K-Means and Fully Connected Neural Network for Child Nutritional Status Classification

Abstract

Stunting remains a persistent child nutrition problem because delayed growth is closely related to long-term health, cognitive, and productivity risks. Manual interpretation of anthropometric measurements using the World Health Organization Z-score standard is clinically valid, yet it becomes inefficient and error-prone when routine records are processed in large numbers. This study develops a child nutritional status classification model by combining K-Means clustering and a fully connected neural network for early identification of stunting, underweight, and wasting. The dataset consisted of toddler anthropometric records from 2021-2024 with sex, age, body weight, and body height attributes. The data were cleaned, standardized, transformed into Z-score indicators, and grouped into 27 clusters representing possible combinations of nutritional status profiles. Cluster membership was then used with Zlen, Zwei, and Zwfl features in a multi-head fully connected neural network. Evaluation on 82 held-out samples showed accuracy values of 91.46% for stunting, 93.90% for underweight, and 98.78% for wasting. Weighted precision, recall, and F1-score were consistently high across the three outputs, while the training curves indicated stable learning without strong overfitting. The proposed hybrid model improves the reliability of child nutrition classification and can support a web-based decision support system for data-driven nutritional screening and intervention planning.

GS Cited Analysis