Employing machine learning models to predict pregnancy termination among adolescent and young women aged 15-24 years in East Africa

Abstract

Pregnancy termination is still a sensitive and continuing public health issue due to several political, economic, religious, and social concerns. This study assesses the applications of machine learning models in the prediction of pregnancy termination using data from eleven national datasets in East Africa. Nine machine learning models, namely: Random Forests (RF), Decision Tree, Logistic Regression, Support Vector Machine, eXtreme Gradient Boosting (XGB), AdaBoost, CatBoost, K-nearest neighbor, and feedforward neural network models were used to predict pregnancy termination, with six evaluation criteria utilized to compare their performance. The pooled prevalence of pregnancy termination in East Africa was found to be 4.56%. All machine learning models had an accuracy of at least 71.8% on average. The RF model provided accuracy, specificity, precision, and AUC of 92.9%, 0.87, 0.91, and 0.93, respectively. The most important variables for predicting pregnancy termination were marital status, age, parity, country of residence, age at first sexual activity, exposure to mass media, and educational attainment. These findings underscore the need for a tailored approach that considers socioeconomic and regional disparities in designing policy initiatives aimed at reducing the rate of pregnancy terminations among younger women in the region.

Keywords: Abortion; Artificial intelligence; Pregnancy termination; Women.

Fig. 1

The main steps of the workflow.

Fig. 2

Data comparison before and after applying the data balancing technique.

Fig. 3

Prevalence of pregnancy termination among adolescent and young women across East African countries.

Fig. 4

Feature selection using Boruta algorithm. v01: age, v02: age at first sex, v03: educational level, v04: religion, v05: working status, v06: smoking status, v08: place of residence, v09: distance to HF, v00: country of residence, v10: marital status, v11: wealth index, v12: mass media exposure, v13: number of children ever born, v14: number of births in the last five years, v15: community level educational status, v16: community level poverty, v17: community level mass media exposure.

Fig. 5

Model performance of each ML model after optimized Hyperparameter tuning.

Fig. 6

ROC curve analysis of each ML model after optimized Hyperparameter tuning.

Fig. 7

Random forest-based feature importance.

Fig. 8

SHAP feature impact on model prediction.

Fig. 9

SHAP force plot.