Klasifikasi Status Gizi Balita Berdasarkan Data Antropometri Menggunakan Algoritma Light Gradient Boosting Machine

Abstract

Toddlers are in their golden age, where malnutrition at this stage is irreversible and can affect health in adulthood. Therefore, monitoring nutritional status is very important. According to the 2025 Jember Regency Data Dashboard, 18,157 toddlers were recorded as stunted, representing 10.28%, 19,839 toddlers were recorded as wasted, representing 11.23%, and 31,050 toddlers were recorded as underweight, representing 17.58%, indicating that nutritional problems among toddlers remain a significant concern in this region. This study aims to develop a classification model for toddler nutritional status based on anthropometric data (gender, age, weight, and height) using the Light Gradient Boosting Machine (LightGBM) algorithm with Bayesian Optimization for hyperparameter tuning, and to evaluate the effect of applying the Synthetic Minority Oversampling Technique (SMOTE) in handling class imbalance. This study also implements the best model into a website-based system using the Flask framework. The dataset used is 131,591 toddler records from the Jember Regency Health Office for the period of December 2025. The preprocessing stages include column selection, handling of missing values (1,097 records removed), removal of duplicate data (11,943 records), outlier handling using the IQR method (1,026 records removed), categorical encoding, and age conversion to months, resulting in a final dataset of 117,525 records split into 80% training and 20% testing data using stratified sampling. Two approaches were compared: LightGBM with Bayesian Optimization (LGBM-BO) and LightGBM with Bayesian Optimization and SMOTE (LGBM-BO-S). The evaluation results showed that the LGBM-BO model provided the best performance with an accuracy of 95.20%, a precision of 95.10%, a recall of 95.20%, and an F1-score of 95.14%. Although SMOTE improved recognition of minority classes, it caused a decrease in overall performance. An overfitting analysis further confirmed that LGBM-BO generalized better, with a training-testing accuracy gap of only 1.30% compared to 2.31% for LGBM-BO-S. The best model was successfully implemented into a website-based system that allows parents and health workers to easily and quickly determine the nutritional status of toddlers. Black box testing showed that all system functions ran as expected, and the system was further validated by a public health center nutritionist, who confirmed that it was easy to use and that the classification results were clearly interpretable.