Developing a machine learning prediction model for postpartum psychiatric admission: Findings from the born in...

We utilised novel prenatal depression screening data in a population sample to develop a machine learning-based prediction model for postpartum psychiatric admissions.

We used health administrative data from Queensland, Australia, containing 945 clinical and demographic variables for 205,495 pregnancies between July 2015 and January 2021. Two machine learning algorithms, gradient-boosted trees and elastic net regularisation, were compared to predict postpartum psychiatric admissions occurring within 12 months of neonatal discharge.

Gradient-boosted trees outperformed elastic net regularisation, achieving good discrimination in the test set [AUC = 0.79, 95 % CI = (0.76, 0.82)]. As expected, the inclusion of the Edinburgh Postnatal Depression Scale ascertained during pregnancy improved the model's predictive performance. Antenatal mental health diagnoses, being a single mother and smoking during pregnancy were also found to be strong risk factors.

This study contained no information on community or outpatient mental health visits and excluded private obstetrics patients.

Our prediction model may enable early identification and timely intervention for at-risk women during pregnancy which, in turn, may reduce the severity of postpartum mental health problems or the need for postpartum psychiatric admissions.