Generalization of the receiver-operating characteristic curve to determine the normal hemoglobin range cutoff points...

Identification of a normal range for biomarkers, based on pregnancy outcomes (caused by their high or low values) is of special importance in clinical studies. As some pregnancy outcomes can happen in both high and low levels of biomarkers, the receiver-operating characteristic (ROC) curve is unsuitable for identifying these levels separately; rather, a statistical method is preferable which identifies both levels simultaneously.

To this effect, our research introduces a generalization of ROC curve (by using a number of related consequences) to identify a normal range for the biomarker. Practically, the study intends to identify a normal range of hemoglobin in the first trimester of pregnancy to prevent adverse outcomes that can be caused by high and low levels of hemoglobin.

The current article introduces an ROC generalization curve to determine a normal range for biomarkers based on a number of pregnancy outcomes, which may occur in high and low levels of biomarkers. Simulated data were also used to compare the current method with the ROC curve method. Our data collected from a cohort study carried out on 600 pregnant women referring to Milad Hospital in Tehran, Iran in 2010. The data comprised hemoglobin level in the first trimester of pregnancy as well as pregnancy outcomes such as preterm delivery, low birth weight, preeclampsia, and gestational diabetes. We calculated an estimation of the normal range of hemoglobin for the study population. Statistical analysis was carried out by R software, version 3.0.2.

Results from the simulation study indicated that, the new method was better than the methods which used two ROC curves separately with regard to sensitivity and specificity. In this method, the level of normal hemoglobin in the first trimester ranged from 10 to 12.4 with sensitivity and specificity levels of 76.2% and 48% respectively, which is higher than previous studies.

With regard to the normal range of biomarkers, our method yielded greater sensitivity and specificity levels than methods using the ROC curve, which separately analyzes the data, particularly in occasions with common consequences in high and low levels of the biomarker.