Predicting the Level of Anemia among Ethiopian Neonatal Using Ensemble Machine Learning Algorithms

Abstract

<title>Abstract</title> Anemia occurs when the body's physiological demands are not met by the quantity of red blood cells, reducing their oxygen-carrying capacity. According to the World Health Organization, neonatal anemia is classified as mild (Hb 10–13.9 g/dL), moderate (Hb 7–9.9 g/dL), normal (Hb 14–24 g/dL), and severe (Hb &lt; 7 g/dL), and is influenced by socioeconomic and demographic factors. Previous studies did not generate actionable rules for policymakers, design or deploy artifacts, or construct multi-class predictive models for neonatal women's anemia based on these factors using machine learning. This study develops a predictive model and prototype for neonatal women's anemia level using data from the Ethiopian Demographic Health Survey (2005–2016). Data preprocessing ensured high-quality input suitable for machine learning. Following a design science research strategy, four experiments were conducted on 42,376 instances with 22 features, splitting data into training and testing sets (80/20) and applying Random Forest, XGBoost, Decision Tree, and CatBoost algorithms, achieving accuracies of 98.37%, 98.22%, 97.92%, and 74.43%, respectively. Random Forest was selected as the best algorithm based on objective and subjective evaluations. Feature importance analysis identified key determinants: age, region, contraceptive use and intention, body mass index, diarrhea, respondent weight, husband’s occupation, cooking fuel type, literacy, wealth index, women’s education, and place of residence. These factors guided artifact design using a Flask framework and deployment on PythonAnywhere cloud platform. Subjective evaluation indicated 88% user acceptance. The model is not integrated with a knowledge-based system; future research should integrate it to develop an intelligent system for automated prediction of neonatal women’s anemia levels.