Details

Autor(en) / Beteiligte

• Guo, Shuxia
• Heinke, Ralf
• Stöckel, Stephan
• Rösch, Petra
• Popp, Jürgen
• Bocklitz, Thomas

Titel

Model transfer for Raman‐spectroscopy‐based bacterial classification

Ist Teil von

• Journal of Raman spectroscopy, 2018-04, Vol.49 (4), p.627-637

Ort / Verlag

Bognor Regis: Wiley Subscription Services, Inc

Erscheinungsjahr

2018

Quelle

Wiley Online Library - AutoHoldings Journals

Beschreibungen/Notizen

• Raman spectroscopy has gained increasing attention in biomedical diagnostics thanks to instrumental developments and chemometric models that enhance the accuracy and speed of this technique. In particular, a model transfer procedure is needed if the chemometric models are utilized to predict a new dataset measured under (secondary) conditions different to the training data (primary). The model transfer methods try to achieve satisfactory prediction on the secondary dataset with minimal or no training samples measured under secondary conditions. Model transfer methods that have been reported are mostly applied for near‐infrared spectroscopy and in regression problems. The investigation of model transfer in Raman spectroscopy and classification is rare. Our recently reported Tikhonov regularization based on partial least squares regression (TR‐PLSR) was utilized for model transfer of Raman‐based classification models for spore species. In the present work, we show that the TR‐PLSR also works for Raman spectra of vegetative bacteria, even though the Raman spectra of 3 species of bacteria were acquired on 3 different Raman spectrometers. Additionally, we report 2 newly developed model transfer methods for Raman spectra: movement of principal components scores and spectral augmentation. 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 movement of principal components scores method yielded results comparable with those of the TR‐PLSR. However, the new methods are superior to TR‐PLSR in 2 ways: No training samples in the secondary conditions are necessary, and the methods are not restricted to partial least squares regression but can also be applied to other models. Both advantages are important in real‐world applications and represent a large step for improving the model transfer of Raman spectra. Two model transfer approaches are reported, which update the primary model with secondary data: the movement of principal component scores and spectral augmentation. These two methods significantly improve the model transferability between different devices, which was benchmarked by Raman spectral datasets measured on multiple devices. Because the label information of the secondary data is not needed, the methods are suitable for biological diagnostics.