Performance criteria based on true and false classification and clinical outcomes. Influence of analytical performance on diagnostic outcome using a single clinical component

Abstract

Background: In the general classical model for diagnoses based on a single analytic component, distributions of healthy and diseased are compared and several investigations of varying analytical performance on the percentage of misclassifications have been published. A new concept based on an alternative type of diagnosing, based on sharp decision limits has been introduced in diagnostic guidelines, but only a few publications on investigation of analytical performance have been seen. Methods: The two diagnostic models (bimodal and unimodal) based on natural logarithmic Gaussian distributions are simulated. Results: In the bimodal model it is possible to evaluate the influence of prevalence of disease in combination with varying analytical performances. In the unimodal model the prevalence is pre-decided by the chosen decision limit. In this model the influence of analytical performance is investigated for diagnosing diabetes using haemoglobin A1c (HbA1c), and for patients with high and low risk for coronary heart disease defined by serum-cholesterol concentrations. Conclusions: For HbA1c the guidelines and recommendations define a maximum inter-laboratory coefficient of variation of 3.5%, but this is in DCCT units (without a true zero-point), so after transformation to IFCC units (which are proportional) it was 5.2%, which allows for analytical bias as high as approximately ±9%. Consequently, analytical quality specifications should be separated as maximum bias and imprecision.