Handheld Raman material identity verification systems were used in this work (TruScan, Ahura Scientific, Wilmington, MA).
Six units manufactured at different times from different runs of parts were used in this study (details provided in subsequent
paragraphs in this article). This instrumentation uses a 785-nm NIR external cavity-stabilized, cooled laser as a light source.
The laser has a maximum power consumption of 1 W, a maximum output of 400 mW, and typically operates at 300 mW output. A single-dispersive
spectrometer, cooled CCD, and dielectric notch filters for Rayleigh rejection make up the rest of the instrument hardware.
The device also has an integrated barcode reader for capturing sample lot–batch–other identifiers and selecting methods. Detailed
specifications are listed in Table I.
Table I: TruScan instrument specifications.
Three TruScan devices (hereafter referred to as the "test devices") were tested with each of the 32 common pharmaceutical
materials listed in the sidebar "Materials tested" (Sigma-Aldrich, St. Louis, MO) to evaluate the applicability of the systems
for incoming inspection.
To create methods, reference spectra first had to be acquired. The reference spectra for the materials were taken on three
TruScan devices (hereafter referred to as "reference devices") using identical data collection software as the test devices
described above. In the context of TruScan, methods are analytical tasks based on stored reference spectra that the instrument's
software executes to determine whether a material's identity can be verified. For each material, a single reference spectrum
was acquired by one of the three reference devices.
To collect each reference spectrum, a sample of the material was placed in a borosilicate glass vial (VWR, West Chester, PA)
and the data were collected through the wall of the vial. The laser aperture of the instrument was placed at the appropriate
distance from the vial using either the nose-cone or vial-holder attachment for correct spacing. For most materials, the acquisition
process was initiated and simply allowed to continue until terminated automatically by the unit's library acquisition software.
In contrast, cellulose, dextrin, trimagnesium phosphate, zinc sulfate, and calcium sulfate are slightly fluorescent, so special
care had to be taken to avoid photobleaching the sample during reference measurement. This can be achieved by terminating
the reference scan before the software automatically discontinues measurement or by periodically moving the sample throughout
data collection. Finally, while reference spectra were collected, it was determined that colloidal silica, talc, sodium carboxy
methyl cellulose, and hydroxy propyl methyl cellulose did not provide an adequate signal to be measured in a practical period
of time for handheld deployment. In particular, colloidal silica had a very weak Raman signal, and the other three materials
were too fluorescent to allow reliable determinations to be made.
Methods were created from the reference spectra by the associated web-based software utility and then loaded into the test
instruments. To prepare test samples to challenge the methods, samples of approximately 2 g each of the 28 remaining materials
were sealed in 2-m-thick polyethylene bags in an effort to emulate expected-use scenarios (measurement through plastic bags)
for incoming material inspection. Three measurements of each sample were made, one measurement with each of the three test
devices. Each test measurement was made using the automatic (or "auto") mode, where the unit's software controls data acquisition
parameters to achieve the necessary spectral signal-to-noise ratio (SNR) in the shortest measurement time possible. The auto-mode
measurements used when running a method compensate for differences in operator positioning, stray-light, lot-to-lot sample
fluorescence, and so forth. Each "unknown" measured spectrum (three for each material) was evaluated against each method for
the test materials using the probability based approach described previously, resulting in a p-value for each unknown-method pair.