Discrepancies in approval documents are often accidental.

Osamu Hiruta, a specially appointed professor at the Joint Research Chair of Quality Assurance and Precision Control at Kumamoto Health Science University (at the time of the interview) and a member of the "Stakeholder Meeting on Stable Securing Measures for Medical Drugs," points out the current situation, saying, "Although there has been an improvement compared to a while ago, more than 10% of pharmaceuticals are still in a state of supply uncertainty, and among the various factors, many are due to discrepancies in approval documents. Discrepancies in approval documents are sometimes referred to as so-called 'illegal manufacturing,' but in reality, it is not only illegal acts such as intentional violations of approval requirements, but also that discrepancies often occur accidentally and lead to recalls."

Regarding the structural reasons for these accidental discrepancies, Mr. Hiruta explained, "The difficulty lies in matching the three documents: the approval document, the product standard, and the manufacturing instructions (actual). Each document has a different purpose, and the writing style, location, and level of detail also differ, making the process of linking and organizing the relevant sections complicated. In change management, updates accumulate, including changes that do not require pharmaceutical procedures, making it difficult to revise in a consistent manner, which leads to unintended inconsistencies. This problem was found very frequently in the mass inspections based on notifications from the government." He illustrated how the accumulation of small, everyday problems leads to discrepancies and increased workload.

Early detection of discrepancies is key; AI can improve efficiency.

To resolve discrepancies, simplifying the information required on approval documents is being discussed. However, this is a long-term approach involving changes to laws and regulations, making it crucial to create a system for early detection of discrepancies. Manual inspection work is burdensome and has its limitations, potentially leading to burnout on the factory floor and impacting production activities. This is where AI is attracting attention.

"By using an AI-powered system to visualize inter-document links and streamline the inspection process, we can quickly increase comprehensiveness even with limited human resources," says Hiruta, expressing his expectations.

One example of this AI-powered approach that has already yielded results is KM Bio's "Manufacturing and Marketing Approval Checker." Developed jointly by KM Bio and Macnica, this approval checker uses AI to automatically extract relevant sections from a vast amount of text, including product specifications and manufacturing instructions, and then cross-references them line by line.

Background of the Approval Document Checker Development

KM Bio, which focuses on human and animal vaccines, plasma-derived products, and neonatal mass screening, is working to ensure a stable supply by expanding globally and developing dual-use facilities. The company's decision to collaborate with Macnica to develop an approval document checker stemmed from the heavy burden of checking for discrepancies in approval documents.

Looking back at the history of the approval document discrepancy problem, in 2015, long-term fraudulent manufacturing was discovered at Kaketsuken (the predecessor of KM Bio), and the following year, in 2016, the government ordered an inspection of all prescription drugs in Japan. Discrepancies were found in about 70% of the products during this comprehensive inspection, and since then, thorough manufacturing in accordance with the approval documents and regular checks have been required.

"Learning from the Kaken incident, we have made strengthening quality, compliance, and governance our top priorities. Currently, we are moving from restoring trust to evolving our business foundation for sustainable growth, and we aim to be a company that can contribute to society in the long term through human resource development and capital investment," Hayashi said, describing the current situation. He also suggested that it was necessary to reduce the workload associated with inspections and improvements precisely because they continue to seriously address discrepancies even after restoring trust. Furthermore, Hatakeyama explained that the project addresses the issue of accidental discrepancies that can recur even after initial revisions due to insufficient impact assessment during document revisions, assumptions, and lack of information sharing. He stated, "The process of reviewing each item in the approval document and comparing it with the corresponding sections in product standards and manufacturing instructions is burdensome to perform alongside normal operations on the manufacturing floor. Inspections require know-how backed by manufacturing experience, making them highly dependent on individuals, and the psychological burden of not wanting discrepancies in the process one is responsible for also hinders proactive action. We started development with the expectation of reducing the workload and increasing efficiency, and by having AI assist in the line-by-line verification process that was previously done based on individual experience, we can convert the 'tacit knowledge' of individual experience into 'explicit knowledge' and share it with less experienced employees."

Man-hours to be reduced by 30-40%, inspections to leave "evidence."

During the actual development process, initially, adapting an existing QC check service using AI was considered, but the cost was too high to meet the required functions, so this was abandoned. Subsequently, while exploring the possibility of developing from scratch, the advancement of AI technology provided a boost, and Macnica Corporation, which had excellent UI development skills, proposed a joint development project. The biggest feature of the approval document checker is that it can be checked line by line using a highly accurate text matching AI, which is expected to reduce time, alleviate labor shortages, improve comprehensiveness, and reduce human error. In KM Bio's verification, the amount of work required for checking approval items in multiple projects for their flagship drug was reduced by approximately 35-38%.

"Previously, it was difficult to keep track of who made which decision and why, and the basis for the decision was sometimes ambiguous when trying to track it down. However, the approval checker can list the results of discrepancies and record the decision-making process, making it easier to present as supporting documentation for GMP requirements," Hatakeyama points out, highlighting benefits that go beyond mere efficiency.

Furthermore, a notable secondary effect is its contribution to employee education and training. The ability to track the relationships between approval documents, standards, and instructions via LinebyLine provides an opportunity to learn about the document structure. Mr. Hiruta, who was involved in the development from an advisory perspective, also stated, "It will be good training for learning the relationships between each document, and it may lead to linking SOPs and making tacit knowledge visible," indicating that the system has the potential to generate a variety of benefits.

A survey conducted by KM Bio among its employees after the system was put into use revealed that "awareness of the importance of discrepancy verification has increased," suggesting that it will contribute to improved compliance with approval documents and the development of a quality culture.

We are also considering expanding the scope beyond just reviewing approval documents.

KM Bio has been using the approval document checker in its production environment since September 2024, but its benefits depend on its widespread adoption in the field. In this respect, Macnica 's support was invaluable, according to Hayashi.

"There was some resistance to introducing a new system, but Macnica responded carefully by holding multiple on-site briefing sessions at each factory. We were able to incorporate feedback from the field into improvements and maintain a commitment to making things even better," said Hayashi.

Macnica is continuously engaging in dialogue with on-site personnel to support the adoption of its system, while simultaneously incorporating their feedback into improvements. Rather than relying solely on cutting-edge AI technology, this approach places the human user at the center, refining the system during operation.

"Going forward, we plan to expand the number of target formulations and broaden the scope of application to include post-revision document review work. We also want to leverage the function of matching discrepancies and connections between documents to explore its application in areas such as validation work and technology transfer. Macnica always presents us with multiple options for the Company challenges, clearly outlining the advantages and disadvantages before making proposals. We hope they will continue to support us in the process of ensuring the widespread adoption of the approval document checker and in developing new applications," the two said, expressing their expectations.