Optuna-optimized stacking models for predicting total phenolic content in NFC compound juices based on routine physicochemical attributes: Interpretability analysis via SHAP.

The Folin-Ciocalteu method for total phenolic content (TPC) in juices is time- consuming and generates chemical waste. Against this context, this study developed an interpretable data-driven framework based on 6 routine physicochemical attributes (RPAs). To this end, color attributes (L⁎, a⁎, b⁎), soluble solids content (SSC), titratable acidity (TA), pH and TPC of 282 apple, mango, orange, and pear NFC compound juices were analyzed. Afterwards, six regression base-models were optimized using the Optuna platform and integrated through a stacking ensemble strategy to capture nonlinear dependencies and feature interactions among RPAs. The resulting stacking model achieved high accuracy, with Rp2 of 0.956, RMSEP of 0.058 mg GAE/mL, and RPD of 4.825. Shapley additive explanation (SHAP) analysis revealed b⁎, a⁎ and pH as dominant predictors. Overall, this study demonstrates that TPC of NFC juices can be accurately predicted using a small number of routine parameters, providing a cost-effective, sustainable, and industrially applicable solution.