An MLOps framework for AI-integrated Quality Management Systems

English

Abstract:

Artificial intelligence (AI) models now perform much of the defect detection that was previously undertaken by human inspectors across pharmaceutical, medical device, semiconductor, and automotive production lines. However, these models are not static. Accuracy can gradually decline as manufacturing conditions change, and when this occurs, regulators require organisations to know precisely what changed and when. This article by Abhineet Johri presents a practical machine learning operations (MLOps) framework for AI-driven quality management systems (QMS), based on five factors: pipelines that check and document data automatically; monitoring that identifies and notifies when issues arise; retraining that is triggered by defined conditions rather than a fixed schedule; a model registry that can provide the information required to answer an auditor’s questions quickly; and deployment methods that enable teams to replace models while minimising risks to production. Each of these factors is considered in relation to the requirements of Title 21 of the US Code of Federal Regulations Part 820, ISO 13485, and ISO 9001, and the overall framework is summarised as a set of practices for adoption. In regulated quality environments, an MLOps framework is no longer simply an optional capability. It is the difference between an AI model that can be defended in an audit and one that cannot.

 

Learning objectives:

• An understanding of why an AI quality model requires ongoing monitoring and maintenance, rather than one-time-only validation before deployment.
• An understanding of a simple three-tier approach to alerts (warning, alert, and critical alert) that can be applied to a dashboard.
• An overview of what FDA and ISO auditors typically require, helping to avoid the need to reconstruct records retrospectively.
• An understanding of different approaches for replacing a model without interrupting production.
• An understanding of how to determine when a model requires retraining based on drift and available data, rather than according to a fixed schedule.
• An awareness of good practices worth considering as part of a process review.

 

About the author:

Abhineet Johri has more than 21 years of experience in the software medical device industry. His prior and current experience includes quality management, product risk management, complaints management and regulatory affairs. He is a consultant on processes and requirements for conducting clinical investigations/multi-reader studies for AI based devices, and has developed strategies to clear AI based medical devices in various countries around the globe. He is a published author, and a senior key expert for various regulatory processes within his organisation including AI/ML based devices. He is also a member of, and contributor to, various industry associations and AI committees such as the Medical Imaging & Technology Alliance (MITA) / MDCI and International Electrotechnical Commission sub-committees (USNC TAG to IEC SC 62B and expert on IEC SC 62B/PT 63524).

 

Specification:

6 pages plus covers, in English, supplied by email as a PDF.

This article has been published in the November 2026 issue of the Journal of Medical Device Regulation.