The identification of spectral signatures is crucial for the classification/profiling of biomedical spectra. Because only limited number of biomedical samples of high dimensionality is typically available, dimensionality reduction techniques (identification of discriminatory features) are essential for robust classifier development. We show, on three real-world biomedical datasets, the potential of a consensus-based identification of important feature subsets, using a genetic algorithm and a sparse linear classifier. When training data are in short supply, the proposed methodology leads to more stable subset identification and higher classification accuracy.