|
Abstract
|
OBJECTIVES: Postpartum depression is a serious health issue beyond mental problem that affects mothers after childbirth. There are no predictive tools available to screen postpartum depression that allow early interventions. We aimed to develop predictive models for postpartum depression using maternal and paternal characteristics based on Machine Learning (ML) approaches.
METHODS: We performed a retrospective cohort study using data from the Pregnancy Risk Assessment Monitoring System (PRAMS), 2012-2013 (n = 72,541). Significant maternal and paternal risk factors in relation to postpartum depression were selected by univariate analyses for the predictive models including maternal age, race/ethnicity, education, marital status, pre-pregnancy body mass index, smoking status, drinking status, previous history of depression, physical activity, number of previous live births, infant alive status, small-for-gestational-age, large-for-gestation-age, and paternal race/ethnicity. The imbalance between postpartum depression and normal cases was addressed by Synthetic Minority Over-sampling Technique (SMOTE). Ten different ML algorithms, including k-Nearest Neighbor, Decision Tree, Support Vector Machine (SVM), Stochastic Gradient Descent, Random Forest, Neural Network, Naïve Bayes, Logistic Regression, CN2 Rule Inducer, and AdaBoost available in Orange data mining toolkit, were employed with a 10-fold cross-validation to evaluate the models.
RESULTS: Removal of records with any missing data and application of SMOTE resulted in a dataset containing 67,902 records. Overall classification accuracies of the ten models ranged from 50.5 (SVM) to 77.7% (AdaBoost). AdaBoost achieved the highest classification accuracy of 77.7% and the highest under the ROC curve (AUC) value of 0.846. Random forest performed similarly as AdaBoost achieving the second best accuracy of 77.1% with an AUC value of 0.845.
CONCLUSIONS: Predictive modeling developed using ML may be used for prediction (screening) tools for postpartum depression in the future studies. FUNDING SOURCES: None.
Language: en |