Dimensionality reduction and main component extraction of mass spectrometry cancer data

Mass spectrometry data have high dimensionality. Dimensionality reduction is a very important step to greatly improve the performance of distinguishing cancer tissue from normal tissue. In this study multilevel wavelet analysis is performed on high dimensional mass spectrometry data. A set of orthogonal wavelet basis of approximation coefficients is extracted to reduce dimensionality of mass spectra and represent main components of mass spectrometry data. The best level of wavelet decomposition of mass spectrometry data is selected based on energy distribution of approximation coefficients. Compared to traditional principal component analysis (PCA) method, which dependents on training samples to build feature space, our proposed method is using wavelet basis to extract main components of mass spectrometry, keeping local properties of data, and computing efficiently. Experiments are conducted on three datasets. The competitive performance is achieved compared to other methods of feature extraction and feature selection.