Describing and predicting contraceptive discontinuation in war-torn settings: a machine learning approach

Conflict has a significant impact on maternal health services, and during the recent war in the Tigray region, nearly 70% of the healthcare facilities were looted, which made it difficult to access family planning services. Thus, this study aimed to assess the prevalence and predictors of contraceptive use discontinuation among women of reproductive-age in war-torn Tigray, Ethiopia, in 2023 using machine learning algorithms. A cross-sectional study of 6,414 women was carried out in the community utilizing stratified cluster sampling. Data were subjected both to descriptive statistics and predictive analysis. Predictive analysis on contraception discontinuation was performed using H2O machine learning methods, specifically ensemble decision tree methods (gradient boosting machine (GBM) and random forest (RF)). H2O is a scalable and distributed platform for machine learning and predictive analysis. For this study, H2O version 3.46 was used and integrated into STATA 19. The H₂O data frame was split into training and testing frames, and 80% of the observations were assigned to the training frame. The five-fold stratified cross-validation (SCV) method was employed to account for an imbalance between classes. The log loss, mean class error, area under the curve (AUC), area under the precision–recall curve (AUCPR), Gini coefficient, mean square error (MSE), and root mean square error (RMSE) were used as performance metrics. In addition, a surrogate tree was generated using Graphviz version 13.1. Variable importance and partial dependence plots were used as global explainability methods. The Shapley additive explanations (SHAP) summary plot and SHAP values plot were used as local explainability methods. The overall prevalence of contraceptive use discontinuation was 16.2%. Based on the SCV result, the AUC for GBM and RF was 0.614 and 0.596, respectively. During testing, the area under the curve for GBM and RF was 0.624 and 0.605, respectively. The Gini coefficient for GBM and RF was 0.227 and 0.191, respectively. Log loss, mean class error, MSE, and RMSE were lower for GBM compared to RF. Based on all performance metrics, GBM provided a better prediction model than RF. The most important variable for predicting contraception use discontinuation was age, followed by parity and the impact of war on the household. The partial dependence and SHAP plots of the GBM model for age and parity revealed that there is a decreasing trend for contraceptive use discontinuation as age and parity increase, respectively. There was a high probability of contraceptive use discontinuation among women who reported war impact, urban dwellers, those displaced during the war, and women who walked for an hour or more to reach a nearby health facility. The conflict had a detrimental effect on the percentage of women who discontinued using contraception. The results show that in a post-conflict situation, the Gradient Boosting Machine (GBM) model is capable of predicting the status of contraceptive discontinuation. The GBM model identified a mix of access, security, and service-related factors that significantly predict contraceptive use discontinuation. Addressing these challenges requires a multi-faceted approach.