Evaluating obstetric risk scores by receiver operating characteristic curves

Abstract

The purpose of obstetric risk scoring systems is to identify pregnancies that will end in poor perinatal outcomes such as perinatal death. The proportion of such pregnancies identified by a risk scoring system can be increased by choosing a more lenient threshold criterion for the score beyond which the pregnancy is regarded as being at risk. This increase in sensitivity is gained at the expense of a decrease in specificity. Thresholds differ between published risk scoring systems, making it difficult to assess their relative merits using the measures of sensitivity and specificity. However, risk scores can be compared by constructing receiver operating characteristic curves, i.e., plotting the true positive rate against the false positive rate for multiple thresholds and calculating the area under the receiver operating characteristic curve (theta). In published systems used in this paper, values of theta, used to rank published obstetric risk scoring systems, ranged from 0.49-0.95. All scores used prepregnancy and early antenatal information. The risk scores with the highest predictive accuracy tended to include factors representing past reproductive experience and those from the late antenatal and intrapartum periods. Studies with statistical weighting of risk factors tended to use fewer late prenatal and intrapartum risk factors.