How AI can keep pharma safe

Andreas Gross - global product manager of Inspection Technology at Syntegon Technology, explains how artificial intelligence (AI) is being used to ensure pharmaceutical products are safe.

Whether in automated driving, computer games or face recognition – artificial intelligence (AI) is transforming our everyday lives. The required software and algorithms are already used in many domains, including visual inspection of manufactured components. So why not apply them to an area where safety plays a crucial role: the production and inspection of drugs for human use.

One of the main challenges in introducing AI consists in transferring applications to complex processes and developing suitable implementation and validation concepts for the strictly regulated pharmaceutical industry. Inspection is a very challenging process. This is especially true for products with difficult characteristics, such as highly viscous parenteral solutions. Those cases usually require long development and optimisation times for vision algorithms before achieving a balanced operational level of detection and false reject rates. AI has the potential of shortening this development period and optimising the desired results more quickly.

Though there are many successful automated visual inspection techniques on the market, in some cases the combination of dense solutions with small containers do not promote the movement of particles, which reduces detection probability. Moreover, agglomerations or other types of inherent morphological features that are similar to particles, as well as bubbles resembling glass particles can cause false rejection of good containers. Especially for high-cost products, every single false reject is one too many.

The one main difference in pharma

While many drug producers and machine manufacturers are considering the use of AI and have issued first studies, reservations about implementation and validation are keeping most companies from using these applications in real production environments. In parallel, machine vision software companies are already offering deep learning vision tools. Hence, manufacturers of automated vision inspection machines must not necessarily develop their own algorithms or neural networks. In fact, existing solutions only require moderate software modifications. Additionally, an upgrade of the vision computers with higher processing power can be realised with GPUs, which are widely available in the gaming industry.

Inspection technology experts can easily perform the required upgrades for visual inspection usage, as was discussed at the 2019 PDA Visual Inspection Forum in Washington D.C. However, there is one crucial point that must be considered to enable validation: in contrast to many other industries, the deep learning model for pharmaceutical use must be “frozen” once the development phase is finalised. It must be static and can no longer change to make it version-controlled for validation. A recent discussion paper published by the FDA about the regulatory framework for Software as a Medical Device (SaMD) provides a good reference for application in areas different from pharmaceutical production.

AI in visual inspection – far more than a vision

There is much more to AI in automated visual inspection than just potential. In fact, some pilot projects have already reached a stage close to production implementation. Syntegon Technology, formerly Bosch Packaging Technology, is currently working on a project to implement deep learning algorithms for the inspection of syringe stopper edges on its AIM 5023 inspection machine. The pharmaceutical industry is known for its conservative approach to innovation. This is mainly due to the very strict regulatory guidelines for process validation – overall a highly positive attribute since the manufactured products have a direct impact on the health and safety of patients.

“It takes a lot of courage and experience regarding software implementation and process validation to push the concept beyond the finish line,” as Dr Jose Zanardi, who is responsible for vision inspection development and applications at Syntegon Technology and closely involved in the project, likes to put it. Here, the company’s long-standing experience in the development of automated visual inspection machines since the early 1970s, combined with the profound expertise in processes, implementation and validation is an important prerequisite.

No “one size fits all” approach

Typically, a “one size fits all” approach will not work in deep learning projects for visual inspection. Instead, the first step should consist in a pre-assessment based on a large amount of images from reference samples. In our example this could be images of good units with bubbles, different stopper positions, products and fill volumes for body inspection, as well as different types of particles intrinsic to the process. Based on the available image data, offline verification studies provide the basis for the integration of deep learning models into the existing software. In the second step, a customer-specific project should be defined with parameters such as product, existing machinery, expectations and timeline.

The principle process does not change, and the recipe parameters are still validated according to GMP requirements. The only changes are the tool used to develop the process and the required hardware. As mentioned above, even the hardware only changes slightly: deep learning requires PCs with GPUs, which are able to process very complex and large amounts of data. In a deterministic deep learning model, small packages are trained up to a certain “level of intelligence” and then frozen. This is especially important regarding validation, regulatory approval and inspection.

A potential trendsetter for the pharmaceutical industry

“We believe that this technology has the potential to achieve detection rates close to 99% in the future while reducing false reject rates dramatically by half or more,” says Zanardi. He is confident that the deep learning application can be implemented in a GMP environment – and will obtain regulatory endorsement for both the qualification strategy and implementation. This will significantly improve the inspection of products that are difficult to inspect, such as lyophilised products or those filled into complex primary packaging such as syringes or double-chamber systems. This will reduce reject rates and subsequently costs in the production of expensive products such as orphan drugs.

USP chapter 1790 specifies that “validation of the automated inspection equipment should be based on comparison with the compendial manual inspection process with an expectation that alternative inspection methods demonstrate equivalent or better performance.” This is definitely true for the current deep learning project. Since the guidelines for implementation of automatic visual inspection machines are applied to AI and all its subsets, the pilot project stands a good chance of advancing into serial production – and of becoming a trendsetter for an entire industry.