Storage-Time-Aware Chemometric Deep Learning Model for Predicting Phenolic Retention and Antioxidant Capacity in...

Kefir offers a suitable fermented dairy matrix for delivering plant-derived bioactives, yet maintaining and predicting phenolic compounds during refrigerated storage remains a technological challenge. This study developed a storage-time-aware chemometric deep learning model to estimate phenolic retention and antioxidant capacity in kefir enriched with free and alginate-encapsulated black grape seed extract. Five formulations were prepared, including control kefir, kefir containing 1% and 3% free extract, and kefir containing 1% and 3% alginate-encapsulated extract. Samples were stored at 4 °C and analyzed on Days 1, 7, and 14 for physicochemical, textural, color and bioactive properties. Total phenolic content and Trolox-equivalent antioxidant capacity were used as target responses, while ten quality descriptors were transformed into latent chemometric features and arranged as storage-time sequences. A convolutional recurrent deep learning architecture was then optimized using the crested porcupine optimizer (CPO) and compared with baseline and ablation models. Alginate encapsulation produced structurally intact beads with high encapsulation efficiency, supporting improved phenolic retention during storage. The proposed model provided the most accurate dual prediction of total phenolic content and antioxidant capacity. It outperformed non-optimized and partially ablated alternatives in terms of reducing error and achieving agreement between the observed and predicted values. Residual, normality and Bland-Altman analyses further supported the reliability of the prediction behavior. Overall, the results demonstrate that integrating alginate encapsulation with storage-time-aware chemometric deep learning provides a practical strategy for monitoring bioactive stability in functional kefir and offers a reproducible modelling framework for fermented dairy products enriched with phenolic-rich by-products.