Model validation was performed using ∼25% of the samples as the evaluation set. Recognition ability was calculated as the percentage of members of the calibration set that were correctly classified, and prediction ability was calculated as the percentage of members of the validation set that were correctly classified. LDA models were constructed employing different numbers of variables (wavenumbers), starting with the entire spectrum and decreasing the number of variables. It was observed that

model recognition ability varied significantly with the number of variables, with the best correlations R428 nmr being provided by eight-variable models. In general the models were satisfactory (average recognition and prediction abilities above 75%) as long as the selected wavenumbers presented high loading values. Therefore, the following wavenumbers, that have been previously reported in other FTIR studies on coffee, were selected for the final models: 2924, 2852, 1743, 1541, 1377, 1076, 910 and 816 cm−1, with possible association to caffeine, carboxylic acids, lipids, chlorogenic acids, trigonelline and carbohydrates. The score plots for the first three discriminant functions are shown in Fig. 4. The first three discriminant functions

accounted for 96.2, 95.2, 95.3 and 97.6% of of the total sample variance, for the models based Selleckchem PR171 on raw spectra, media-centered spectra, normalized spectra and first derivatives, respectively. A clear separation of all groups (non-defective, black, immature, dark sour and light sour) can be observed for the models based on DR spectra (see Figs 4a–c), whereas some level of group overlapping was observed for the model based on spectra derivatives (Fig. 4d). The calculated

values of each discriminant function at the group centroids are displayed in Table 1. It is interesting to point out that, for all the developed models, the first three discriminant functions are enough to provide Ixazomib mouse sample classification. For example, considering the model based on the raw spectra, it can be observed that non-defective coffees present positive values for DF1 and DF2 and negative values for DF3, whereas black beans present negative values for DF1, DF2 and DF3. The corresponding values obtained for correct classification rates for each specific model and group are shown in Table 2. Recognition and prediction abilities were quite similar for all the developed models. The data were further evaluated in order to develop a more generic classification model, i.e., only one discrimination function that would provide discrimination between non-defective and defective beans, without separating the defects into specific groups. The classification functions and respective correct classification rates are shown in Table 3. Respective average values of recognition and prediction abilities were 96.4 and 100%, for the model based on raw spectra, 97.