Explainable Machine Learning for Early Malnutrition Risk Prediction Using Clinical and Socioeconomic Determinants

Quinteros-Navarro D.; Soto M.; Quispe-Arbildo D.; Ponce D.; Marquez C.; Rojas M.; Maguiña A. · Ingenierie des Systemes d'Information (2026) · DOI: 10.18280/isi.310222

Resumen

Early identification of malnutrition risk remains a major challenge in heterogeneous clinical settings, where conventional screening tools may have limited predictive accuracy and adaptability. This retrospective proof-of-concept study developed and internally validated machine learning models for malnutrition risk prediction using 1,000 anonymized patient records from a public referral hospital in Lima, Peru. The dataset integrated demographic, clinical, biochemical, and socioeconomic variables, including age, body mass index, serum albumin, food insecurity, and household income. Random Forest, XGBoost, and deep neural network models were evaluated using stratified 10-fold cross-validation and compared with conventional nutritional screening approaches. The deep neural network achieved the highest discriminative performance (AUC = 0.90), whereas XGBoost provided the most balanced overall performance and more stable calibration. Explainability analysis based on SHAP identified serum albumin, body mass index, and socioeconomic indicators as the strongest contributors to risk prediction. These findings indicate that machine learning models can improve early malnutrition risk stratification by integrating multidimensional determinants beyond conventional screening variables. However, external validation in independent multicenter cohorts is required before clinical deployment. Copyright:

Instituciones: Universidad Tecnológica del Perú; Universidad Nacional Federico Villarreal