Challenges and Strategies for Implementing Automated Visual Inspection for Biopharmaceuticals

This article is part of PharmTech's supplement "Injectable Drug Delivery."

Manufacture of sterile parenteral drug products involves a series of unit operations (1) and aseptic processing conducted under strict requirements with respect to product quality. The manufacturing process is designed and validated to address such requirements and to ensure supply of safe and efficacious products. Visually inspecting each filled and sealed container for foreign contaminants or particulates ensures that these high standards are met and the final drug product is safe for patient use (2).

The United States Pharmacopeia (USP) provides guidance with respect to the inspection process for injectable drug products (2, 3). According to USP General Chapter <1>, the injection process shall be designed and qualified to ensure that every lot of all parenteral preparations is essentially free from visible particulates and every container whose contents shows evidence of visible particulates shall be rejected. Two methods are primarily employed by the pharmaceutical industry to address the need for visual inspection of filled and sealed containers: manual visual inspection (MVI) relying on human capability and machine based automated visual inspection (AVI).

Benefits of automated visual inspection

As the name suggests, a manual inspection relies on the ability of human operators to detect foreign contaminants in the filled containers. The inspection requires trained and certified inspectors to perform the task. Use of inspection aids such as contrasting colors and magnifying glass can improve the accuracy of human inspection. In spite of this, the subjectivity involved with manual inspection impacts effectiveness and the speed with which the inspection can be done. In addition, the process cannot be validated. Achieving required inspection throughput for a large commercial lot would require larger number of inspectors, which can add to labor costs.

Automated inspection systems, on the other hand, rely on a machine to detect visible particulates. Compared with manual inspection, an AVI process is more consistent and can be more cost effective over a longer time period of use. The AVI system requires qualification and validation, which ensure that the performance is consistent and similar to or better than human inspection. Several comprehensive studies of Knapp and coworkers [4, 5] highlight the probabilistic nature of the inspection process and provide a mathematical framework for comparing the performance of an automated inspection system with human capability.

Automated inspection process: technologies and principles

The automated inspection machine (AIM) used in this study contains a light-transmission double-check system for detecting particles in filled and sealed drug-product containers. The AIM uses a static division (SD) system that divides the photo detector into independent bits that span a detection window from the base of the container to just below the meniscus. The first step in the inspection process is the spinning of the container at a specified speed. As the vial spins, the liquid inside the vial forms a vortex and, because of the centrifugal force, imparts momentum to insoluble particles. These suspended particles are forced toward the container wall. The vial is then stopped with precise timing through the application of brakes on the machine. Because of frictional drag, the vortex collapses, thereby lifting and rotating the suspended particles. The image of moving particles is projected onto the SD sensor and can be sensed through variation in the intensity of the transmitted light which is converted to an electric signal from the affected bits. The amount of change in the electric signal is proportional to the size of the particle and is compared with a preset sensitivity level. If the signal exceeds the threshold established by the preset sensitivity level, the vial is deemed faulty by the machine and is sent to the defect bin. Cosmetic defects such as scratches or stains on the vial surface do not result in any movement during inspection and are not detected by the SD sensor.

Industry also uses a camera-based system to detect defects in filled drug-product containers. Unlike the SD sensor, which relies on light transmittance, a camera system uses light reflection to detect particles. Because the judgment of the camera system depends on the intensity of the reflected light, its performance is dependent on particle reflectivity and color. In addition to moving particles, a camera-based system can also pick up the light reflected from surface scratches and other container defects. Depending on the sensitivity of the system, this can result in increased false rejects. Alternatively, the system can be calibrated to detect specified cosmetic defects.

In addition to cosmetic defects, the potential benefit of a camera-based system includes the improved performance at lower fill levels. For very low fill volumes, the inspection window for a SD sensor-based system is greatly reduced, thereby resulting in deteriorated performance. Such challenges can be addressed by strategic placement of cameras to target a low fill-volume window. Hybrid systems that seek to combine the benefit of camera-based and SD sensor-based technology are being developed to provide improved performance for both particle and cosmetic defects.

