# Machine learning to predict the role of CHWs in shifting birth preferences away from homebirth in India

## TL;DR

This study uses machine learning to predict birth locations in India, focusing on how community health workers can influence decisions based on women's perceptions of violence.

## Contribution

The study identifies actionable insights for improving public health policy through efficient targeting of community health worker interventions.

## Key findings

- Random Forest achieved the highest test AUC (0.991) and accuracy (96.7%) among evaluated models.
- Improved prediction of homebirth cases can help prioritize community health worker contact and shift birth preferences.

## Abstract

This study utilized well-known supervised machine learning algorithms to NFHS‑5 data of West Bengal, India, to predict the place of birth (home vs facility) by integrating CHW (community health worker) contact factors and women participant’s perceptions about intimate partner violence (IPV). Although the study applied modelling techniques from conventional ML literature, the overarching contribution was identifying avenues to enhance public health policy response (e.g., efficient targeting of home visits and counselling by ANM/ASHA). The study concludes that, identifying likely homebirth cases among women with IPV-related poor perceptions applying improved prediction can enhance prioritising of CHW-contact and alter birth preference. The study improves minority-class learning using SMOTE on weighted NFHS data keeping in mind the complex survey design and SMOTE limitations. With respect to the ML model performance, Random Forest produced the highest test AUC (0.991) and accuracy (96.7%) among the 5 evaluated classifiers—LR (base), RF, MNB, k-NN, SVM and 0.950 with stable accuracy of 96% on hold-out data. The study does not bring methodological novelty in the underlying algorithms but generated actionable insights for equitable CHW allocation for efficient targeting using standard cross-sectional survey data.