Machine learning-optimized perinatal depression screening: Maximum impact, minimal burden.

Area of Science:

• Perinatal mental health
• Clinical informatics
• Machine learning applications

Background:

• Perinatal depression affects a significant portion of pregnant and postpartum women, with increased prevalence post-COVID-19.
• Rapid identification of affected women is a clinical priority.
• Existing screening tools like the Edinburgh Postnatal Depression Scale (EPDS) are lengthy, posing challenges in busy clinical settings.

Purpose of the Study:

• To develop and validate a machine learning (ML) framework for predicting full 10-item EPDS scores using abbreviated question subsets.
• To optimize screening brevity while maintaining predictive accuracy for perinatal depression.

Main Methods:

• Utilized National Clinical Cohort Collaborative (N3C) data (n=22,924) to train ML models predicting full EPDS scores from 2-5 item combinations.
• Validated models across diverse cohorts, including postpartum women (n=7,750) and an external pregnancy population (n=1,217).
• Assessed generalizability using the PHQ-9 (n=398,606) and evaluated clinical utility via decision curve analysis.

Main Results:

• Optimal 2-question EPDS subsets (Q4+Q8, Q5+Q8) achieved high predictive accuracy (R²=0.70).
• Binary classification models demonstrated strong performance (sensitivity 0.68-0.72, specificity 0.98-0.99).
• The ML framework showed robust generalization across validation cohorts and outperformed traditional scoring methods in clinical utility.

Conclusions:

• A machine learning framework using only two EPDS questions can maintain predictive accuracy comparable to the full 10-item scale.
• This abbreviated approach significantly reduces assessment burden for patients and clinicians.
• Implementation of this ML-driven screening could enhance perinatal depression identification in clinical practice, potentially reaching millions of women annually.