Improving near-infrared prediction model robustness with support vector machine regression: a pharmaceutical tablet...

Area of Science:

• Chemometrics
• Spectroscopy
• Machine Learning

Background:

• Near-infrared (NIR) calibration models are sensitive to variations in raw materials and manufacturing processes.
• Lack of variability in calibration, test, and validation sets limits model performance and robustness.
• Nonlinearity, often caused by light path-length differences due to particle size or density variations, is a major interference in NIR spectroscopy.

Purpose of the Study:

• To evaluate the effectiveness of Support Vector Machine (SVM) regression in handling nonlinearity.
• To assess SVM's ability to enhance the robustness of NIR calibration models.
• To compare SVM regression with Partial Least Squares (PLS) regression in real-world manufacturing scenarios.

Main Methods:

• Support Vector Machine (SVM) regression was employed to model NIR spectral data.
• The performance of SVM regression was compared against Partial Least Squares (PLS) regression.
• Models were evaluated under conditions with variability not initially present in the calibration set.

Main Results:

• SVM regression demonstrated greater robustness compared to PLS regression when faced with physical (particle size) and chemical (moisture) variations.
• SVM regression showed improved linearity in predicted values.
• SVM models were less affected by un-modeled variability present in the test set.

Conclusions:

• SVM regression is a promising method for improving the robustness and accuracy of NIR calibration models, particularly in the presence of nonlinearity and process variations.
• While SVM offers advantages over PLS, further development of user-friendly tools is needed for widespread adoption in the pharmaceutical industry.