Pixel-based identification of Raman hyperspectral data: application to pharmaceutical tablets impurities detection

In the pharmaceutical field, analysis of tablets by Raman hyperspectral imaging is widely
used for quality control purpose and has been now included in the general chapters of the European
Pharmacopeia. However, data obtained can be consequent to analyze and implies to use
appropriated chemometric tools. Most of the time, factorial decomposition methods (i.e
Multivariate Curve Resolution – Alternating Least Squares (MCR-ALS)) can be applied, excepted
for the analysis of big data matrices, as well as in the presence of many constituents. Moreover,
even when the composition is known, the MCR resolution can be challenging because some low
variances sources can be diluted in the process of unmixing and can hardly be resolved unless
information on the expected sample composition. Moreover, it can exist minor compounds
presents in a few pixels which can be missed in the MCR process. To bypass these limitations, one
possibility can be to step back to the analysis of individual pixels, which somehow would be the
most efficient method for database matching. The objective of the present study is to develop a
pixel-based identification (PBI) approach to elucidate chemical composition of Raman
hyperspectral images of complex pharmaceutical formulations. The proposed approach relies on the
identification of Essential Spectral pixels (ESP).
The proposed study was evaluated on both known and unknown tablets composition. The known
formulations were made of polymorphic forms of carbamazepine (case 1) and piroxicam (case 2) to
mimic minor compounds (from 0.1%w/w to 5%w/w). The seven unknown samples were falsified
chloroquine (case 3) which were seized during the COVID-19 pandemic [3]. Raman hyperspectral
imaging analyses of samples were performed with a LabRAM HR Evolution (Horiba scientific)
equipped with an EMCCD detector (1600 × 200-pixels sensor) (Andor Technology Ltd.), a Leica
50x Fluotar LWD objective and a 785 nm laser with a power of 45mW at sample (XTRA II single
frequency diode laser, Toptica Photonics AG). For both case 1 and 3, the whole tablet surface was
analyzed with a 150 x 150 mapping and a step size of 87µm (total map size of ~13 x 13 mm²). For
case 2, the middle of the tablet surface was mapped with a step size of 5.5 μm over a 5.5 x 5.5 mm², providing a 1000 x 1000 mapping. Three different approaches were then evaluated on each map to select pixels: (i) Kennard-Stone randomized, (ii) randomized selection and (iii) ESP selection, which were subsequently matched with the in-house database by using correlation coefficient (CC).
For case 1 and case 2, the ESP approach compared to the other pixel-selection algorithms has
shown the best results in terms of correlation coefficient but also with the smallest analysis time,
with 50 seconds for the classical data size and 2 minutes for the big map size. The ESP approach
was thus applied on falsified medicines and enabled to get the entire sample composition even for
complex formulation (from 4 to 9 chemical compounds) with correlation coefficient superior to
0.80. After gathering the ESP, a classical least squares was applied and allowed to show that even
chemical information localised in a unique pixel had been elucidated, as it can be seen in Figure 1.
The proposed study highlighted the potential of the PBI approach for chemical identification
purposes. It has been shown that, for known samples, both tiny and huge amount of data can be
analyzed without the need of the entire map, by selecting only a few percentages of pixels (~8% of the initial data). The proposed methodology allowed to keep even the chemical information which
was not in the in-house database which is very interesting in case of falsified medicines purposes.
The global conclusion of this study is about the potential applicability of the methodology to other
hyperspectral imaging techniques or matrices. Indeed, thanks to the inherent properties of the
essential spectral pixel algorithm, the only requirement for PBI is to have at least one pure pixel by
component. In case of mixed spectra, the use of the ESP could be a pre-processing step like, to
reduce data dimensionality, which has been successfully demonstrated.