ProSensus, Inc.

NIR imaging is a powerful method to capture chemical information. This study, performed for a major pharmaceutical company, aimed to:

• predict tablet hardness
• predict average concentration of certain APIs (active pharmaceutical ingredients)
• estimate the distribution of these APIs in the tablet

 

NIR images

A Spectral Dimensions imaging system was used to capture spectra in the 1250nm - 2250nm range, in steps of 10nm. This leads to a 3-dimensional data matrix covering 101 wavelength bands.
Left a grayscale NIR tablet image at the 2090nm wavelength [black dots are dead pixels] and right the spectra, across all wavelengths, at two locations on the tablet.

Analysis of results

The NIR data cube is pre-processed to remove bad pixel responses and to maximize subsequent information extraction.

Predicting hardness

The processed NIR data was used to predict tablet hardness. It is interesting that the spectra, which are affected by compression of the tablet material, are able to successfully predict the hardness with less than 10% error.

Predicting average API

The average API concentration was also successfully predicted, with errors no greater than 5%.

Predicting the average API concentration

Assessing API distribution

The API distribution is a rapid way to measure tablet production quality; decisions affecting the process design and its control can be made from these predictions.

Several methods are available for predicting multiple APIs from the pure component spectra. A common method is to use a weighted sum approach, with linear regression to estimate the coefficients. This is inefficient, considering that the pure component spectra are so highly correlated; and it may lead to spurious predictions.

ProSensus has developed an approach, which we call un-mixing the spectra, to avoid problems related to correlated spectra. Our models successfully predict the API distribution within a tablet, and these models are easily re-trained when new APIs are investigated.