Irrespective of the technology selected for automated inspection, several operational parameters (e.g., machine settings) and product properties play a key role in determining the performance of the system. Detailed characterization and optimization of these parameters is critical to developing an AVI process. Each technology needs to be qualified for its ability to detect faulty containers and to ensure that non-defect containers will not be rejected. This qualification requires a series of experiments using standard defect sets to challenge the AIM. Careful selection of experimental conditions (i.e., defect sets and machine settings) is important to minimize the number of evaluations and still generate conclusive data for entire process space. This study uses a SD sensor-based AIM to evaluate the effect of key process parameters on machine performance for inspection of liquid products in vials. Tested parameters include machine settings, formulation properties, and fill configuration.

Methods and materials

Eisai Automated Inspection Machine (Model 587-2, Eisai Machinery of USA, Allendale, NJ) was used to conduct an automated inspection. Mimic solutions were prepared by adding an estimated amount of PEG 1000 into desired buffer solution to reach a target viscosity value. Upon proper mixing, a sample was tested for viscosity confirmation. The mimic solution was then filtered through a 0.22μm filter and aspectically filled into vials followed by manual inspection to ensure absence of foreign particles. These clean vials were spiked with standard glass beads of three different sizes: 70μm, 100μm and 400μm (Duke Scientific Soda Lime Glass, Palo Alto, CA). Each vial was seeded with a single glass bead. The seeded vials were manually inspected a second time to ensure that each vial only had one particle and no other foreign contaminants. Two different buffer formulations (with and without polysorbate) were used to prepare PEG solutions of different viscosities. Formulations comprised of a commonly used stabilizing excipient (sucrose or sorbitol) and a buffering agent (acetate). The formulations are listed as follows and are referred to accordingly in the remaining sections of this article:

Formulation A = PEG 1000 in acetate buffer + sucrose + Polysorbate 20

Formulation B = PEG 1000 in acetate buffer + sorbitol

Experimental design. Vials seeded with particles were inspected on the Eisai machine through a double-check system comprising two inspection stations. A set of 24 vials (six clean and six each seeded with 70μm, 100μm and 400μm particles) were inspected at each inspection station. Each vial was inspected 32 times in a single run. A total of three runs were conducted to estimate statistical error in the performance. For selected experiments, a design of experiment (DOE) was performed using a space-filling design and spanning a range of spin and brake settings for a fill configuration of 1.7 mL in 3 cc. All DOE experiments were conducted for solutions with different viscosities using PEG solution in Formulation B.

Data analysis. Inspection data obtained from the Eisai AIM were analyzed by calculating the detection rate (%DR) as follows:

The detection rate can then be calcuated for each vial type (clean, 70μm, 100μm and 400μm) by averaging the %DR values for all six vials in each group. For runs that were conducted in triplicates, the average and standard deviation were also computed for the detection rate for each particle size. For the data generated through DOEs, contour plots were generated to depict detection rates for 400 μm-particle size at various spin and brake settings.

Results and discussion

Systematic studies were conducted to evaluate the effect of key process parameters, including machine settings, product properties, and fill configuration on the performance of the AVI system.

Role of machine parameters. The key operating parameters for the AVI process included the machine settings used during inspection (see Table I).

Table I: List of machine parameters potentially impacting detection rates.

Impact of machine sensitivity.The mechanism of operation of a SD sensor-based system involved sensing the variations in the electric signal (voltage level) as a result of light interference by foreign particles. Sensitivity of the machine refers to the threshold DC voltage signal that should be exceeded to judge the vial as faulty. A higher sensitivity would improve the machine's ability to detect foreign particles, especially those of smaller sizes as shown in Figure 2. However, there is a threshold sensitivity above which the improved detection rates come at the cost of false rejection. Even clean vials with no foreign contaminants were judged by the machine as faulty (see Figure 2). Careful selection of the sensitivity levels should be made to maximize the detection rates while minimizing the business costs associated with rejection of nondefective vials.

