*3.3. Whole 1H and 13C-NMR Modeling Using Feed-Forward Predictive Artificial Neural Networks*

The use of feed-forward predictive neural networks on the NMR data as a method for classification was assessed using three different forms of validation, namely 33.3% data holdback, CV-10 k-fold, and excluded row validation. Since ANNs are more powerful than any other classification method in terms of their flexibility and noise insensitivity, the algorithm was fitted on the training set using the entire NMR spectrum without any form of variable selection. Table 4a,b shows % accuracy and % predictability for the different forms of cross-validation carried out on different spectral pretreatments of 13C NMR and 1H NMR under the zg30 and NOESY NMR spectra. In general, contrary to what was observed in PLS-DA, it was observed that models obtained under 1H NMR models had higher rates of classification when compared to 13C NMR. Similarly, to what was observed in PLS-DA, raw data derived from the NOESY experiment had a higher model performance throughout the three different validation methods used when compared to the zg30 experiment. With respect to validation it was observed that, irrelevant to the spectrum used, the 33% holdback cross-validation resulted in overfitted models which were identified as spectral transformations that had very good training models but failed to predict new samples with the exception of MSC. In general, the best performing cross-validation method was the excluded row validation followed by k-fold validation. This could be attributed to the fact that these cross-validation methods are not completely random as the 33% holdback. In the case of the excluded row validation the samples were selected in such a way that, the groups contained approximately equal amounts of local and foreign EVOOs in the training stage. Thus, the models obtained where equally capable of recognising and predicting local and foreign EVOOs.


**Table 4.** Application of artificial neural network (ANN) on the (**a**) 1H NMR and (**b**) 13C NMR data using three forms of cross-validation.


**Table 4.** *Cont.*
