Abstract
This study assessed the effectiveness of various multivariate calibration models in predicting the sensory evaluation scores of specialty coffee produced in Panama. The predictions were based on seven key physicochemical parameters of the beverage, considering the processing method used (natural or washed). To construct the models, three algorithms, Multiple Linear Regression (MLR), Principal Component Regression (PCR), and Partial Least Squares Regression (PLSR), were employed, analyzing data sets for natural, washed, and combined processing methods. Model quality was evaluated using metrics such as the coefficient of determination (R(2)), root-mean-square error (RMSE) for cross-validation and prediction, and the residual predictive deviation (RPD). Among the physicochemical parameters, titratable acidity, soluble solids, and protein content showed a positive correlation with sensory scores, whereas pH exhibited an inverse relationship. The best-performing MLR and PCR models were those for the natural process, achieving R(2)p, RMSEp, and RPD values of 0.8293, 0.4239, and 2.34 for MLR and 0.7233, 0.5322, and 1.86 for PCR, respectively. Across all algorithms, models built exclusively with data from a single processing method consistently outperformed those that combined samples from both processes. PLSR models further demonstrated this trend, with R(2)p values of 0.7639 and 0.8306, RMSEp of 0.6891 and 0.3948, and RPD values of 2.07 and 2.51 for the washed and natural processes, respectively. In conclusion, the study highlights the critical importance of considering processing methods when developing multivariate models to predict the sensory evaluation scores of specialty coffee. Models built with samples from a uniform processing method yielded significantly better performance than those developed using mixed-process data sets.