Prediction and interpretation of factors associated with low birth weight using machine learning with MINSA data, Peru

Andrade-Girón D.; Neri-Ayala A.; Aguilar-Luna-Victoria M.; Oscuvicla-Tapia E.; Peña A.; Cuevas-Huari E. · Revista Cientifica de Sistemas e Informatica (2026) · DOI: 10.51252/rcsi.v6i2.1388

Resumen

Low birth weight is a significant public health issue in Peru. The objective was to develop, evaluate, and interpret machine learning models to predict it using national records from the Ministry of Health. A retrospective observational study was conducted involving 4,873,146 births registered between 2015 and 2025. Sixteen predictors—covering maternal, obstetric, neonatal, territorial, and healthcare-related factors—were used. CatBoost, LightGBM, and XGBoost were compared using Average Precision and ten stratified partitions from 2023. The selected model was trained on data from 2015–2022, calibrated using 2023 data, and temporally evaluated on data from 2024 and 2025. CatBoost achieved the best performance, with an Average Precision of 0.680 and an ROC-AUC of 0.910 in 2023; in 2024 and 2025, it achieved ROC-AUC values of 0.913 and 0.915, respectively. SHAP analysis identified gestational duration as the most influential predictor; its removal substantially reduced performance. The results demonstrate that MINSA records enable the development of interpretable and temporally stable models to support perinatal surveillance. © Los autores.

Instituciones: Universidad Nacional José Faustino Sánchez Carrión