A Model Of Estimating Fetal Weight Based On Artificial Neural Network

?Objective?The estimated fetal weight(EFW)before birth plays an important role in helping obstetricians to make management decisions and determine whether caesarean section is necessary.Improving the accuracy of EFW can also decrease the ratio of perinatal complications and perinatal mortality.Quite a few prediction approaches have been developed,but none of them can achieve the accuracy so that clinical practice can employ.Currently,the most widely used estimating fetal weight method is an approach based on fetal sonography data.However,these EFW formulae are based on the data from western countries(mostly Caucasian population).?Methods?In recent years,artificial neural network has attained significant prediction power in many fields,and it also have been used to estimate fetal weight,but none have been applied in the clinic.In this study,we proposed a more accurate method that used an artificial neural network model with heterogenous big data of pregnancy examinations,physiological parameters and ultrasound parameters.We use some feature selection methods to select features and make the model more concise to apply in clinic.?Results?The results demonstrated that the neural network model can decrease both the root-mean-square error(RMSE)and mean absolute percentage error(MAPE)of EFW.The RMSE of the neural network model is 244.88 g,which is 10.4% better than the best traditional model.The overall estimation accuracy rate is increased from64.95%(the traditional method)to 70.43%,a 5.48% incrase.Due to the high correlation between the weight of current and previous neonates from a same mother,a neural network model was trained by the multipara samples separately.In addition,comparing the frequency distribution of the estimated weight absolute error between the neural network model and the traditional formula have shown that there is much less serious error(> 500g)appeared in the EFW of the neural network model,reducing the false-positive and false negative rates to a large extend.The neural network model did not follow the traditional approach in predicting macrosomia.The area under the curve(AUC)of the relative operating characteristic(ROC)curve was0.917,comparing to the existing best 0.894.?Conclusion?Our model used a larger sample size and relatively fewer parameters to make the model easier to be applied in the clinic.Further research can continue to train the model by recruiting more data for clinical external verification.In conclusion,we believe that artificial neural network,as a mature method for establishing prediction models,can be widely used in not only estimated fetal weight but also other clinical research in obstetrics.