Augmentation of vibrational spectroscopic datasets using local covariance-based sampling in latent space to make prediction models that are more tolerant to spectral variations caused by the physical properties of samples

Abstract

An augmentation strategy for vibrational spectroscopic datasets using local covariance-based sampling in latent space is investigated to build a prediction model that is more tolerant to the spectral variation of samples caused by their physical properties. The strategy is based on the expansion of an original training dataset by adding newly generated spectra that go beyond the boundary of the original domain and the use of the expanded dataset for modeling. In this way, the generated spectra in the expanded domain emulate variations in the spectra caused by the physical properties of the sample. A data augmentation method that simultaneously leverages the capabilities of the Synthetic Minority Oversampling Technique (SMOTE) and the Mixup, called Local Covariance-based Augmentation (LoCA), is developed. To evaluate the utility of LoCA, Raman spectra of paracetamol tablets with four different packing densities and near-infrared (NIR) spectra of bovine serum albumin (BSA) powder samples with three different particle sizes are employed. The incorporation of LoCA-generated spectra for training is effective in building models for predicting the paracetamol and BSA concentrations that are more tolerant to variations in the spectra induced by differences in packing density and particle size, respectively. In overall, LoCA combining the local geometry-awareness of SMOTE and the input–output joint interpolation of Mixup for the augmentation in a latent space is beneficial to secure accuracy for vibrational spectroscopic analysis of solid samples under variation of their physical presentations.