In HCOs, AI/ML technologies are often initiated as siloed research or quality improvement initiatives. However, when these AI technologies demonstrate potential for implementation in patient care, development teams may encounter substantial challenges and backtracking to meet the rigorous quality and regulatory requirements12,13. Similarly, HCO governance and leadership may possess a strong foundation in scientific rigor and clinical studies; however, without targeted qualifications and training, they may find themselves unprepared to offer institutional support, regulatory oversight, or mobilize teams toward interdisciplinary scientific validation of AI/ML–enabled technologies required for regulatory submissions and deployment of SaMD. Consequently, the unpreparedness of HCOs exacerbates the translation gap between research activities and the practical implementation of clinical solutions14. The absence of a systematic approach to ensuring the effectiveness of practices and perpetuating them throughout the organization can lead to operational inefficiencies or harm. Thus, HCOs must first contend with a culture shift when faced with quality control rigor inherent to industry-aligned software development and deployment, specifically design controls, version control, installation qualification, operational qualification, performance qualification, that primarily focuses on end-user acceptance testing and the product meeting its intended purpose (improving clinical outcomes or processes compared to the standard of care or the current state), and the traceability and auditability of proof records (Table 1).
Consider that even in cases where a regulatory submission is not within the scope, it remains imperative to adhere to practices encompassing ethical and quality principles. Examples of such principles identified by the Coalition for Health AI and the National Institute for Standards and Technology (NIST) include effectiveness, safety, fairness, equity, accountability, transparency, privacy, and security3,7,15,16,17,18,19,20. It is also feasible that the AI/ML technology could transition from a non-regulated state to a regulated one due to updated regulations or an expanded scope. In that case, a proactive approach to streamlining the conversion from a non-regulatory to a regulatory standard should address the delicate balance of meeting baseline requirements while maintaining a least-burdensome transition to regulatory compliance.
As utilized by the FDA for regulating SaMD, a proactive culture of quality recognizes the same practices familiar to research scientists well-versed in informatics, translational science, and AI/ML framework development. For example, the FDA has published good machine learning practices (GMLP)21 that enumerate its expectations across the entire AI/ML life cycle grounded in emerging AI/ML science. The FDA’s regulatory framework allows for a stepwise product realization approach that HCOs can follow to augment this culture shift. This stepwise approach implements ethical and quality principles by design into the AI product lifecycle, fostering downstream compliance while allowing development teams to innovate and continuously improve and refine their products. Using this approach allows for freedom to iterate at early research stages. As the product evolves, the team is prepared for the next stage, where prospectively planned development, risk management, and industry-standard design controls are initiated. At this stage, the model becomes a product, incorporating all the software and functionality needed for the model to work as intended in its clinical setting. QMS procedures outline practices, and the records generated during this stage create the level of evidence expected by industry and regulators22,23. HCOs may either maintain dedicated quality teams responsible for conducting testing or employ alternative structures designed to carry out independent reviews and audits.
Upon deployment, QMS rigor increases again to account for standardized post-deployment monitoring and change management practices embedded in QMS procedures (Fig. 2). By increasing formal QMS consistency as the AI/ML gets closer to clinical deployment, the QMS can minimize disruption to current research practices and embolden HCO scientists with a clear pathway as they continue to prove their software safe, effective, and ethical for clinical deployment.