Pupillometry is a promising technique for the potential diagnosis of several neurological pathologies. However, its potential is not fully explored yet, especially for prediction purposes and results interpretation. In this work, we analyzed 100 pupillometric curves obtained by 12 subjects, applying both advanced signal processing techniques and physics methods to extract typically collected features and newly proposed ones. We used machine learning techniques for the classification of Optic Neuritis (ON) vs. Healthy subjects, controlling for overfitting and ranking the features by random permutation, following their importance in prediction. All the extracted features, except one, turned out to have significant importance for prediction, with an average accuracy of 76%, showing the complexity of the processes involved in the pupillary light response. Furthermore, we provided a possible neurological interpretation of this new set of pupillometry features in relation to ON vs. Healthy classification.