Photonic Data Science

The chemometric analysis of Raman spectra of biological materials is hampered by spectral variations due to the instrumental setup that overlay the subtle biological changes of interest. Thus, an established statistical model may fail when applied to Raman spectra of samples acquired with a different device. Therefore, model transfer strategies are essential. Herein we report a model transfer approach based on extended multiplicative signal correction (EMSC). As opposed to existing model transfer methods, the EMSC based approach does not require group information on the secondary data sets, thus no extra measurements are required. The proposed model-transfer approach is a preprocessing procedure and can be combined with any method for regression and classification. The performance of EMSC as a model transfer method was demonstrated with a data set of Raman spectra of three Bacillus bacteria spore species ( B. mycoides, B. subtilis, and B. thuringiensis), which were acquired on four Raman spectrometers. A three-group classification by partial least-squares discriminant analysis (PLS-DA) with leave-one-device-out external cross-validation (LODCV) was performed. The mean sensitivities of the prediction on the independent device were considerably improved by the EMSC method. Besides the mean sensitivity, the model transferability was additionally benchmarked by the newly defined numeric markers: (1) relative Pearson's correlation coefficient and (2) relative Fisher's discriminant ratio. We show that these markers have led to consistent conclusions compared to the mean sensitivity of the classification. The advantage of our defined markers is that the evaluation is more effective and objective, because it is independent of the classification models.