Novel Adaptive Feature-Penalized Ridge Logistic Regression Model: A Supervised Machine Learning Advancements over Traditional Ridge Logistic Regression for Composite Malnutrition Diagnosis Among Under-Five Children in Mozambique

In Mozambique, child malnutrition is not only a pressing public health issue but also one of the most complex, multi-faceted, persistent problems, where stunting, underweight, and wasting continue to defy numerous efforts. Efforts to understand and tackle the issues of malnutrition have been hindered by traditional statistical techniques such as logistic regression due to multicollinearity and interdependence of these outcomes situated within an intricate causal framework. We present a new statistical technique: Adaptive Feature-Penalized Ridge Logistic Regression (AFPR-RM), which models a composite malnutrition outcome by capturing stunting (chronic), underweight (intermediate), and wasting (acute) into a singular term for children below 5 years of age. Using the 2022 Mozambique Demographic and Health Survey (DHS), we extract 3,953 observations. We present Ridge Logistic Regression (RLR) model which overcomes the inadequacy demonstrated by classical logistic regression via multicollinearity. AFPR-RM is advanced by incorporating a dual-adaptive penalty structure: one driven by the data-derived signal strength and the other, feature scaling determined from domain-specific considerations. Together, these create a multiplicative penalty framework, which we deem a novel contribution to the literature on penalised regression. AFPR-RM outperforms standard Ridge Logistic Regression in classification accuracy, model interpretability, and diagnostic stability, as shown by empirical comparison. This study aims to assess the classification, accuracy, calibration and diagnostics of the AFPR-RM against the baseline Ridge Logistic Regression model and determine the best model for guiding public health interventions on child malnutrition across its indicators.