Explainable machine learning algorithms to identify predictors of intention to use family planning among women of reproductive-age in Ethiopia: Evidence from the Performance Monitoring and Accountability (PMA) 2021 survey data set

Abstract

Introduction: Reducing maternal and infant mortality, preventing unintended pregnancies and improving the health of women and their families are all strongly associated with use of family planning (FP). It is widely believed that intentions are a strong predictor of behaviours, and many interventions that aim to change behaviour, including those targeting FP use, rely on evaluating programme effectiveness through analysis of behavioural intentions. Understanding a woman's intention to use FP is crucial in predicting and promoting its actual use. Thus, using explainable machine learning algorithms, this study aimed to identify the key determinants of intention to use FP among women of reproductive age in Ethiopia.

Methods: Secondary data from the Ethiopian Performance Monitoring and Accountability 2021 survey were analysed using R and Python on Google Colab. Eight machine learning classifiers were employed to identify significant determinants of intention to use FP in a weighted sample of 5993 women. Performance metrics evaluated these classifiers. Data preparation techniques, such as feature engineering, handling missing values and addressing imbalanced categories, were applied. A SHAP (SHapley Additive exPlanations) analysis identified the most influential predictors, clarifying their impact on model outcomes.

Results: Using 10-fold cross-validation and balanced training data, the random forest model achieved an accuracy of 77.0% (95% CI 74.73%, 79.33%) and an area under the curve of 85.0% (95% CI 81.43%, 88.63%), making it the most effective model. The SHAP analysis revealed the key determinants of intention to use FP, including age at first use of FP, partner's age, marital status, religion, pregnancy status, unmet needs for FP, family size and household relationship dynamics.

Conclusions and recommendations: This research highlights the sociodemographic, economic and personal factors influencing intention to use FP in Ethiopia. Addressing barriers such as perceived side effects, unmet needs for FP and partner involvement can improve FP uptake. Insights from this study can inform targeted interventions and policies to enhance the health and well-being of women in Ethiopia.

Keywords: Community Health; Health Services Accessibility; Public Health.

Figure 1

Flow chart of data preparation and analysis plan. ANN, artificial neural network; XGB, eXtreme Gradient Boosting; KNN, K-nearest neighbour; LR, logistic regression; NB, Naïve Bayes; PMA, Performance Monitoring and Accountability; RF, random forest; SHAP, SHapley Additive exPlanations; SVM, support vector machine.

Figure 2

Comparison of the random forest model’s predictions on the test data. AUC, area under the ROC curve; RF, random forest; ROC, receiver operating characteristic.

Figure 3

Global importance plot of the optimised random forest model. fp_ever_used_1=yes; partner_age_2=above 40 years; maritall_status_5=single/unmarried for marital status; religion_4=Muslim; pregnant_1=yes; ever_pregnant_1=yes, more_children_pregnant_2=no; unmete_need_for_limiting_1=no unmet need; relationship_3=son/daughter. SHAP, SHapley Additive exPlanations.

Figure 4

Beeswarm plot ranked by mean absolute SHAP value generated by the optimised random forest model. fp_ever_used_1=yes; partner_age_2=above 40 years; maritall_status_5=single/unmarried for marital status; religion 4=Muslim; pregnant_1=yes; ever pregnant_1=yes; partner_age_1=between 21 and 39 years; more_children pregnant_2=yes; unmete_need_for_limiting_1=having unmet need; age_at_first_use_30=above 30 years. SHAP, SHapley Additive exPlanations.