will also be compared to several other non-deterministic methods, including boosting and spike and slab variable selection. All of these techniques can be used to develop complex models while maintaining good prediction error and are ideal for high dimensional problems where traditional methods breakdown. Although this project will focus on risk assessment among patients undergoing cardiac surgery, it is important to recognize that we are primarily interested in the value of LAD as a means of analyzing very large and complex data sets within a medical sphere. Hence, the applicability of this work goes beyond determination of risk of patients undergoing cardiac surgery.
Data used for this study will consist of cardiac surgery data from the Cleveland Clinic Foundation Cardiovascular Information Registry (CVIR). Four cohorts of data will be assembled; Cohort I: 18,914 CABG patients between 1990 and 2000; Cohort II: 6952 patients undergoing aortic valve replacement; Cohort III: 2979 patients undergoing mitral valve replacement; Cohort IV: 10,482 patients undergoing mitral valve repair. The primary endpoint will be long term total mortality; for valve surgery patients it will be active follow-up.
