In this doctoral dissertation the problem of pain level recognition based on facial images is considered. As it represents a complex problem from the domain of pattern recognition, various machine learning approaches are commonly utilised in an attempt to solve it. Considering the fact that the classification procedure involves several steps, as a part of this dissertation several ideas are proposed that can be implemented in individual steps, thusly improving the pain level recognition. A feature extraction procedure which is based on the shearlet transform, a method for feature selection through individual contribution tracking and an algorithm for radial basis network classification model design that is based on the artificial bee colony algorithm are presented and implemented. Each of these proposals has been experimentally evaluated on data coming from a clinical environment, and its behaviour compared with alternative approaches employed in the literature for the same tasks. The results of the conducted analysis suggest that the feature extraction procedure leads to better pain level recognition using several classifiers, while achieving the aforementioned using a comparingly lower number of features. Further more, the method for feature selection provides feature subsets that are comparingly of lower dimensionality while maintaining or improving the classification quality. Finally, the enhanced algorithm proposed for the radial basis network design outputs networks that give favourable results regarding classification performance when compared with the networks designed by other algorithms utilised in the experimental analysis.