Explainable Artificial Intelligence Models for Birth Weight Prediction Based on Maternal Parameters in Ethiopia

Abnormal birth can have negative impacts on both mothers and infants, leading to complex deliveries and potential long-term health issues such as obesity during childhood and adolescence. This study aimed to develop explainable artificial intelligence models for the prediction of birth weight based on maternal parameters in Ethiopia. The data set was collected from the Ethiopian Demographic Health Survey (EDHS). Homogenous ensemble machine learning algorithms with class decomposition (one versus one and one versus the rest) and without class decomposition were employed after step-forward feature selection techniques. The model was further explained using the LIME, Eli5, and SHAP XAI tools to achieve model-specific interpretability. The performance of the models was evaluated and compared using the standard metrics of accuracy, precision, recall, and F1 score. The overall accuracy of the random forest, bagging, cat boosting, and extra tree with one-versus-rest methods was 95.0%, 95.3%, 95.8%, and 97.4%, respectively. The experimental results showed that the extra tree with one versus the rest outperformed with 97.4% accuracy compared to others, and the researchers decided to use the extra tree further using the XAI tool. The presented XAI techniques provide useful information to understand abnormal birth weight risk and predict results for family members and healthcare providers. Based on the XAI results, it was clear that the greatest risk factors for birth weight prediction were age, sex of the child, birth order, number of households, residency, education level, and religion.