Genetic support vector machines as powerful tools for the analysis of biomedical Raman spectra

The growing number of applications of Raman spectroscopy in medicine necessitates the development of robust and accurate processing methods. The two major tasks for which Raman spectra are used are quantifying chemical species in a sample (regression) and discriminating chemically distinct samples (classification). Conventionally, linear techniques, primarily projection to latent structures (PLS), are used to perform these tasks. However, there are a number of nonlinearities that may arise when acquiring the Raman spectra of biomedical samples, such as scattering differences between tissues or auto-fluorescence variances, which makes nonlinear methods more suitable. To this end, we compared kernelized support vector machines (SVM) to PLS for a number of biomedical Raman datasets. Additionally, this work develops a genetic SVM, wherein the parameters of a SVM are selected by a classical genetic algorithm instead of the conventional grid search. This facilitates the use of complex kernels, which yield higher performance than simple kernel functions. We have found that this genetic SVM outperforms PLS in all of the regression tasks examined in this paper, while yielding equivalent results for classification tasks. Additionally, we have found that the genetic algorithm provides significant time savings in the optimization of the SVM parameters over grid search.