PCA-Enhanced Machine Learning Framework for Child Stunting Prediction Using Household and Socioeconomic Factors

Abstract

Childhood stunting continues to pose a major public health concern because its underlying determinants arise from complex household and socioeconomic interactions that are difficult to capture using conventional analytical approaches. Although machine-learning techniques have shown considerable potential for health prediction, limited attention has been given to understanding how different levels of dimensionality reduction influence classifier performance when analysing high-dimensional survey data. Addressing this gap, this study developed a Principal Component Analysis (PCA)-enhanced machine-learning framework for childhood stunting prediction using secondary data from the 2023 Indonesian Ministry of Health survey in Riau Province. Following preprocessing, 2,976 valid observations with 16 predictor variables were transformed into 117 numerical features, after which PCA generated three feature representations retaining 89.30%, 94.28%, and 98.44% of the total variance. Twelve supervised machine-learning algorithms were subsequently evaluated using precision, recall, F1-score, and area under the receiver operating characteristic curve (AUC). The empirical results demonstrated that preserving a greater proportion of variance improved predictive performance across most classifiers. Among all evaluated models, K-Nearest Neighbours combined with 45 principal components achieved the strongest overall performance, yielding a precision of 0.710, recall of 0.771, F1-score of 0.739, and AUC of 0.809. These findings provide empirical evidence that integrating PCA with machine-learning algorithms offers a reproducible and computationally efficient framework for supporting evidence-based nutritional surveillance and advancing data-driven childhood stunting prediction.