Machine Learning-Based Stunting Classification Using Puskesmas Nutritional Data: A Comparative Study of Six Algorithms with Hyperparameter Tuning

Abstract

Stunting detection at Indonesia's community health centers (Puskesmas) still relies heavily on manual anthropometric assessment a process that is slow, inconsistent, and difficult to scale. This study applies machine learning to automate that classification using 40,071 real-world records from the Jeneponto Regency Health Department (2021–2024), covering three nutritional categories: Normal, Stunting, and Severe, derived from Height-for-Age Z-Scores per WHO standards. Six classifiers were compared: Logistic Regression (LR), Decision Tree (DT), Random Forest (RF), Support Vector Machine (SVM), K-Nearest Neighbors (KNN), and Gradient Boosting (GB). Each was tested in both default and tuned configurations, with hyperparameters optimised via GridSearchCV and stratified k-fold cross-validation. Preprocessing included invalid value removal, target leakage prevention, label encoding, StandardScaler normalisation, and stratified 80:20 splitting. Performance was measured using weighted Accuracy, Precision, Recall, and F1-Score. At baseline, RF led with 97.04% accuracy and 97.05% F1-Score, while LR trailed at 77.35% a gap that points to a non-linear classification boundary that linear models cannot handle. After tuning, GB came out ahead with 97.54% accuracy and F1-Score, overtaking RF at 97.22%. That reversal matters: it shows that GB's sequential architecture holds more capacity, but only delivers it when key parameters are properly configured. Tuned GB shows real promise as a decision-support tool for Puskesmas-level stunting screening at scale.