407: Risk assessment for adverse neonatal outcomes in pregnancies complicated by gestational diabetes using machine learning

Abstract

To assess machine learning and neural network model ability to predict adverse neonatal outcome in pregnancies complicated by gestational diabetes Machine learning was used to create pregnancy complication prediction model in diabetic women. Our model was coded using Python3.6 with Keras framework based on Google's TensorFlow. The model was implemented using a 4-layers fully-connected neural network: data were fed after batch normalization; another layer of dropout was applied to prevent the model from memorizing the training samples and overfitting the data. Data were retrieved from the medical records of women diagnosed with gestational diabetes mellitus, who delivered in our medical center (2012-2016). The following baseline characteristics were included: maternal age, body mass index, parity, gravity, oral glucose test results, diabetes treatment and glycemic control. We defined a composite neonatal adverse outcome including any of the following: large for gestational age neonate, shoulder dystocia, umbilical vein pH< 7.2, neonatal intensive care unit admission, respiratory distress syndrome, hyperbilirubinemia and polycythemia. For the machine training phase 70% of the cohort was randomly selected, each sample included baseline parameters and the composite outcome. We have then used the rest of the samples to evaluate our model’s accuracy. The baseline parameters were fed into the trained model and the predicted outcome was compared to the actual outcome 452 women with gestational diabetes mellitus were included. The incidence of composite neonatal adverse outcome was 40%. On the randomly picked samples used to evaluate our model, we demonstrated an accuracy rate of 91% in predicting adverse outcome A state-of-the-art machine learning algorithm presented promising ability to predict adverse neonatal outcome among women with gestational diabetes mellitus. The algorithm provides an opportunity to identify at-risk patients who may benefit from early fetal monitoring and intervention