Predicting prenatal depression and assessing model bias using machine learning models.

Area of Science:

• Reproductive Health
• Computational Medicine
• Health Disparities

Background:

• Perinatal depression (PND) affects 10-20% of pregnant individuals, with higher prevalence in Black and Latina women who face disparities in diagnosis and treatment.
• Existing machine learning (ML) models for PND prediction, often trained on data from majority populations, exhibit biases and underperformance when applied to racial and ethnic minorities.
• This study addresses the need to evaluate ML model efficacy and fairness in predicting early pregnancy depression specifically within a predominantly low-income, minority patient population.

Approach:

• Utilized Electronic Medical Records (EMRs) from 5,875 patients at an urban hospital serving a large Black and Hispanic population.
• Assessed depressive symptom severity using the PHQ-9 self-reported questionnaire.
• Investigated multiple ML classifiers, employed SHAP for interpretability, and quantified prediction bias using Disparate Impact and Equal Opportunity Difference metrics.

Key Points:

• Identified known PND predictors (e.g., unplanned pregnancy, single status) and novel factors (e.g., self-reported pain, low prenatal vitamin intake, asthma, male fetus, low platelet count).
• The ML model (Elastic Net) achieved moderate predictive performance (ROCAUC=0.67) but demonstrated lower accuracy for the 75% low-income minority sample (54% Black, 27% Latina).
• Model performance was notably lower for minority women, indicating a significant bias against these groups.

Conclusions:

• Machine learning models show potential for moderately predicting perinatal depression in early pregnancy using EMR data.
• Significant performance disparities exist, with ML models demonstrating inherent biases against low-income and minority women.
• Further research is crucial to develop equitable and accurate ML tools for diverse maternal health populations.