<p>This paper presents a hybrid framework for predictive quality control in manufacturing, with a primary focus on addressing the lack of systematic feature selection in industrial data-driven applications. Validation was conducted through an industrial case study in plastic bottle production. The principal contribution is the integration of Response Surface Methodology (RSM) as a statistically rigorous feature selection mechanism, through which significant process variables and interactions are identified and subsequently used to inform machine learning (ML) model development. A comparative evaluation was performed using twelve months of real production data and several ML techniques, including Multi-Layer Perceptron (MLP), Random Forest, and XGBoost classifiers. Although all examined models exhibited satisfactory predictive capability, the Random Forest classifier achieved the highest performance, with an accuracy of 97.5% and an F1-score of 0.969. Following deployment of the proposed framework over a three-month period, the defect rate was reduced from a baseline value of 45% to below 5%, accompanied by a marked improvement in overall production efficiency. The results demonstrate that coupling statistically driven feature selection with ensemble-based learning yields a robust and transferable solution for predictive quality control in noisy, high-volume manufacturing environments. </p>