In recent years, the application of Artificial Intelligence (AI) technology in the field of medical devices has been rapidly advancing, providing convenient and efficient technical means for disease diagnosis, chronic disease monitoring and management. The registration and approval of AI medical devices is a prerequisite for determining whether AI technology can be productized, whether it can gain market access and be used in large-scale clinical applications. Unlike traditional medical devices, AI medical devices are "learnable" and can be continuously trained to optimize software performance by inputting real-world clinical data. This means that if you don't lock the learning function of the product, the performance of the product will change with the increase in the number of times it is used, and this feature has become a major challenge for the review and approval of AI medical devices and post-market regulatory work.
In 2017, the U.S. FDA was the first in the world to approve the marketing application of an AI medical device product. Since then, the FDA has continued to conduct pre-market review of AI medical devices through the conventional medical device approval pathway based on risk classification, and on the other hand, the FDA has been committed to exploring innovative review modes to realize the safety and efficacy of AI medical devices, allowing them to "learn" and "evolve" to a certain extent. "The newest addition to the list is the newest addition to the list, which is the newest addition to the list.
The Action Plan includes the following five elements:
I. Software Management Framework for Artificial Intelligence in Medical Devices
Currently, the pre-approval framework has not been formally implemented, and the FDA plans to issue guidance in 2021 - Draft Guidance on Scheduled Change Control Plans. -Draft Guidance on the Predetermined Change Control Plan‖ to clarify the details of the framework.
II. Good Machine Learning Quality Management Practices (GMLP)
In the Action Plan, FDA proposes to collaborate with the Artificial Intelligence Working Group of the Institute of Electrical and Electronics Engineers, the International Organization for Standardization, and several other organizations in the next phase of its work on the machine learning standards included in the GMLP, Best Practices, etc., with the aim of guiding industry development while restraining manufacturers through standards and specifications, and promoting further standardization of AI medical device regulation.
Third, patient-centered to improve product transparency
In recent years, the FDA has received a large number of inquiries and suggestions from stakeholders regarding the writing of AI medical device labels, suggesting that it is difficult to clearly explain in the labels the data used to train the algorithm, the input and output data, the logical operations, the product performance evidence, and other content. FDA believes that it is essential for users to understand the capabilities of AI medical devices and the benefits, risks, and limitations of medical devices, which will help increase user trust in the products and technologies. As a next step, the FDA will organize workshops to guide companies in improving the transparency of medical devices to users through medical device labeling, make recommendations on what information should be included in companies' product labels, and promote the improvement of AI medical device transparency to users through the development of standards and the advancement of patient-centered programs.
IV. Regulatory Science Approaches to Algorithmic Bias and Robustness
Because AI medical device systems are trained with past medical data, and medical data collection is susceptible to biases that affect algorithmic accuracy due to ethnicity, ethics, and the economic level of the patient, the FDA will promote AI medical device systems through a collaborative effort with its Center of Excellence for Regulatory Science and Innovation (CERSI). Innovation Center of Excellence (CERSI), the University of California, Stanford University, and other academic institutions to improve regulatory science and AI product evaluation capabilities by conducting research on ways to eliminate algorithmic bias and improve robustness and resilience to reduce the impact of changing clinical data inputs on algorithms.
V. Real-World Performance (RWP)
Data collection of products in real-world use can help companies better understand and improve product performance, identify product risks and use issues in time to reduce product risk, and also benefit/risk evaluations for product application materials It can also provide data support for the benefit/risk evaluation in the product application materials. In the next phase, FDA will work with willing companies to carry out real-world performance monitoring of medical devices and develop a monitoring framework for collecting real-world product performance data and determining models and parameters.
In addition, the Action Plan describes an AI cardiac ultrasound diagnostic software that was approved for marketing by the FDA in February 2020 through the De Novo pathway. This software, unlike other AI ultrasound software previously marketed in the U.S., is the first FDA-approved AI software that can guide users through ultrasound image acquisition. The software can be trained through machine learning to differentiate between acceptable and unacceptable ultrasound scans to quickly capture high-quality ultrasound images, and laypersons such as nurses can be trained to obtain image images with a similar level of quality as those obtained by ultrasound imaging professionals. The availability of this software also marks another step forward in improving the diagnostic accuracy and effectiveness of artificial intelligence software.
Currently, the Action Plan focuses on AI-enabled standalone software (SaMD), but the FDA has indicated that the review ideas presented in the plan may also be used in the future for product reviews of types of products, such as medical devices equipped with AI-enabled software (SiMD).
Article Chinese Academy of Medical Sciences, Institute of Medical Informatics, Ouyang Zhaolian? Yan Shu
Source Chinese Medicine News