Use of Global Symmetries in Automated Signal Class Recognition by a Bayesian Method

Automated or semiautomated pattern recognition in multidimensional NMR spectroscopy is strongly hampered by the large number of noise and artifact peaks occurring under practical conditions. A general Bayesian method which is able to assign probabilities that observed peaks are members of given signal classes (e.g., the class of true resonance peaks or the class of noise and artifact peaks) was proposed previously. The discriminative power of this approach is dependent on the choice of the properties characterizing the peaks. The automated class recognition is improved by the addition of a nonlocal feature, the similarities of peak shapes in symmetry-related positions. It turns out that this additional property strongly decreases the overlap of the multivariate probability distributions for true signals and noise and hence largely increases the discrimination of true resonance peaks from noise and artifacts.