Machine Learning-Enabled NIR Spectroscopy in Assessing Powder Blend Uniformity: Clear-Up Disparities and Biases...

Area of Science:

• Pharmaceutical Sciences
• Analytical Chemistry
• Spectroscopy

Background:

• Near-Infrared (NIR) spectroscopy is a valuable non-destructive technique for assessing blend uniformity (BU) in pharmaceutical powder blends.
• NIR spectra often contain overlapping physical and chemical information, complicating the accurate assessment of blend uniformity.
• Distinguishing chemical information from physical effects is crucial for reliable BU analysis.

Purpose of the Study:

• To develop and validate a machine learning-integrated workflow for NIR spectral analysis to improve blend uniformity assessment.
• To identify sources of variability impacting BU results and mitigate bias in NIR spectroscopic measurements.
• To enhance the accuracy and reliability of NIR-based BU assessment for active pharmaceutical ingredients (APIs) in powder blends.

Main Methods:

• Preparation of calibration samples of amlodipine (API) with varying concentrations using a gravimetric approach.
• NIR spectroscopic analysis of samples, followed by High-Performance Liquid Chromatography (HPLC) for validation.
• Application of data quality metrics (DQM) and bias-variance decomposition (BVD) to investigate and overcome bias.
• Implementation of clustered regression (non-parametric and linear) and various cross-validation techniques (hold-out, k-fold, bootstrapping).

Main Results:

• Established NIR-based blend homogeneity with a low mean absolute error (MAE) of 0.674 ± 0.218 w/w.
• Bootstrapping-based cross-validation demonstrated MAE of ±3.5% w/w for model generalizability and ±1.5% w/w for model transferability.
• Successfully implemented a workflow integrating machine learning with NIR spectral analysis to improve BU assessment accuracy.

Conclusions:

• The developed workflow effectively deconvolutes physical and chemical information in NIR spectra for enhanced BU assessment.
• Machine learning and advanced data analytics significantly improve the accuracy and reliability of NIR-based pharmaceutical blend uniformity analysis.
• The study provides a robust methodology for lifecycle management and transferability of NIR methods in pharmaceutical quality control.