Classification of magnetic resonance spectra using parallel randomized feature selection

Feature selection is a useful preprocessing strategy when dealing with the classification and interpretation of high-dimensional biomedical data, especially when the sample size is small. A classification technique, exploiting parallelization efficiencies, is presented where a set of multi-layer perceptrons are trained on randomly selected feature subsets with varying cardinality. This technique is tested using high-dimensional biomedical spectra acquired from a magnetic resonance spectrometer. The classification results are benchmarked against a conventional multi-layer perceptron architecture as well as linear discriminant analysis. The new technique had a significantly lower classification error than either of the benchmarks.