EEMizer: Automated modeling of fluorescence EEM data

Authors: Rasmus Bro a,*, Maider Vidal b
aDepartment of Food Science, Quality and Technology, Faculty of Life, Sciences, University of Copenhagen, Rolighedsvej 30, DK-1958, Frederiksberg C, Denmark
bDepartamento de Qu?mica Aplicada, Facultad de Qu?mica, Universidad, del Pa?s Vasco, Apdo. 1072,20080 San Sebasti?n, Spain
Summary
For PARAFAC modeling, there should be sufficient number of samples, sufficient resolution of excitation and emission, and that EEM data should be meaningful in context of Beers law. After fulfillment of these conditions, we decide for which wavelengths to include, how to handle Rayleigh scattering, number of components to use, and samples to exclude.
Algorithm EEMizer used in this paper is shown below
For FAC = 1 to maxFAC
    For SCATWIDTH = [5,10,15,25]
        For LOWEXREMOVE = 1 to maxValue
            Determine optimal model by removing outliers
        end
    end
end

Three different data sets have been used for EEMizer. First is called Dorrit data consisting of 27 samples, which are mixtures of hydroquinone, tryptophan, phenylalanine and dopa. Second data set consists of 65 samples, called DOM. Third data set consists of 338 samples, known as Fermentation data.
For Dorrit data, 4 components are adequate. Loadings of 5 component model are not sound. The wavenlengths removed for this data set are 15 as chosen by EEMizer. For DOM data, a 3 component model looks suitable, whereas, 6 component model is adequate in Fermentation data set. Twelve data set were removed as outliers as loadings look reasonable after removal of these outliers. EEMizer proved a good automation in modeling of EEMs in PARAFAC modeling.
The results for three data sets are shown below in figures 1, 2, and 3 respectively.
 
Figure 1. Factors in Dorrit Data
  
Figure 2. Factors and Loadings in DOM data
  
Figure 3. Loadings and Factors in Fermentation data

Reviewed by:     Aamir Alaud-din
            aamiralauddin@gist.ac.kr