Development and validation of a machine learning-based postpartum depression prediction model: A nationwide cohort...

Area of Science:

• Reproductive Medicine
• Psychiatry
• Data Science

Background:

• Current postpartum depression (PPD) screening relies on subjective self-reports, lacking objective tools for early risk identification.
• There is a need for integrative, objective methods to identify women at high risk for PPD before its clinical manifestation.

Purpose of the Study:

• To develop and validate a machine learning (ML) model for predicting PPD risk using electronic health record (EHR) data.
• To identify novel predictors of PPD through ML analysis of comprehensive EHR information.

Main Methods:

• A nationwide longitudinal cohort of 214,359 births (2008-2015) from Israel's largest health maintenance organization was utilized.
• A gradient-boosted decision tree algorithm was applied to sociodemographic, clinical, and obstetric features extracted from EHRs.
• PPD was defined as a new diagnosis or antidepressant prescription within one year postpartum.

Main Results:

• The ML model achieved an AUC of 0.712 in the validation set, with 34.9% sensitivity and 90.5% specificity at the 90th percentile risk threshold.
• The model identified PPD cases at a rate over three times higher than the overall cohort prevalence.
• Key predictors included prior depression history and unique patterns in blood test results.

Conclusions:

• Machine learning models using EHR data can enhance PPD screening by identifying high-risk individuals.
• This approach can facilitate timely preventive interventions for women at risk of developing PPD.
• Integrating ML predictions with clinical practice may significantly improve PPD management and outcomes.