Machine Learning Models for the Prediction of Postpartum Depression: Application and Comparison Based on a Cohort...

Area of Science:

• Reproductive Health
• Mental Health Analytics
• Machine Learning in Healthcare

Background:

• Postpartum depression (PPD) presents a significant public health challenge.
• Early identification and intervention are facilitated by predictive modeling using antenatal data.

Purpose of the Study:

• To compare four machine learning models for PPD prediction using data collected during pregnancy.
• To identify the most influential factors for PPD prediction.

Main Methods:

• Utilized data from 508 women, including demographics, social factors, and mental health during pregnancy.
• Employed two feature selection methods (expert consultation, FFS-RF) and two algorithms (SVM, RF) to build four PPD prediction models.
• Assessed model performance using the Edinburgh Postnatal Depression Scale score post-delivery.

Main Results:

• No significant difference in prediction performance between feature selection methods, but FFS-RF reduced the number of factors by 10.
• The Support Vector Machine (SVM) model combined with Filter Feature Selection using Random Forest (FFS-RF) demonstrated the best predictive performance (AUC=0.78).
• Key predictors identified by the Random Forest (RF) algorithm included psychological resilience, third-trimester depression, and income level.

Conclusions:

• FFS-RF is effective for dimensionality reduction compared to expert consultation.
• SVM is a suitable algorithm for PPD prediction in studies with smaller sample sizes.
• Maternal psychological resilience is a critical factor for PPD prevention.