The main objective of this study is to present a comparison of the different architectures applicable to classifying levels of attention with a database of eye trajectories. The classification was performed using Machine Learning (ML) models such as Logistic Regression (LR) and Ridge Classifier (RC), both used in supervised classification tasks due to the ease of development. Likewise, convolutional neural networks (CNN) such as VGG16, Resnet, and ConvNext were used, as well as a visual transformer that focused on using attention mechanisms. Emphasis is placed on key performance metrics (Accuracy, Precision, F1, etc.) to analyze the behavior of the models, demonstrating that traditional models when they execute simple tasks still have a good performance against advanced models that despite giving very similar results, such as a 95% mean precision with the ML models to conventional CNNs with a 97% mean precision, have the disadvantage of computational expense and execution time, on the other hand, ML models despite being easy to implement models showed that in classification tasks they cannot compete against specific models to work with images.