Addressing Class Imbalance in Low Birth Weight Prediction: A Quantile-Based Outcome Redefinition Framework Using...

Low birth weight (LBW), defined by the World Health Organization as birth weight below 2500 g, remains a major public health concern. This study evaluates whether a quantile-based outcome redefinition can improve predictive severity stratification among newborns already classified as LBW. We analyzed deidentified SIVIGILA surveillance records from Medellín, Colombia (2011-2021; n = 10,185), all corresponding to confirmed LBW births. The analysis therefore does not model LBW versus normal birth weight. Instead, the binary outcome was redefined within the LBW cohort using the first quartile of birth weight (Q1 = 2240 g), distinguishing a Q1-defined severe LBW stratum from a moderate LBW stratum (2240-2499 g). The WHO 2500 g threshold is retained as the clinical and public-health definition of LBW; the Q1 threshold is used only for risk stratification, surveillance, and predictive modeling. Models based on logistic regression, class weighting, SMOTE, and Elastic Net were compared using discrimination, classification, and calibration metrics. The proposed outcome redefinition strategy led to higher sensitivity (0.218 to 0.632), F1 score (0.330 to 0.496), and AUC (0.624 to 0.726) compared with baseline models. A sensitivity-specificity tradeoff analysis across classification thresholds is also presented to support practical implementation in epidemiological surveillance contexts. Overall, the findings demonstrate that redefining the outcome variable using distribution-based thresholds can be an effective and interpretable strategy for mitigating class imbalance in rare-event epidemiological data, providing a practical framework for improving predictive modeling in maternal and neonatal health surveillance systems.