Work

Predictors of weight gain in under-five children with severe acute malnutrition: A machine learning analysis of a tertiary care hospital dataset

Year: 2026

Type: article

Abstract: BACKGROUND & AIMS: Nutritionally rehabilitating children admitted to hospitals with severe acute malnutrition (SAM) and acute illness can pose significant challenges. Understanding the predictors of weight gain in this vulnerable group remains limited. Advanced analytical methods leveraging machine learning (ML) approaches adeptly handle the complexity of the data involved and identify predictors of clinical outcomes in a comprehensive and interpretable manner. This work aimed to predict weight gain in hospitalized children with SAM using explainable ML approaches. METHODS: We included 2183 children aged 0-59 months with SAM hospitalized in the Dhaka Hospital at the International Centre for Diarrhoeal Disease Research, Bangladesh (icddr,b) between 2011 and 2019. Service data were collected during hospitalization and analyzed retrospectively. Seven ML models were developed to predict weight gain from hospital admission to discharge (continuous outcome) and 5% or 15% weight gain during hospitalization (binary outcomes). Explainability analysis was conducted using Shapley values to interpret the most important predictors. RESULTS: Among the ML algorithms evaluated, Random Forest and XGBoost demonstrated consistently superior performance across both continuous and binary outcomes. The ML models identified MUAC at admission, age, baseline height-for-age z-score (HAZ), and admission weight as the most important predictors of weight gain when analyzed as a continuous outcome. When weight gain was treated as a binary outcome, admission anthropometry, including MUAC, WHZ, and WAZ, followed by age, formula feeding, and family income, were found among the most important predictors of weight gain during hospitalization. Shapley explainability analysis revealed that baseline anthropometric measurements (e.g., MUAC and weight), duration of formula feeding, and household roof condition, had a positive contribution, while age had a negative contribution on>15% weight gain. Weight gain models had moderate-low performance (RMSE = 0.38-0.63, MAE = 0.24-0.34). For weight gain categories of >5% or >15% during hospitalization, the average accuracy was 60% (50%-70%), while sensitivity, specificity, AUC, and positive and negative predictive values also varied between 50% and 70%. CONCLUSIONS: Tree-based ensemble methods showed the greatest potential for future clinical application. This analysis, using explainable ML models, provided valuable insights into the role of baseline anthropometry and the importance of feeding practices in attaining weight from hospital admission to discharge in children under five treated for SAM. This finding has important implications for future research on childhood feeding practices, developing personalized nutrition strategies, and community-based detection of children with SAM.

Source: Clinical Nutrition ESPEN

DOI: 10.1016/j.clnesp.2026.103119