Photonic Data Science

Raman spectroscopy is gaining increasing attention in biomedical diagnostics thanks to instrumental development and chemometric approaches enhancing the accuracy and speed of this technique. Meanwhile, it is demanding to construct a statistical model based on one dataset (primary conditions) and use it to predict another dataset measured under different (secondary) conditions. Thus, model transfer becomes extremely important to improve prediction with minimal or no training samples measured under secondary conditions. Methods that have been proposed and applied for near-infrared spectroscopy, for example, spectral standardization, lead to poor performance in Raman spectroscopy. This is because Raman bands are sharper and more sensitive to noise introduced by the spectral standardization. Our recently reported Tikhonov regularization based on a partial least squares regression (TR-PLSR) approached this problem. In the present work, we showed that the TR-PLSR model transfer also works for Raman spectra of vegetative bacteria. This was demonstrated by the Raman spectra of three species of bacteria acquired on three different Raman spectrometers. Afterward, we report two newly developed model transfer methods: movement of principal components scores (MS) and spectral augmentation (SA). Both methods were validated based on the Raman spectra of bacterial spores and vegetative bacteria, where a significant improvement of the model transferability was observed. The MS method yielded comparable results to the TR-PLSR. However, the new methods are superior to TR-PLSR in two ways: first, no training samples with the secondary conditions are necessary, and second, the methods are not restricted to PLSR but can also be applied to other models. Both advantages are important in real applications and are a big step to enhance the performance of model transfer.