Xin Xu1, Hongzhuan Tan2, Shujin Zhou3, Yue He1, Lin Shen1, Yi Liu1, Li Hu1, Xiaojuan Wang1, Xun Li1. 1. School of Public Health, Central South University, Changsha 410008, China. 2. School of Public Health, Central South University, Changsha 410008, China. Email: tanhz99@qq.com. 3. Liuyang Hospital for Maternal and Child Health Care.
Abstract
OBJECTIVE: To establish a practical and effective model in predicting the premature birth, using the Back-Propagation Artificial Neural Network (BPANN). METHODS: This was a prospective cohort study. Data was gathered from pregnant women selected by cluster sampling method from 2010 to 2012 in Liuyang city, Hunan province and was randomly divided into training sample (to establish the prediction models), validation sample (to select the optimal network) and testing sample (to evaluate the prediction models) by ratio of 2:1:1. BPANN and logistic regression analysis were used to establish models while ROC was applied to evaluate the 'prediction models'. RESULTS: Among the 6 270 pregnant women, 265 premature births were seen, with the premature incidence as 4.22%. The 7 variables which entered into the forecasting model would include abnormal uterine or uterine deformity, parity, number of pregnancies, gestational hypertension, placenta previa, premature rupture of membrane and regular prenatal examination. Sensitivity, specificity, agreement rate and area under the ROC curve of BPANN were 67.65% , 84.87%, 84.12% and 0.795, respectively. However, the sensitivity, specificity, agreement rate and area under the ROC curve of logistic regression were 64.71%, 85.60%, 84.69% and 0.783, respectively. CONCLUSION: The newly established BPANN model was practical and reliable, which proved that this model was slightly better than the logistic regression in the prediction of premature birth.
OBJECTIVE: To establish a practical and effective model in predicting the premature birth, using the Back-Propagation Artificial Neural Network (BPANN). METHODS: This was a prospective cohort study. Data was gathered from pregnant women selected by cluster sampling method from 2010 to 2012 in Liuyang city, Hunan province and was randomly divided into training sample (to establish the prediction models), validation sample (to select the optimal network) and testing sample (to evaluate the prediction models) by ratio of 2:1:1. BPANN and logistic regression analysis were used to establish models while ROC was applied to evaluate the 'prediction models'. RESULTS: Among the 6 270 pregnant women, 265 premature births were seen, with the premature incidence as 4.22%. The 7 variables which entered into the forecasting model would include abnormal uterine or uterine deformity, parity, number of pregnancies, gestational hypertension, placenta previa, premature rupture of membrane and regular prenatal examination. Sensitivity, specificity, agreement rate and area under the ROC curve of BPANN were 67.65% , 84.87%, 84.12% and 0.795, respectively. However, the sensitivity, specificity, agreement rate and area under the ROC curve of logistic regression were 64.71%, 85.60%, 84.69% and 0.783, respectively. CONCLUSION: The newly established BPANN model was practical and reliable, which proved that this model was slightly better than the logistic regression in the prediction of premature birth.