Feature selection and classification of high-resolution NMR spectra in the complex wavelet transform domain

Successful identification of the important metabolite features in high-resolution nuclear magnetic resonance (NMR) spectra is a crucial task for the discovery of biomarkers that have the potential for early diagnosis of disease and subsequent monitoring of its progression. Although a number of traditional features extraction/selection methods are available, most of them have been conducted in the original frequency domain and disregarded the fact that an NMR spectrum comprises a number of local bumps and peaks with different scales. In the present study a complex wavelet transform that can handle multiscale information efficiently and has an energy shift-insensitive property is proposed as a method to improve feature extraction and classification in NMR spectra. Furthermore, a multiple testing procedure based on a false discovery rate (FDR) was used to identify important metabolite features in the complex wavelet domain. Experimental results with real NMR spectra showed that classification models constructed with the complex wavelet coefficients selected by the FDR-based procedure yield lower rates of misclassification than models constructed with original features and conventional wavelet coefficients.