Predicting women's intention to use contraceptives in East Africa: a machine learning analysis of predictors

Intention to use contraceptives reflects an individual's or couple's plan to adopt contraceptive methods, supporting women's reproductive autonomy. It is associated with reduced unintended pregnancies, unsafe abortions, and high fertility rates, thereby improving maternal and infant health outcomes. Machine learning approaches can strengthen prediction accuracy and support evidence-based reproductive health planning. This study aimed to predict women's intention to use contraceptives in East Africa using machine-learning methods.

We analyzed Demographic and Health Survey (DHS) data from 11 East African countries (2015-2024) in a community-based cross-sectional design. Country-specific sampling weights, stratification, and clustering were applied to account for the complex survey design. Missing data were imputed using the KNNImputer, and predictors were harmonized across surveys. Data preprocessing included cleaning, transformation, integration, and one-hot encoding, with an 80/20 train-test split. Seven machine learning algorithms were evaluated: adaptive boosting, CatBoost, random forest, light gradient boosting, extreme gradient boosting, logistic regression, and decision tree. Hyperparameters for CatBoost were tuned using Bayesian optimization under stratified 10-fold cross-validation. Model transportability was assessed using leave-one-country-out cross-validation.

Among 123,290 reproductive-age women in East Africa, 51.23% reported intention to use contraceptives. Boosted tree algorithms performed best, particularly CatBoost achieving the highest discrimination with an AUC of 80.09% and an accuracy of 73.48%. Leave-one-country-out cross-validation confirmed moderate transportability with an AUC 74%, while calibration analysis showed reliable probability estimates (Brier score 0.179). SHAP feature importance identified employment, education, age, hearing about family planning, pregnancy, breastfeeding, barriers to healthcare access, and fertility preference as the most influential predictors of contraceptive intention.

This study provides a transparent, reproducible framework for pooled DHS prediction modeling, offering actionable insights for policymakers and health planners while serving as a methodological foundation for future applied work.