EP17.14: Prediction of stillbirths using machine learning on first trimester clinical data

To explore the performance of machine learning techniques in predicting stillbirth using clinical information collected at the first prenatal visit. Retrospective observational prognostic accuracy study including 396,220 pregnancy outcomes occurred between 2014 and 2018. Data were extracted from the Victorian Agency for Health Information. The primary outcome was stillbirth occurrence. Features included: maternal age, BMI, parity, assisted reproduction, fetal sex, presence of congenital abnormalities, type 1 or 2 diabetes mellitus, pre-existing hypertension, previous stillbirth, pregnant woman's nationality, indigenous status, history of smoking before 20 weeks of pregnancy, alcohol, cannabis, opioid, or amphetamine use, and IRSD quintile. Logistic regression, naïve Bayes, random forest, bagging, extra, light gradient and adaptive boosting only or combined with isolation forest and minimum covariance determinant were used. Multilayer perceptron and autoencoders were used as deep learning strategies. Shapley Additive Explanations (SHAP) values were utilised to standardise the identification and measurement of feature importance. The rate of stillbirths was 3.0 per 1,000 births. The recall varied between 42.7-48.7%, while the specificity ranged between 67.3-73.7%. Autoencoders improved recall in detecting stillbirth by 10-20%, reaching a sensitivity of 62%. Combining both machine and deep learning, achieved the highest AUROC (59.2%), balancing recall and specificity. Higher social disadvantage, higher BMI, alcohol, opioid, amphetamine use, or smoking, and maternal country of birth from Southern/Eastern Europe and Sub-Saharan Africa were the most relevant predictors of stillbirths. Combining traditional classifiers, one-class classification algorithms and deep learning models increased prediction accuracy. However, there was a higher false positive rate. Incorporating more phenotypic or genotypic characteristics into prediction, as well as new information that arises during the pregnancy, may be essential to improve model performance.