Early Prediction of Postpartum Depression Using Machine Learning on Electronic Health Records

Postpartum depression (PPD) is an issue that occurs in
a good percentage of women in the world with a result of poor
maternal and neonatal outcomes in case of its absence. Such a
problem lies in the fact that, as early as possible, it is difficult to
detect because the headache symptoms manifest too insensitively
and not all the clinical data are considered over a long period of
time. The purpose of the study was to create a machine learning
platform based on the use of electronic health records (EHR) to
predict PPD early. The cohort of women between 18 and 45 who
gave live birth was examined retrospectively and those with prior
severe psychiatric conditions were excluded. Formatted EHR data
were obtained, processed, and engineered to longitudinal features
representing laboratory data, vital signs, medical care use, and
behavioral data. Temporal and static features were used to train
machine learning models such as logistic regression, random
forest/XGBoost. The proposed method attained an AUROC of
0.79, AUPRC of 0.41, sensitivity of 84, and specificity of 65 which
was better than conventional static models. Combining machine
learning with temporal feature engineering allows identifying
high-risk women early and developing a platform of proactive
interventions to offset the negative maternal and infant outcomes.