Machine learning prediction models for postpartum depression: A multicenter study in Japan

Aim: Postpartum depression (PPD) and perinatal mental health care are of growing importance worldwide. Here we aimed to develop and validate machine learning models for the prediction of PPD, and to evaluate the usefulness of the recently adopted 2-week postpartum checkup in some parts of Japan for the identification of women at high risk of PPD. Methods: A multicenter retrospective study was conducted using the clinical data of 10 013 women who delivered at ≥35 weeks of gestation at 12 maternity care hospitals in Japan. PPD was defined as an Edinburgh Postnatal Depression Scale score of ≥9 points at 4 weeks postpartum. We developed prediction models using conventional logistic regression and four machine learning algorithms based on the information that can be routinely collected in daily clinical practice. The model performance was evaluated using the area under the receiver operating characteristic curve (AUROC). Results: In the machine learning models developed using clinical data before discharge, the AUROCs were similar to those in the conventional logistic regression models (AUROC, 0.569–0.630 vs. 0.626). The incorporation of additional 2-week postpartum checkup data into the model significantly improved the predictive performance for PPD compared to that without in the Ridge regression and Elastic net (AUROC, 0.702 vs. 0.630 [p < 0.01] and 0.701 vs. 0.628 [p < 0.01], respectively). Conclusions: Our machine learning models did not achieve better predictive performance for PPD than conventional logistic regression models. However, we demonstrated the usefulness of the 2-week postpartum checkup for the identification of women at high risk of PPD.