LAUSR

Accurate fetal weight estimation is important for both fetuses and their mothers. The low birth weight (LBW, birth weight < 2500 g) and high birth weight (HBW, birth weight ≥ 4000 g) fetuses and their mothers are linked to both short and long-term health outcomes such as high perinatal mortality rate, various complications, and chronic disease in life. Because of the imbalanced small data sets and body size heterogeneities between different fetal weight groups, it is difficult for the commonly used regression formulas to get a satisfying performance, especially for the HBW and LBW fetuses. The aim of this paper is to propose a machine learning solution to improve fetal weight estimation accuracy and to help the clinicians identify potential risks before delivery. A clinical data set of 7875 singleton fetuses were analyzed. The synthetic minority over-sampling technique (SMOTE) was employed to solve the imbalanced learning problem. Then, the support vector machine (SVM) algorithm was utilized for fetal weight classification. Finally, the deep belief network (DBN) was employed to estimate the fetal weight based on different ultrasound parameters. The estimation result of the proposed model showed a mean absolute percent error (MAPE) of 6.09 ± 5.06% and mean absolute error (MAE) of 198.55±158.63g. It demonstrated that our model outperformed the commonly-used regression formulas, especially for the HBW and LBW fetuses.