Fetal weight estimation for prediction of fetal macrosomia: does additional clinical and demographic data using pattern recognition algorithm improve detection?

OBJECTIVE

The aim of this study was to test whether pattern recognition classifiers with multiple clinical and sonographic variables could improve ultrasound prediction of fetal macrosomia over prediction which relies on the commonly used formulas for the sonographic estimation of fetal weight.

METHODS

THE SVM ALGORITHM WAS USED FOR BINARY CLASSIFICATION BETWEEN TWO CATEGORIES OF WEIGHT ESTIMATION: >4000gr and <4000gr. Clinical and sononographic input variables of 100 pregnancies suspected of having LGA fetuses were tested.

RESULTS

Thirteen out of 38 features were selected as contributing variables that distinguish birth weights of below 4000gr and of 4000gr and above. Considering 4000gr. as a cutoff weight the pattern recognition algorithm predicted macrosomia with a sensitivity of 81%, specificity of 73%, positive predictive value of 81% and negative predictive value of 73%. The comparative figures according to the combined criteria based on two commonly used formulas generated from regression analysis were 88.1%, 34%, 65.8%, 66.7%.

CONCLUSIONS

The SVM algorithm provides a comparable prediction of LGA fetuses as other commonly used formulas generated from regression analysis. The better specificity and better positive predictive value suggest potential value for this method and further accumulation of data may improve the reliability of this approach.