Comparação de modelos de regressão linear ponderados para avaliação de curvas de calibração de resíduos e contaminantes em alimentos

Abstract

This work aims to propose different weighting models for linear fitting adjustments using the Weighted Least Squares (WLS) methods for calibration curves employed in predicting concentrations of residues and contaminants in food within the scope of the Ministry of Agriculture and Livestock (MAPA) inspections, conducted at the Federal Laboratory of Agricultural Defense of Minas Gerais (LFDA-MG). The weighting models used were k/si², 1/x, 1/x², 1/y, and 1/y², in addition to the ordinary model without weighting. Besides the weighting models, this work also proposes evaluation metrics for the quality of the adjustments using weighted ANOVA, the ratio of predicted concentration to nominal concentration (ERx), the ratio of predicted instrumental response to nominal response (ERy), the correlation coefficient (R²), and the real correlation coefficient (R²real). The evaluation was divided into 4 groups of substances/analite based on their chemical nature and measurement procedures. The 4 groups are: Veterinary Drug Residues (VDR), Inorganic Contaminants (IC), Mycotoxins (MT), and Pesticides (PC). The results showed that in all 4 groups, there were curves that did not meet the proposed criteria in all the evaluated weighting models. Notably, mycotoxins showed approximately 85% of the curves outside the criterion in the ERx metric, and inorganic contaminants presented around 60% of the curves outside the criterion in the weighted ANOVA considering all the evaluated weighting models. Additionally, it was clearly observed that the mycotoxins group presented the highest differences between R² and R²real, indicating high variability in the instrumental responses, while the inorganic contaminants group showed good precision (low variability) but significant evidence of poor fitting in most of the evaluated curves.