RESUMEN
Objective:To improve the accuracy of prediction of macrosomia by application of machine learning.Methods:Ultrasound measurement data and fetal birth weight of macrosomia and normal birth weight neonates were collected during January 2015 to December 2016 in Mindong Hospital Affiliated to Fujian Medical University.Ultrasound built-in Hadlock formula,multiple linear regression,k-nearest neighbor,support vector machine,random forest were evaluated and compared to predict macrosomia using actual fetal birth weight as the gold standard.Results:The sensitivity of built-in Hadlock formula to predict macrosomia was 40.86％ and Youden index was 39.95％.The sensitivity of the multivariate linear regression was 60.22％ and the Youden index was 58.85％.The sensitivity of the k-nearest neighbor was 86.21％ and the Youden index was 75.10％.The sensitivity of the support vector machine was 86.21％ and the Youden index was 73.51％.The sensitivity of the random forest was 81.03％ and the Youden index was 71.51％.The Youden index of multivariate linear regression was significantly bigger than that of built-in Hadlock(u =3.64,P ＜0.001).The Youden index of k-nearest neighbor,support vector machine and random forest was significantly bigger and built-in Hadlock and multivariate linear regression (P＜0.001,P＜ 0.05).Conclusions:The machine learning is of high accuracy and great value of application.