A Machine Learning Approach to Screening for Maternal Intimate Partner Violence In Pediatric Settings

Background and Purpose: Intimate partner violence (IPV) affects one in three Peruvian women of reproductive age annually, with over half reporting lifetime exposure, yet 44% of survivors never seek help. Current screening relies on self-disclosure, constrained by stigma, fear, and lack of trusted provider encounters. Meta-analytic evidence links maternal IPV to low birth weight, preterm birth, and inadequate breastfeeding, while regional studies associate IPV with elevated childhood anemia, though other child health associations remain inconsistent. These findings suggest that routinely collected child health and birth history data may serve as indirect indicators of maternal IPV victimization. Pediatric child visits offer a frequent, consistent contact point for families who may never seek IPV services. This study aims to develop a machine learning model that identifies maternal IPV victimization using child health indicators available during routine pediatric visits, without requiring direct parental disclosure.

Methods: We used pooled data from Peru’s Demographic and Health Survey (DHS; ENDES) 2016-2024. A total of 140,355 child-mother records from 2016-2022 were used for model development and 39,642 records from 2023-2024 were held-out for external validation. ENDES links child biomarker, immunization, and morbidity data to maternal IPV victimization assessed through a standardized DHS domestic violence module. The outcome was any lifetime IPV victimization (emotional, physical, or sexual). Predictors were restricted to indicators observable during standard pediatric visits (excluding any IPV question): child anthropometric Z-scores, hemoglobin, anemia severity, birth weight, breastfeeding duration, acute respiratory infection, fever, diarrhea, vaccination status, child age, sex, birth order, and household characteristics (wealth index, rural residence, maternal education). We evaluated gradient-boosted tree models (HistGradientBoosting, XGBoost, LightGBM, and CatBoost) and ensemble models (soft voting and stacking), with hyperparameter tuning via randomized search and Optuna where applicable. Discrimination was evaluated via AUC-ROC, with sensitivity, specificity, and predicted values calculated at thresholds prioritizing high sensitivity.

Results: We evaluated six models across multiple sample specifications. Performance was consistent across classifiers, with test AUCs ranging from 0.689 to 0.697. The best-performing model (stacking ensemble, youngest-child-only sample; N = 123,444 training, 35,764 validation) achieved a cross-validated AUC of 0.697 (SD = 0.004) and external test AUC of 0.697, demonstrating stable generalization to held-out data. At the selected threshold, sensitivity was 0.755, specificity was 0.512, PPV was 0.264, and NPV was 0.900. The model correctly flagged three in four IPV-victimized mothers, though with a high false-positive rate reflecting the tradeoffs of a sensitivity-prioritized screening approach. Observed discrimination is consistent with comparable proxy-based social determinant prediction models reported in the literature (AUC: 0.65-0.78).

Conclusions and Implications: This proof-of-concept study demonstrates that pediatric clinical data reflect measurable patterns of maternal IPV victimization. The findings suggest a novel pathway for identifying survivors who may not disclose during conventional screening approaches (i.e., self-report). For social work practice, this approach could reduce reliance on self-disclosure as the primary means of IPV identification. It could also improve access to support services for survivors. Future implementation efforts should integrate this approach into trauma-informed referral pathways spanning pediatric, maternal, and family service settings.