Interpretable One-Class Classification of Raman Spectra Using Prediction Bands Estimated by Wavelet Regression

Area of Science:

• Spectroscopy
• Chemometrics
• Analytical Chemistry

Background:

• Traditional one-class classification (OCC) methods struggle with noisy spectral data.
• Previous work applied band-based OCC to Near-Infrared (NIR) spectra using Principal Component (PC) analysis.
• Raman spectra present unique challenges due to their sharper features and higher noise levels compared to NIR spectra.

Purpose of the Study:

• To develop and validate a novel band-based one-class classifier specifically for Raman spectra.
• To adapt and improve upon existing OCC methodologies for handling the characteristics of Raman spectral data.
• To enhance the interpretability of spectral classification by documenting deviations in the wavelength space.

Main Methods:

• Developed a hybrid wavelet and Principal Component (wPC) expansion for spectral transformation and denoising.
• Employed a multinormal prediction model to forecast future wPC scores of unseen spectra.
• Back-transformed predicted wPC scores to generate spectral trajectories in the wavelength space, defining the acceptance band.

Main Results:

• The proposed wPC-based band-classification method successfully classified the first derivatives of pharmaceutical Raman spectra.
• The methodology effectively captures sharp spectral features and provides efficient denoising capabilities.
• Deviations from critical trajectories were documented in the wavelength space, improving classification interpretability.

Conclusions:

• The novel band-based OCC classifier is effective for analyzing Raman spectra.
• The integration of wavelets and PCs offers a robust approach for spectral denoising and feature extraction.
• This method provides an interpretable framework for identifying target chemicals and understanding spectral variations.