Figure 1: Mechanism of visible particle detection through a static-division sensor-based AVI system. (ALL FIGURES ARE COURTESY OF THE AUTHORS)

Impact of spin speed and brake position. The key requirement for tuning the ability of the machine to detect particles was to adjust the balance between the vial rotation and the precise timing and position when the vial was stopped and inspected. In an optimal set-up, the liquid surface (see Figure 1) would have been completely restored right before its inspection. If the rotation of the vial is terminated too late (i.e., represented by larger brake setting here) the liquid level may not be restored at the time of inspection and the moving meniscus can send the faulty signal to the detector resulting in a false reject. On the other hand, if the rotation is terminated too early, the liquid would have slowed down and the foreign particle would have started sinking even before reaching the inspection station. This slowdown could result in the machine missing a defective vial and wrongly classifying it as an "accept." Figure 3 shows that the detection rates for all particle sizes were improved when the brake settings were increased from 7 to 9, signifying a better performance when inspection is conducted closer to the termination of rotation. Similar improvement in performance was observed when the spin speed was increased from 1600 rpm to 2200 rpm. Higher rpm rates seemed to impart larger momentum to the liquid and suspended particles and thereby keeps them moving for a longer time, making them easier to detect. The magnitude of this effect could depend on product property (e.g., viscosity) and fill configuration (e.g., container size and fill volume) as discussed in the following sections.

Figure 2: Comparison of detection rates for particles of different sizes when using two different levels of sensitivity (Formulation A). The impact of sensitivity on detection rate for smaller particles is larger and higher sensitivity may result in false rejects (i.e., rejection of clean vials).

Role of product properties. In addition to the machine settings, the product properties can have a significant impact on the performance of the AVI system. Solution properties such as density, viscosity, and surface tension govern the movement of the foreign particle in the flow-field generated by spinning the the vial. The speed at which meniscus recovers, as well as the time it takes for the foreign particle to descend and stop after the application of brakes, is dependent on these solution properties. Figure 4 shows deterioration in performance of the AVI system as the product viscosity is increased. As solution becomes viscous, the particle motion relative to the solution is arrested and it becomes difficult for the machine to detect. Detection rates can improve by increasing the spin speed and brake settings as shown in the figure by the three traces of color (blue, red, and green). The Figure 4 inset shows a closeup of the data set for 2.3 cP. It was observed that the two formulations with same viscosity but different density and surface tensions (Formulation A is represented in yellow and has a density of 1.046 g/mL and surface tension of 48 mN/m2; Formulation B is represented in red and has density of 1.033 g/mL and surface tension of 61 mN/m2 at room temperature) exhibit slightly different detection rates (about 7% variation). This finding demonstrates that variation in physical properties other than viscosity can effect the way the particles are suspended and move during inspection, thereby affecting their detection rates. However, as the spin speed was increased, the overall performance improved and the differences between the detections rates for the two formulations was reduced.

Figure 3: Effect of spin speed and brake settings on the detection rate of the machine. A setting of 1600 x 7 represents spin speed of 1600 rpm and brake setting of 7. Performance improves as spin speed is increased and inspection is performed quickly after stopping the vial.

In addition to the physical properties discussed above, other inherent properties of the protein solution could affect the ability of the machine to differentiate between and true and false rejects. If the protein has a propensity to form or trap particulates, such protein particles can be perceived by the AVI system as rejects. In such cases, kinetics of particulate formation should be characterized to assess the feasibility of using automated inspection. The liquid formulation may also have propensity to form micro air bubbles which can be perceived by the inspection system as foreign particulates. Beccause air bubbles have a tendency to rise to the meniscus, inspection view height can be carefully selected to avoid any interference with the bubbles. Some AIMs use a pre-spin to facilitate removal of air bubbles before the inspection spin. Such AVI process issues could be very product specific and may cause costly delays during performance qualification. Selection of an appropriate mimic solution for development runs is therefore critical to identify and trouble-shoot such problems early during product development.

Figure 4: Product properties can have a significant impact on machine performance. Detection rates (average of 100 μm and 400 μm particles) are reduced at higher viscosity and improve with increased speed and brake settings. Solutions with similar viscosity can also exhibit differences in detection rates based due to differences in density and surface tension. A setting of 1600 x 7 represents spin speed of 1600 rpm and brake setting of 7. Yellow fill color represents Formulation A while red represents Formulation B.

