Interpretable machine learning for predicting delays in seeking abortion among reproductive-aged women in Ethiopia: A study using EDHS 2016 data

Abstract

Delayed access to abortion care in Ethiopia poses significant public health risks, yet it has not been studied using advanced machine learning models with interpretable techniques. This study aims to identify its key predictors through Shapley Additive Explanations (SHAP) values. The study used data from the 2016 Ethiopian Demographic and Health Survey. Data preprocessing tasks such as feature engineering, lumping, filtering, and encoding were performed before model building. Eight machine learning models, including LightGBM, Support Vector Machine, Random Forest, and XGBoost, were employed to predict delays in seeking an abortion. SHAP analysis was used to interpret feature importance and understand individual variable contributions. The prevalence of delayed abortion seeking was 1109 (54.3%). The Random Forest model performed the best, with an accuracy of 91.8% (95% CI: 89.3, 93.8) and an AUC of 97.6, effectively predicting delays in abortion-seeking behavior. SHAP analysis revealed that age (women aged 40-49), regional factors (residing in the Somali and Amhara regions), and lack of media exposure were strong positive contributors to delays. In contrast, urban residence and living in Addis Ababa were associated with a lower likelihood of delay. Alcohol consumption also showed a positive association with delay. The study identifies key factors influencing delays in seeking abortion services in Ethiopia, highlighting the importance of targeted interventions, especially for older women and those in rural regions. These findings offer valuable insights for designing public health initiatives aimed at reducing unsafe abortion-related maternal morbidity and mortality.