Non-Parametric Combined Reference Regions and Prediction of Clinical Risk

Abstract

Background: Many clinical decisions depend on estimating patient risk of clinical outcomes by interpreting test results relative to reference intervals, but standard application of reference intervals suffers from two major limitations that reduce the accuracy of clinical decisions: (1) each test result is assessed separately relative to a univariate reference interval, ignoring the rich pathophysiologic information in multivariate relationships, and (2) reference intervals are intended to reflect a population's biological characteristics and are not calibrated for outcome prediction.

Methods: We developed a combined reference region (CRR), derived CRRs for some pairs of complete blood count (CBC) indices (RBC, MCH, RDW, WBC, PLT), and assessed whether the CRR could enhance the univariate reference interval's prediction of a general clinical outcome, 5-year mortality risk (MR).

Results: The CRR significantly improved MR estimation for 21/21 patient subsets defined by current univariate reference intervals. The CRR identified individuals with >2-fold increase in MR in many cases and uniformly improved the accuracy for all five pairs of tests considered. Overall, the 95% CRR identified individuals with a >7× increase in 5-year MR.

Conclusions: The CRR enhances the accuracy of the prediction of 5-year MR relative to current univariate reference intervals. The CRR generalizes to higher numbers of tests or biomarkers, as well as to clinical outcomes more specific than MR, and may provide a general way to use existing data to enhance the accuracy and precision of clinical decisions.

Keywords: Big Data; Complete Blood Count; Laboratory Medicine; Reference Intervals; Statistics.