Abstract
OBJECTIVE: Women with intersecting identities, such as being both Black and disabled, face heightened risk of antenatal depression, yet few studies examine its nuanced mechanisms. To capture complex, interactive associations among risk factors, we applied explainable machine learning to predict antenatal depression and identify key predictors among non-Hispanic Black (NHB) and non-Hispanic White (NHW) women with and without disabilities in the U.S. METHODS: Using 2019 Pregnancy Risk Assessment Monitoring System data merged with its disability supplement (n = 23,104), we developed random forest models for four subgroups defined by race/ethnicity and disability status. Model performance was evaluated using repeated 10-fold cross-validation with AUC. Variable importance and its stability were assessed through 50 refits of the final models with optimal hyperparameters. RESULTS: NHW women and women with disabilities had higher rates of antenatal depression. Model performance was strong across subgroups (AUC: 0.79-0.89). Depression before pregnancy was the strongest predictor, followed by hypertension during pregnancy or smoking across subgroups. Having at least one disability contributed more strongly to prediction among NHB women, whereas depression screening was uniquely predictive among NHW women. CONCLUSIONS: Antenatal depression risk is shaped by women's intersecting identities. Nuanced subgroup differences should inform more targeted and equitable prevention strategies.