Derivation of level-specific reference change values (RCV) from a health screening database and optimization of their thresholds based on clinical utility.

Abstract

Reference change values (RCV) are used for judging the significance of changes between any two measurements. Based on the within-individual CV (CVI), RCV is conventionally computed as the 95% confidence limit (CL) of the changes: RCV=1.96×2CVI. $RCV = 1.96 \times \sqrt 2 C{V_I}.$ However, the appropriateness of assuming a constant CVI and using the 95% CL for RCV remains controversial. The level-specific CVI and RCV were estimated for 20 screening tests using a database composed of results from 13,545 health-screening attendees over a 17-year period, after preliminary exclusion of individuals taking medications or having unusual changes in body mass index (BMI). A rational CL for RCV was explored in reference to a clinical score for the metabolic syndrome, sMS, which was derived based on a logistic regression model consisting of tests related to metabolic syndrome. The effect of adjusting CL for the RCV on diagnostic efficacies of detecting between-year change in sMS was evaluated. Test level dependency of CVI was apparent for some screening tests which have distributions with prominent skewing. The use of level-specific RCV was thus essential for them. The sensitivity for detecting a critical change in sMS based on the RCV set at 95%CL was extremely low in the majority of tests. However, by lowering CL stepwise from 95 to 75%, the sensitivity improved greatly without much change in specificity and positive predictive value. Loss-and-gain analysis showed that CL for RCV set around 80% gave the lowest loss, assuming a policy of reducing false negative judgment. Level specific CVI and RCV were necessary in tests with skewed distributions. RCV using 80%-90% CL is suitable in health screening for diseases that require early intervention for changes.