Performance evaluation of preprocessing techniques utilizing expert information in multivariate calibration.

Area of Science:

• Analytical Chemistry
• Chemometrics
• Spectroscopy

Background:

• Partial Least Squares (PLS) regression is crucial for Near Infrared (NIR) spectroscopy data analysis.
• PLS model performance depends on calibration data representing analyte and interferent variations.
• Changes in interferent concentrations can degrade prediction accuracy if not accounted for during calibration.

Purpose of the Study:

• To compare various preprocessing techniques that incorporate expert knowledge for PLS models.
• To evaluate the effectiveness of expert knowledge-driven preprocessing against methods without it.
• To assess performance on datasets with differing interferent concentration ranges between calibration and test sets.

Main Methods:

• Comparison of multiple expert knowledge-based preprocessing techniques.
• Evaluation against standard preprocessing methods lacking expert knowledge.
• Application of techniques to two experimental datasets with distinct interferent profiles.

Main Results:

• Expert knowledge integration in preprocessing demonstrably improved prediction performance.
• Extended Multiplicative Signal Correction (EMSC) using pure component spectra yielded a ~32% prediction error reduction on dataset-1.
• Spectral Interferent Subtraction (SIS) utilizing interferent information achieved a ~63% prediction error reduction on dataset-2.

Conclusions:

• Preprocessing techniques leveraging expert knowledge enhance PLS model robustness and prediction accuracy in NIR spectroscopy.
• Specific methods like EMSC and SIS, when informed by expert knowledge of interferents, offer substantial improvements.
• This study validates the utility of incorporating prior chemical knowledge into spectral data preprocessing.