Researchers combined terahertz (THz) spectroscopy with chemometric and machine learning algorithms to identify and quantify active pharmaceutical ingredients (APIs) in multicomponent pharmaceutical samples.
This research Analysis methodreported a prediction accuracy of up to R² = 0.999 and a reduction in prediction time of as much as 90%.
Wang Xiaoning Others. Principal component analysis (PCA) was used in parallel with a machine learning model to interpret the THz spectral data. The researchers reported that the resulting approach was able to perform both qualitative identification and quantitative content analysis while assessing robustness using pharmaceutical systems containing multiple components.
Terahertz spectroscopy measures the interaction between electromagnetic radiation and a sample in the terahertz frequency range. This technique has the potential to distinguish between APIs and excipients without destroying the sample, as different molecular structures and crystal forms can produce characteristic spectral responses.
Viewed through the lens of process analytical technology (PAT), this combination could provide a path to rapid at-line, or potentially in-line, measurements of API concentrations. Faster analysis could allow manufacturers to identify variations during production instead of waiting for traditional laboratory test results.
This combination could provide a path to rapid at-line, or potentially in-line, measurements of API concentrations. ”
The use of machine learning is important because drug spectra can contain overlapping signals resulting from the physical properties of the API, excipients, and sample. Algorithms trained to recognize these complex relationships have the potential to extract quantitative information that is difficult to obtain with traditional spectral interpretation.
However, the reported performance represents an experimental proof of concept rather than a validated drug manufacturing method. Adoption of GMP will require testing across independent batches, equipment, manufacturing sites, and anticipated sources of process variation. Manufacturers must also demonstrate model lifecycle management, data integrity, and continued suitability of calibration datasets.
Changing the raw materials, formulation, or processing conditions can affect the spectral response and therefore model performance. These risks must be addressed through prospective validation and continuous model monitoring.
Nevertheless, our results demonstrate that combining fast nondestructive spectroscopy and machine learning has the potential to extend the capabilities of pharmaceutical PAT. If successfully implemented in manufacturing, this approach could support earlier process intervention, improved content uniformity, and more data-rich control strategies.
