A Robust Ensemble Machine Learning Framework with Automated Hyperparameter Optimization for Childhood Stunting Prediction

Abstract

Childhood stunting remains a major public health problem because of its long-term impact on physical growth, cognitive development, and future quality of life. Although machine learning has been increasingly applied for stunting prediction, existing studies primarily focus on predictive accuracy, while model stability and robust classifier selection under repeated experimental conditions remain less explored. This study proposes an integrated machine learning framework combining Optuna-based hyperparameter optimization, repeated Monte Carlo evaluation, composite score-based model selection, weighted soft voting ensemble learning, and statistical validation. Seven heterogeneous classifiers were independently optimized and evaluated across 100 Monte Carlo iterations. Composite score analysis under multiple weighting scenarios, supported by standard deviation and 95% confidence interval estimation, was used to identify the most robust classifiers. The selected models were subsequently integrated into a weighted soft voting ensemble framework. Experimental results showed that the proposed ensemble achieved the highest predictive accuracy of 0.9040 with stable performance distribution (SD = 0.0056). Statistical validation further confirmed significant performance improvement compared with standalone classifiers. The proposed framework demonstrates a reliable and robust approach for machine learning-based childhood stunting prediction.