Improving machine learning models through explainable AI for predicting the level of dietary diversity among Ethiopian preschool children.

Abstract

Background: Child nutrition in Ethiopia is a significant concern, particularly for preschool-aged children. Children must have a varied diet to ensure they receive all the essential nutrients for good health. Unfortunately, many children in Ethiopia lack access to a range of foods, which can lead to malnutrition and other health issues. While machine learning (ML) has the potential to analyse extensive datasets, the lack of transparency in these models can impede their effectiveness in real-world applications, especially in public health. This research aims to enhance machine learning models by integrating Explainable AI (XAI) methods to more accurately predict the level of dietary diversity in Ethiopian preschool children.

Methods: To Improve the ML Model for Predicting the Level of Dietary Diversity among Ethiopian Preschool Children. We employed an ensemble ML approach with XAI. The Ethiopian demographic health survey collected a dataset consisting of dietary information and relevant socioeconomic variables. The data were preprocessed to obtain quality data that are suitable for the ensemble ML algorithms to develop a model. We applied filter (chi-square and mutual information) and wrapper (sequential backwards) feature selection methods to identify the most influential factors for dietary diversity (DD). Ethiopia demographic health survey (from 2011 to 2019). Datasets were used. We developed a predictive model using a decision tree, random forest, gradient boosting, light gradient boosting, CatBoost, and XGBClassifier. We evaluated it using accuracy, precision, recall, F1_score, and receiver operating characteristic (ROC)-based evaluation techniques.

Results: The ensemble ML models exhibited robust predictive performance, and light gradient boosting outperformed the other ensemble ML algorithms by 95.3%. The explainability of the Light Gradient Boosting Ensemble Model was determined using Eli5 and LIME. The child's age, household wealth index, household region, source of drinking water, frequency of listening to the radio, and mother's education level were the most crucial variables for the prediction of Minimum Dietary Diversity (MDD) in Ethiopia.

Conclusions: The research effectively demonstrated that integrating Explainable AI with machine learning can accurately predict dietary diversity in preschoolers in Ethiopia. The results of this study have significant implications for stakeholders in child development and nutrition, as well as for policymakers and medical experts. Targeted interventions and policies to enhance the nutritional health of Ethiopian preschool children are made possible by the explainable AI model that has been constructed.

Trial registration: Retrospectively registered.