A new algorithm for improving fetal weight estimation from ultrasound data at term

Abstract

Objective: The purpose of this retrospective study was to find a method of improving the accuracy of fetal birth weight estimation on the basis of traditional ultrasonographic measurements of the head, thorax, and femur at term. In this context, we analyzed a novel regression method comparing to existing algorithms.

Methods: The delivery records of two hospitals were searched for women who delivered macrosomic infants, and the patients' medical records were retrospectively reviewed in order to derive clinical and ultrasonographic data at term. A total of 223 patients with macrosomic infants (birth weight > 4,000 g) were identified. These patients were complemented by data for 212 women who had ultrasound fetal assessments of less than 4,000 g. We used the method of isotonic regression to construct a birth weight prediction function that increases monotonically with each of the input variables and which minimizes the empirical quadratic loss.

Results: A suspicion of macrosomia was based on a history of macrosomia, fundal height, and sonographic weight estimation >4,000 g. The mean period between ultrasound weight estimation and delivery was 7.2 days. The ability of the biometric algorithms developed to predict fetal weight at term ranged between a mean absolute error of 312 and 344 g, given a confidence interval of 95%. We demonstrate that predictions of birth weight on the basis of ultrasound data can be improved significantly, if an isotonic regression model is used instead of a linear regression model.

Conclusions: This study demonstrates that ultrasound detection of macrosomia can be improved using the isotonic regression method.

Fig. 1

Distribution of the regression residuals for the six regression methods shown in Fig. 1, using untransformed data

Fig. 2

Six regression methods and their performance as predictors of macrosomia. The receiver operating characteristics (ROC) of all of the methods appear very similar—i.e., all of the methods perform equally well