A design of experiments can be conducted to characterize the performance of the AVI system over a wide range of operational parameters. A design space can then be created over the range that gives acceptable performance. Such design space characterization offers the assurance of a consistent and robust process. Figure 5 shows results of a DOE study conducted over a wide range of spin speed and brake settings for a fill volume of 1.7 ml in a 3cc vial using Formulation B. The contour colors represent detection rates measured by the automated inspection system for solutions of different viscosities. As discussed above the performance clearly deteriorates with increase in solution viscosity. While higher spin speed and higher brake settings results in improved detection rates, other operational issues can come into play under such conditions. For example, very high spin speed may cause the vials to shoot out of the spindles making the process operationally unfriendly. Very high brake settings (inspecting very close to the termination of vial rotation) may not provide adequate time for the meniscus to recover. The meniscus, in turn, puts a shadow on the sensor and can be wrongly classified as a defect. All these operational issues should be carefully characterized to create a design space that offers acceptable performance during commercial manufacturing.

Figure 5: Contour plots representing the machine performance over a wide range of spin and brake settings for solutions of different viscosities. The black area represents the operational parameter range that was not studied.

Role of fill configuration. In addition to machine settings and formulation properties, fill configuration of the final drug product presentation (i.e., size and shape of container and liquid fill volume) plays a significant role in determining the performance of an automated inspection system. The radius of the container has a direct impact on the shape of the vortex formed when the vial is spun and the recovery of the meniscus when brakes are applied. Syringe barrels usually have smaller radii than vials and pose a bigger challenge for inspection. Higher spin speeds are needed for syringes to obtain performance comparable with vials. The height of liquid level plays an equally important role as well. For a given vial size, as fill volume is reduced, the liquid level is lowered and the size of the inspection window (i.e., distance between base and meniscus level) shrinks. Figure 6 demonstrates that the automated machine's performance was consistently better for larger fill volumes for each of the three vial sizes. Inspecting very close to the meniscus level may result in false rejects due to the meniscus shadow being perceived by the sensor as foreign particle. Careful selection of inspection view height is critical to minimizing such false rejects while maximizing the size of the inspection window.

Figure 6: Comparison of detection rates for vials of different sizes and fill volumes. Smaller fill volumes are consistently more challenging for automated inspection.

Conclusions

AVI of injectable drug products offers several advantages over manual inspection, including process consistency, speed, and potential cost effectiveness. This study investigated the role of machine settings, formulation properties, and fill configuration on the performance of an AVI. Higher spin speed and brake settings were shown to improve detection rates of the AVI system. Product properties such as viscosity, density, and surface tension affect the manner and duration of particle suspension in solution and thereby affect process performance. Other inherent solution properties such propensity to form air-bubbles and/or protein particles can also cause potential interference with the inspection system resulting in false rejections. Low fill volumes are also challenging because of the smaller inspection window. It is suggested that any equipment qualification or process characterization work should evaluate the system performance over a wide range of these process parameters and solution properties to arrive at a robust and consistent visual inspection process. DOEs can be conducted to study these parameters and any potential interactions. Formulation properties and fill configurations can be bracketed to minimize the number of experiments.

Acknowledgments

The authors wish to thank Aarti Gidh, Deborah Shnek, Erwin Freund, and Ed Walls in process development at Amgen for useful discussions and suggestions for this paper. We also thank Jeff Stephens, Ari Levy, and Damien Villanueva in clinical manufacturing at Amgen for providing valuable experimental support toward the execution of these studies.

Nitin Rathore*, is a senior scientist, Cylia Chen is a senior associate scientist, Oscar Gonzalez is a senior engineer, and Wenchang Ji is principal scientist, all in drug product and device development at Amgen, Thousand Oaks, CA, nrathore@amgen.com, tel. 805.313.6393.

*To whom all correspondence should be addressed.

References

1. N. Rathore and R. Rajan, Biotechnol. Prog., 24 (3), 504–514 (2008).

2. T.A. Barber, Control of Particulate Matter Contamination in Healthcare Manufacturing (CRC Press, 1999).

3. C. Jones, presentation before the PDA Visual Inspection Forum (Bethesda, MD, 2007).

4. J.Z. Knapp and L.R. Abramson, Jrnl. of Parenteral Sci. and Technol., 44 (2), 74–107 (1990).

5. J.Z. Knapp, PDA Jrnl. of Pharma. Sci. and Technol., 75 (2), 131–147 (2007).