Explainable hybrid machine learning model for predicting stunting and identifying key risk factors among Ethiopian children under five

Abstract Childhood stunting, caused by poor nutrition, infections, and inadequate stimulation, impairs growth and development. This leads to reduced cognitive function and academic performance, increased risk of illness and death due to weakened immunity, lower economic productivity, and higher susceptibility to chronic diseases in adulthood. Childhood stunting is a major issue in low-income countries like Ethiopia. Children’s stunting prevalence in Ethiopia is 35.4%, highlighting the need for interventions. The World Health Organization (WHO) classifies stunting as normal, moderate, and severe. The study aims to develop an advanced explainable hybrid predictive model by combining two top-performing algorithms, Extra Tree (ensemble learning) and Multilayer Perceptron (deep learning), to leverage their complementary strengths. This approach enhances accuracy and reveals the inner workings of AI decision-making. The central goal of this study is to move beyond black box models by using post-hoc explainability methods like Local Interpretable Model-agnostic Explanations. This is intended to reveal the inner workings of the artificial intelligence decision-making and identify the key risk factors associated with stunting to build trust for public health policy design. In this study, data from the 2019 Ethiopian Demographic and Health Survey dataset were used for the experiment. The data was preprocessed to ensure high quality for analysis and to build a model predicting childhood stunting. This study employed an experimental research methodology to conduct five experiments, involving 11,121 instances and 17 features. While we developed the model using a balanced dataset, we evaluated it on the original dataset to avoid inflating performance metrics. The experiments showed that the hybrid model constructed from Extra Tree and Multilayer Perceptron performed better than others, achieving an accuracy of 94%. The hybrid model identified age_of_child, Region, Birth_Interval, and number of children under five in the household as the most influential predictors of stunting status in this dataset. It is essential to note that these are predictive risk factors, representing strong statistical associations within the cross-sectional data, and should not be interpreted as definitive causal drivers of stunting.