1039: Machine learning approach to fetal weight estimation

Abstract

Fetal Weight Estimation (FWE) is important for ensuring the safety of both the fetus and mother during pregnancy and delivery. Current estimation methods are typically based on manual ultrasound (US) measurements. These are obtained during various stages of pregnancy, binned into quantiles, and compared to reference growth curves. It is well established that this method is inaccurate and there is a strong debate in the Maternal Fetal Medicine community regarding what should be the right approach. We propose a novel approach of fetal weight estimation, which makes use all the multidimensional data rather than a predetermined quantile, in order to create a standardized model. The study was performed in a large referral center, from January 2011 to December 2016. In every parturient presented to the delivery room, an US estimation of fetal weight was performed by a trained physician. We have studied 6972 records. These measurements were linked with postpartum data. The samples included the following information: admission demographic data, gestational age, US measurements (BPD, FL, AC, HC), newborn sex. Since we used 'real life data', exclusion criteria included examinations with missing data or illogical values. After normalizing the data, we applied supervised learning algorithms such as Decision Tree Regressor, Support Vector Regressor, Ada Boost Regressor and Linear Regressor to create a new FWE formula. We then applied the K-Means algorithm on the data to create clusters based on the measurements and created five different estimators for the different clusters. We used a training dataset with 4880 entries, a 10-fold cross validation set and a test dataset with 2092 entries. To evaluate the quality of the models, we examined the mean absolute error (MAE) and with respect to the actual birthweight (in grams) and compared them with three Hadlock formulas. We also calculated the R2 value for each model. When using machine learning algorithms on the entire dataset to create one equation for FWE, the clustered estimator achieved the best results. The results of the analysis are summarized in Figure 1 and Table 1. Figure 1 - Comparison of Hadlock formulas to machine learning algorithms in predicting birthweight (Train 4880, Test 2092 samples). Table 1 - R2 values for the different models: A machine learning algorithm approach to FWE creates models that reach a higher level of accuracy compared to the currently used methods such as Hadlock 1-3.View Large Image Figure ViewerDownload Hi-res image Download (PPT)