| BackgroundMacrosomia refers to live born newborns with birth weight≥4 000 g,which is one of the common adverse outcomes of newborns,accounting for about 9%of all live births worldwide.In China,the prevalence rate of macrosomia from 2010 to 2014 was about 8.70%.The occurrence of macrosomia seriously affects the immediate and long-term health of mothers and newborns.It can increase the cesarean section,postpartum hemorrhage,shoulder dystocia and perinatal asphyxia,some risk of maternal and fetal complications,may also affect neonatal growth and development,and increase the adolescent obesity and overweight in adulthood the incidence of high blood pressure,diabetes.Therefore,it is of great public health significance to identify fetuses with rapid growth tendency before delivery.Currently,clinicians often use the Hadlock formula embedded in ultrasound to estimate fetal body mass.However,previous studies have shown low accuracy in ultrasound diagnosis of macrosomia(sensitivity:12%-75%,specificity:68%-99%).Therefore,it is helpful to reduce the incidence of macrosomia if the pregnant women who are prone to macrosomia are identified and early intervention is carried out according to some characteristics in early pregnancy.Machine learning is a multidisciplinary discipline,including probability theory,statistics,proximity theory,algorithm complexity and other disciplines.Algorithms can be used to guide a computer to use existing data to obtain a suitable model and use that model to make decisions about new situations.Commonly used machine learning methods include decision tree,support vector machine,random forest,neural engine network,etc.,among which support vector machine and random forest have been widely used in the medical field and have good stability for medical data fitting.Therefore,this study analyzed the growth and development characteristics of macrosomia in the fetal period,screened the risk factors related to macrosomia,and used Logistic regression,support vector machine and random forest method to build a prediction model of macrosomia in early pregnancy,providing a basis for the early identification of macrosomia and nutrition guidance during pregnancy.ObjectivesBased on the national health care big data center in the north of maternal database,evaluation of the existing macrosomia diagnosis model and analysis of macrosomia prenatal growth trajectory and its influencing factors,using the random forest model construction of macrosomia prediction model,improve the sensitivity of macrosomia prediction and specific degree to realize the early prediction of macrosomia in the early stages of pregnancy and the purpose of clinical intervention.Methods1.Data collectionIn the maternal database of The North Center for Big Data of National Health and Medical Care,the data of maternity examination and delivery of pregnant women in Jinan city from June 2017 to May 2018 were collected.The basic information of pregnant women included the age of pregnant women,pre-pregnancy body mass index,number of pregnancies,and birth rank.The prenatal information of pregnant women includes extra-pelvic measurements of pregnant women and b-ultrasound results of different gestational weeks;Delivery information for pregnant women included the last menstruation and delivery time,gestational age,newborn weight,and other hospitalized delivery outcomes.2.Statistical analysis(1)Mean ± standard deviation was used to describe the central tendency and discrete tendency of the measurement data conforming to normal distribution,and t test was used to compare the difference between the two groups,α=0.05.For classification data,composition ratio was used to describe the internal composition,and 2 test was used for inter-group comparison.(2)According to the predicted weight of the last prenatal B-ultrasound examination,the real birth weight was used as the gold standard for the diagnosis of macrosomia,and the sensitivity,specificity,and positive predictive value of the last prenatal B-ultrasound examination for the prediction of macrosomia were calculated by evaluating the authenticity of screening test.(3)Descriptive epidemiological analysis was used to compare the development of bultrasound indexes of macrosomia and normal-weight infants delivered in the study site during the study period and describe the development of the fetus.Single factor analysis and multifactor analysis were carried out on the risk factors affecting the occurrence of macrosomia,and the specific relationship between and was calculated at the same time.(4)Under sampling method is used to extract the same as the number of macrosomia,will match good research object according to the ratio of 7:3 were randomly divided into training set and test set,based on the basic information and neonatal mother pregnant 12 weeks of physical examination data,using random forest methods in training focus macrosomia prediction model was constructed,and is verified in the test set,The classification accuracy of random forest prediction method was calculated and the ROC curve was drawn.According to the characteristics of random forest model,the importance of influencing factors to predict the occurrence of macrosomia was ranked.(5)Will match good research object according to the ratio of 7:3 was randomly divided into training set and test set,based on the basic information and neonatal mother pregnant 12 weeks of physical examination data,using logistic regression and support vector machine(SVM)method in the training set to build macrosomia prediction model,and is verified in the test set,computing the two kinds of forecasting methods of classification accuracy,And draw the ROC curve.The control data not included in the model were randomly divided into nine parts,and the random forest model was cross validated.Results1.A total of 4260 neonates were included in the study to evaluate the predictive ability of B-ultrasonography for macrosomia,including 405(9.5%)macrosomia.The classification accuracy,sensitivity,specificity,positive predictive value and negative predictive value of macrosomia were 63.6%,29.6%,97.6%,57.7%and 93.1%respectively.2.A total of 4260 pregnant women were included in fetal growth trajectory analysis,among which macrosomia accounted for 9.5%;The results of fetal growth track analysis of macrosomia and normal infants showed that the biparietal diameter,femoral neck,head circumference and abdominal circumference of macrosomia were significantly larger than those of normal weight infants after 20 weeks of gestation,and the differences were statistically significant.Maternal age,pre-pregnancy BMI,number of pregnancies,and fetal rank were the influential factors of large size.A pre-pregnancy BMI of 26.5 to 32 was associated with a greater risk of having a baby.The risk of macrosomia decreased with a pre-pregnancy BMI greater than 32.3.405 macrosomia were selected as the case group,and 405 normal-weight infants were matched as the control group from 3855 normal-weight infants by under sampling method.The accuracy,sensitivity and specificity of random forest model for macrocephalus classification were 92.6%,88.4%,96.7%,96.4%and 89.4%respectively.By calculating the average decrease of Gini coefficient of each variable in the random forest model,the influencing factors for predicting the occurrence of macrosomia were obtained as follows:interspinous iliac diameter,transverse outlet diameter,interspinous iliac diameter,external sacral pubic diameter,prepregnancy BMI,age,parity and gestation number.4.The classification accuracy,sensitivity and specificity of the prediction model based on Logistic regression were 74.5%,68.6%,70.2%,70.8%positive predictive value and 37.9%negative predictive value.The classification accuracy,sensitivity and specificity of the prediction model based on support vector mechanism are 92.9%,82.6%and 95.9%respectively.The positive predictive value was 89.4%and the negative predictive value was 79.3%.5.The accuracy of the model verified for 9 times was 89.4%,92.2%,92.1%,95.0%,94.0%,92.7%,95.5%,93.2%,94.2%,respectively.The verification results are more robust than the constructed random forest model,indicating that the selected control has good representativeness.Conclusions1.The last prenatal b-ultrasound information currently used in clinical practice predicts a high rate of missed diagnosis of macrosomia.2.After 20 weeks of gestation,the developmental indexes of macrosomia significantly increased.3.The prediction effect of random forest algorithm is obviously better than that of support vector machine and logistic regression model.4.The shape of pregnant women’s pelvis plays an important role in predicting the occurrence of macrosomia.Using the measurements of pregnant women’s pelvis and the characteristics of pregnant women can predict macrosomia in early pregnancy.5.BMI between 26.5 and 32 before pregnancy increases the risk of having a baby.The risk of macrosomia decreased with a pre-pregnancy BMI greater than 32. |
PDF Full Text Request