Photonic Data Science

One of the most important issues for the application of Ramanspectroscopy for biological diagnostics is how to deal efficiently with large datasets. The best solution is chemometrics, where statistical models are built based on a certain number of known samples and used topredict unknown datasets in future. However, the prediction may fail if the new datasets are measured under different conditions as those used for establishing the model. In this case, model transfer methods are required to obtain high prediction accuracy for both datasets. Knownmodel transfer methods, for instance standard calibration and trainingmodels with datasets measured under multiple conditions, do not providesatisfactory results. Therefore, we studied two approaches to improvemodel transfer: wavenumber adjustment by a genetic algorithm (GA) after the standard calibration and model updating based on the Tikhonov regularization (TR). We based our investigation on Raman spectra of three spore species measured on four spectrometers. The methods were tested regarding two aspects. First, the wavenumber alignment is checked by computing Euclidean distances between the mean Raman spectra from different devices. Second, we evaluated the model transferability by means of the accuracy of a three-class classification system. According to the results, the model transferability was significantly improved by the wavenumber adjustment, even though the Euclidean distances were almost the same compared with those after the standard calibration. Forthe TR2 method the model transferability was dramatically improved by updating current models with very few samples from the new datasets. This improvement was not significantly lowered even if no spectral standardization was implemented beforehand. Nevertheless, the modelt ransferability was enhanced by combining different model transform mechanisms.