**4. Discussion**

#### *4.1. Conclusions for the Tanh Activation Function*

We found that the ANN showed the most optimal behavior under the tanh activation function for the training stage. The reference value was 0.879 for the precision test that varied the training epoch and hidden size parameters. Precision and cost behaviors were as expected, considering that the cost decreased and the precision increased for all the evaluations proposed under different parameters. Another relevant conclusion is that, according to the ROC analysis, the classes that are least likely to be identified under these ANN parameters are classes 2 and 6.

#### *4.2. Conclusions About the Dataset*

Regarding the dataset, we can conclude that, for future work, it is advisable to consider more CTX-M contigs. In this study, the 10 most representative groups were considered, ye<sup>t</sup> some of the groups were not representative enough to be able to carry out a stratified cross validation. This was particularly true for the experimentation in the validation stage, in which 20% of the initial dataset was used for this validation. Regarding the dataset, we can conclude that more CTX-M contigs should be considered for future studies.
