Analisis Performa Jaringan Saraf Tiruan Backpropagation Menggunakan Optimizer SGD, RMSProp dan Adam untuk Klasifikasi Stunting pada Balita

Abstract

Stunting is a linear growth disorder in toddlers caused by chronic malnutrition during critical developmental periods and remains a significant national health problem in Indonesia. According to the 2022 Indonesian Nutritional Status Survey (SSGI), national stunting prevalence reached 21.6%, surpassing the World Health Organization (WHO) recommended ceiling of 20%. This study evaluates three optimizers—Stochastic Gradient Descent (SGD), RMSProp, and Adam—within a Backpropagation Artificial Neural Network (ANN) for classifying toddler stunting status. The dataset comprises 1,454 anthropometric records (sex, age, and height) collected from UPT Puskesmas Kampung Baru, Medan City, covering 2020–2023. Preprocessing included MinMax scaling (fitted exclusively on training data to prevent leakage) and SMOTETomek resampling to address severe class imbalance. Forty-eight model configurations were evaluated via grid search across a 3-32-1 architecture (ReLU hidden layer, sigmoid output). Evaluation metrics comprised accuracy, precision, recall, F1-score, specificity, AUC, and Matthews correlation coefficient (MCC). RMSProp achieved perfect scores on all metrics (1.000) in the shortest execution time (133.42 s). Adam achieved equivalent classification performance with similarly rapid convergence (135.28 s). SGD attained perfect recall (1.000) only when training data were simultaneously balanced and normalized, empirically demonstrating that non-adaptive optimizers require both interventions to compete with adaptive counterparts.