Avoiding Overestimation and the ‘Black Box’ Problem in Biofluids Multivariate Analysis by Raman Spectroscopy: Interpretation and Transparency With the SP-LIME Algorithm

Abstract

Raman spectroscopy, in combination with multivariate analysis, is a powerful analytical tool for solving regression and classification problems in various fields—from materials science to clinical practice. However, in practical applications, experimental studies and the implementation of Raman spectroscopy present numerous challenges, including multicollinearity in spectral data and the ‘black box’ problem of complex analytical models. To avoid these problems, the proposed classification and regression models require proper interpretation. This study makes use of a comparative analysis of explanation methods based on the SP-LIME (local interpretable model-agnostic explanations with submodular pick) algorithm of a bilinear model (projection onto latent structures [PLS]) and a nonlinear model (one-dimensional convolutional neural network [CNN]). The models to be interpreted are trained to solve the regression task of the blood serum Raman characteristics and the urea levels. Effective SP-LIME evaluation of the blood Raman spectra revealed that in urea analysis for both PLS and CNN models, the important band is at 1003 cm⁻¹. This approach is based on the value of the root mean square error estimation only when a single Raman band is analyzed. The aim of this paper is to develop an approach to explain the operation of the analytical models and provides the way to reveal the exact Raman bands with the biggest impact on the model performance.