Hydrogen-1 magnetic resonance spectroscopy (1H-MRS) allows noninvasive in vivo quantification of metabolite concentrations in brain tissue. In this work two of the most aggressive brain tumors are studied with the purpose of differentiating them. The challenging aspect in this task resides in that their radiological appearance is often similar, despite the fact that treatment of patients suffering these conditions is quite different. Efforts to differentiate between these two profiles are getting increasing attention, mainly because the consequences of performing an incorrect diagnosis. Due to the high dimensionality, initiatives oriented to reduce the description complexity become important. In this work we present a feature selection algorithm that generates relevant subsets of spectral frequencies. Experimental results deliver models that are both simple in terms of numbers of frequencies and show good generalization capabilities.