Imaging the Blending Process
Hyperspectral imaging can be used to optimize blending, by monitoring the distribution of excipients and APIs in formulation.
Blending, one of the most basic of pharmaceutical unit operations, can also be one of the most challenging to control. Solid formulations contain multiple ingredients beyond the active pharmaceutical ingredients: fillers, tabletting agents, disintegrants, and absorption enhancers or agents that slow down and control absorption. Ingredients from different vendors may behave differently due to their particle size and shape and other factors, and their tendency to form aggregates.
Materials must be chosen to ensure the desired flow characteristics, potency, proper dissolution profile, and absorption of specific formulations. Proper particle size grades of the ingredients must be selected to produce an optimum blend for capsule filling.
There are strong economic drivers for optimizing blending. Reinforcing these is the pharmaceutical quality by design (QbD) framework advanced by FDA, which requires a deeper understanding of pharmaceutical manufacturing processes, how ingredients blend and how blending progresses through different stages.
imMix System
For these studies we used the inMix system, which consists of a push-broom SWIR HS camera (Specim Ltd., Oulu Finland), which is positioned to view the blend inside the rotating blender
The blender is rotated by a computer-controlled motor. For the collection of HS data on all or certain specified rotations, the blender is slowed down and the camera is programmed to scan the blend covering the window. HS data collected through the optical window on the blender is turned into composition maps (indicating spatial dispersion), which are used to predict ingredients throughout the blending process. The very large amount of imaging data is condensed to a limited number of useful micromixing parameters.
The camera is protected in a stainless steel housing; the motor, power supply, and camera controller and other electronics are housed in another stainless steel module. The instrument is compatible with different sizes and different types and sizes of blenders, as the camera position is adjustable. To empty the blender, the front housing including the camera module can be easily slid out from under the blender. A white reference for background measurements and devices to establish proper focus and help measure pure components or smaller quantities of materials can be attached to the blender.
The SWIR (Short Wave Infrared) HS camera, with a wavelength range of 1000-2500 nm, is equipped with an OLES Macro SWIR HS lens (Specim Ltd., Oulu Finland), viewing approximately a 1 cm line with 30 µm optical resolution. A stationary directing mirror is used to direct the reflected light 90 degrees back to the camera lens. The blender with a window is illuminated with two quartz halogen lamps which are protected by another glass window. The height of the camera and optical parts can be adjusted. This is necessary to bring the blend into sharp focus, as well as adjusting for the differences among the different types and sizes of blenders used.
The blender, shown in Figure, is a one liter IBC (intermediate bulk container) blender, with filling and emptying ports, and whose emptying port is equipped with a removable 1-inch diameter sapphire observation window. The blender is attached to a rotating shaft, which is rotated by a computer-controlled motor. For frequent filling and emptying of the blender, the camera and optical parts of the blend monitor are mounted on a base plate, attached to a slide mechanism to move the entire front housing out of the way of the emptying port.
The data collection software allows the user to view live camera data, store white and dark references, and adjust camera settings such as exposure time and frame rate. The blend speed and measurement resolution can be selected, and blend rotations, where the image should be measured, can be specified.
Analysis of Hyperspectral Data
The analysis software allows the prediction of the composition of any of the ingredients for any of the blender rotations using the Science-Based Calibration (SBC) [7] or partial least squares (PLS) methods. The SBC method requires the input of the pure analytes’ spectra which can be collected with the system or imported as a single spectrum obtained from other instruments.
The large amount of HS data is automatically collected, sequentially arranged by blender rotation, and analyzed. The composition maps provide the first level of data compression, resulting in the images of the predicted compositions of each ingredient. In subsequent analysis, the images are compressed into parameters that are meaningful for the blending process and displayed as a function of the rotation of the blender.
Various image analyses can be performed with the prediction images, including standard statistical measures such as image average, standard deviation, relative standard deviation, and the fraction of pixels above/below/within a certain threshold, and spatial uniformity measures such as the distribution of aggregate sizes. The analysis software allows the prediction images for selected rotations and components to be viewed, compared, and saved.
In one blending experiment, 20% acetaminophen was blended with 39% methyl cellulose, 39% lactose, and 2% magnesium stearate. The blending was monitored at every rotation of the blender up to 200 rotations. It can be observed on Figure 5 that lactose is evenly blended by about 30 rotations of the blender, and that there is some evidence of re-agglomeration in the 80-90 rotation range. Even though there is some statistical variability from turn to turn, in this experiment, lactose seems to break up again over the consequent hundred rotations.
From the various image-processing options, the fraction of pixels that are within range of the nominal composition is shown in Figure, ranging from zero to one.
Each pixel is much smaller than the unit dose and is even smaller than the usual aggregate sizes, thus it is a good metric to show the progression of the blending. Of the main ingredients, cellulose and lactose reach their best uniformity at around 25 turns, whereas the acetaminophen breaks up more slowly, improving until about 160 turns in this experiment. These differences would not have been revealed using single-point near-infrared monitoring, which obtains one average spectrum for each rotation of the blender. It can also be observed that in this example the cellulose and lactose are both slowly getting less homogeneously blended, although these changes probably do not affect the quality of the blend as much as the API getting more uniformly distributed.
In another experiment, a blend containing 50% acetylsalicylic acid, 19% methyl cellulose, and 29% lactose was blended for 100 turns. The blender was stopped and 2% magnesium stearate was added and the blending continued for another 100 turns. In Figure, it is seen that the addition of magnesium stearate successfully broke up the remaining aggregates, significantly lowering the median size of lactose aggregates.
A HSI camera system proved useful in monitoring the distribution and aggregate sizes of APIs and excipients in solid pharmaceutical formulations. It was also able to measure changes in the different ingredients in a 1-L test blender over the time of the blending. The results are useful for pharmaceutical formulation development for troubleshooting blending problems, qualifying the blending behavior of excipients with different compositions or different characteristics and from different sources.