Development of hybrid stacking machine learning for evaluating parameters affecting refrigerated shrimp coated with chitosan-loaded Salvia officinalis nanoemulsions

• The effect of chitosan coating incorporated with Salvia officinalis L. essential oil nano-emulsion (1 and 3 %) on the refrigerated shrimp’s quality was investigated. • A hybrid stacking ensemble method was proposed to assess treatment performance over time. • Chitosan coating loaded with 3 % nano-emulsion effectively inhibited lipid oxidation. • Coated shrimps had better microbial quality and higher sensorial scores. • The hybrid model consistently achieved high R² values and the individual models exhibited signs of overfitting. This study investigated the effects of chitosan coatings containing Salvia officinalis L. essential oil nanoemulsion (NEO) at concentrations of 1 % and 3 % NEO (Ch-NEO-1 and Ch-NEO-3) on refrigerated shrimp quality for 12 days. A hybrid stacking ensemble method was proposed to assess treatment performance over time. It aggregates a support vector machine (SVM), extreme gradient boosting (XGBoost), and random forest (RF) as a meta-learner. The droplet size of the EO nanoemulsion was approximately 156 nm, and FTIR spectroscopy confirmed the encapsulation of the EO. The results showed that coating shrimp with Ch-NEO-3 positively inhibited lipid oxidation. The chemical results showed that the pH of all groups increased during storage, and the pH of the control group reached 11.14 ± 0.13, while it was 8.67±0.91 for the Ch-NEO-3 treated group. Moreover, PV and TBARS values of Ch-NEO-1 (14.51±0.02 meq O2 /kg lipid and 0.59± 0.01 mg MDA/kg) and Ch-NEO-3 (14.32±0.04 and 0.56± 0.01 mg MDA/kg) treated groups were significantly lower than those of control (1.41± 0.09 mg MDA/kg) after 12 days ( p < 0.05). In addition, active Ch-NEO-3 coating showed antibacterial effects and lowered TVC, LAB and Enterobacteriaceae to 6.38± 0.11 log CFU/g, 2.06± 0.03 log CFU/g, and 2.47±0.12 log CFU/g, respectively, which was in line with the TVBN results. Shrimp coated with Ch and Ch-NEOs also have higher sensor scores than the control. The hybrid model consistently achieved high R² values, such as R 2 -train=0.986 and R 2 -test=0.986 for pH and R 2 -train=0.958 and R 2 -test=0.997 for overall acceptability, while maintaining low mean absolute error (MAE) values, notably 0.105 for pH and 0.138 for overall acceptability. In contrast, the individual models exhibited signs of overfitting; for example, for TABRS, XGBoost achieved an impressive R² of 0.998 in training but decreased to 0.695 in testing, and RF showed a significant gap between its training (R² = 0.926) and testing (R² = 0.595) performances. Similarly, SVR struggled with overfitting, yielding R² values of 0.924 for training and 0.830 for testing in the color assessment. The hybrid model's ability to maintain closely aligned R² values for training and testing across various properties underscores its robustness and reliability, highlighting its effectiveness over the standalone models.