Representation of Mass Spectra as Pattern Vectors

A great variety of methods have been used for the representation of mass spectra as pattern vectors, for preprocessing and feature selection. An evident method for the representation of a low resolution mass spectrum in a multidimensional space is to use directly the original mass spectrum as a vector. Peak heights x. at integral mass numbers i correspond to the vector components. Elimination of very small peaks (e.g. lower than 0.5 or 1 % of the base peak) is useful [1, 165, 1933. 

TABLE 10. Pattern recognition applications in mass spectrometry with specialized data sets. 

Cluster analysis and factor analysis of sulfur--containing compounds 

In order to reduce the dynamic range and to improve the classification result the square root or the logarithm of x are frequently used. New logarithmic features x. which lie in the intensity range 0 to 100 can be obtained by equation (148) C2503- 

From the information theory (Chapter 11.6.1) it is known that a signal contains most information if all possible values of the signal have equal probability. To generate pattern vectors with equally distributed features intensity levels must be created for each mass number. Each level should have the same probability in the spectral file. If there are k intensity levels and d mass numbers a set of (k-1)d threshold values must be stored for spectral preprocessing C250D. 

---