Machine learning-based prediction of antioxidant capacity in red wines during storage using phenolic compounds, color, and physicochemical profiles

Abstract

The aim of this study was to predict the antioxidant capacity of red wines produced from three grape varieties using machine learning models based on phenolic compounds, color characteristics, and key physicochemical parameters throughout aging, as determined by the DPPH method. The dataset consists of a total of 297 observations, with three replicates each for 99 samples obtained under different storage conditions. To evaluate model performance more realistically, a group-based cross-validation (GroupKFold) approach was used to prevent data leakage resulting from repeated measurements. Random Forest (RF), Extreme Gradient Boosting (XGBoost), and Support Vector Regression (SVR) were compared as machine learning models. According to cross-validation results, the XGBoost and RF models showed similar levels of prediction performance. XGBoost was selected as the final model due to its marginal advantage in terms of the coefficient of determination and its suitability for SHAP-based interpretability analysis. The SVR model, however, exhibited lower and less consistent performance. The SHAP analysis revealed that color intensity, pH, epicatechin, gallocatechin, and certain anthocyanin derivatives were the most significant predictors of DPPH activity. SHAP dependence analyses indicated that the effects of these variables on antioxidant capacity were nonlinear and interactive. The findings indicate that antioxidant capacity during storage should be evaluated not only in terms of time and temperature, but also in relation to changes in the phenolic profile and color characteristics. The models should currently be regarded as exploratory tools for preliminary estimation and decision support rather than as replacements for direct DPPH measurement. Future studies should evaluate their validity using larger, more diverse, and independent wine datasets and investigate whether additional biochemical and process-related predictors improve predictive performance and generalizability.