Predicting Prenatal Depression and Assessing Model Bias Using Machine Learning Models

Area of Science:

• Reproductive Health
• Medical Informatics
• Health Disparities

Background:

• Perinatal depression affects 10-20% of pregnant individuals, with higher prevalence and lower diagnosis rates in Black and Latina women.
• Existing machine learning (ML) models for postpartum depression often lack diversity, leading to biased predictions.
• This study addresses the underrepresentation of racial/ethnic minorities in ML models for predicting perinatal depression.

Purpose of the Study:

• To evaluate the efficacy of ML models in predicting early pregnancy depression among racial/ethnic minority women.
• To leverage electronic medical record (EMR) data for depression risk prediction in diverse populations.
• To identify potential biases in ML model performance across different racial/ethnic groups.

Main Methods:

• Utilized EMR data from 5875 women at a U.S. urban hospital serving low-income Black and Hispanic populations.
• Assessed depressive symptom severity using the Patient Health Questionnaire-9 (PHQ-9).
• Employed multiple ML classifiers and Shapley additive explanations for interpretation and bias assessment using four metrics.

Main Results:

• The best ML model (elastic net) achieved low predictive performance (AUC = 0.61).
• Identified known risk factors (unplanned pregnancy, single status) and novel factors (pain, low vitamin intake, asthma, male fetus, low platelets).
• The model performed worse for Black (AUC=57%) and Latina (AUC=59%) women compared to White women (AUC=64%), despite the sample's minority focus.

Conclusions:

• EMR-based ML models offer moderate potential for predicting early pregnancy depression.
• Significant performance biases were observed against low-income minority women.
• Addressing data diversity and model fairness is critical for equitable perinatal mental healthcare.