Penrod Blog

New FDA AI and Machine Learning Regulations: What Do they Mean for Medical Devices

Last Tuesday, the FDA released an action plan to further define the regulatory framework surrounding the use of artificial intelligence (AI) and machine learning (ML) in medical devices.

The infusion of AI and machine learning into the design of advanced medical devices is just beginning, and the potential to improve patient outcomes is undoubtedly exciting. Much of the potential comes down to its promise to learn and improve performance, making it suitable for application as software as a medical device (SaMD). The main difference between AI/ML-based SaMD and traditional software-based devices is that the software in AI/ML-based SaMD can be constantly updated.

AI/ML-based SaMD is a field that is growing rapidly, yet the existing regulations are inadequate to address the safety and effectiveness challenges posed by such devices. The FDA hopes that a regulatory framework tailored for this technology will facilitate the development of safe and effective devices that improve healthcare delivery.

In this article, we'll discuss:

Differences between fixed and adaptive devices
New implications of the framework
Scope and shortcomings of pre-specification

Some of the proposals from Tuesday’s updated framework depend on whether the SaMD is locked or adaptive.

Most of the approved SaMDs have locked algorithms, meaning that algorithm changes beyond the original market authorization require premarket review. Adaptive AI/ML technologies require a new Total Product Lifecycle (TPLC) regulatory approach that allows the device to improve continually while ensuring patient safety. The proposed framework anticipates three types of device modifications that will require premarket review. The three categories are changes in:

  • Clinical and analytical performance
  • The inputs used by the algorithm
  • The intended use of the device

The TPLC Approach for AI/ML-based SaMD

The TPLC approach is applicable to AI/ML-based SaMD due to their ability to adapt and improve during real-world use. The FDA proposes a novel TPLC approach that will balance the risks and benefits of AI/ML-based SaMD that is premised on the four principles of:

  • Establishing clear expectations on quality systems and acceptable ML practices
  • Performing premarket review of those devices that require premarket submission to ensure reasonable assurance of safety and effectiveness.
  • Establish clear standards for manufacturers to manage risks throughout the product lifecycle continually.
  • Manufacturers will be expected to monitor their devices and utilize sound risk management approaches in the development, validation, and execution of algorithm changes.
  • Facilitate increased transparency to the FDA and users by providing real-world performance reports to maintain continued assurance of device safety and effectiveness.

Quality Systems and Good Machine Learning Practices (GMLP)

The quality systems proposal is an extension of the FDA’s requirement that medical device manufacturers establish a quality assurance system that ensures the development, delivery, and maintenance of high-quality products throughout the product lifecycle. The devices should also conform to the applicable standards and regulations. The devices should undergo analytical and clinical validation as provided in the clinical evaluation guidance.

Since AI/ML-based SaMD are subject to modifications, the FDA proposes that manufacturers adopt GMLP best practices by ensuring that:

  • The data used is relevant to the clinical problem and the latest clinical practices.
  • The data obtained is consistent, clinically relevant, and generalizable in a manner that aligns with the intended use of the device and modification plans.
  • Adequate distinction between training, tuning, and test datasets
  • Adequate levels of clarity and transparency about the algorithm and the output.

Premarket Assurance of Safety and Effectiveness

The proposed framework gives device makers the option of submitting a modification plan during the initial premarket review of the product. The review and determination of the acceptability of the plan would provide reasonable assurance of the safety and effectiveness of the device.

The proposed framework is based on the principle of a predetermined change control plan. This plan will includes:

  • A SaMD Pre-Specification that outlines anticipated modifications
  • A Algorithm Change Protocol that outlines the methodology being used to implement the changes in a controlled manner to protect patients

SaMD Pre-Specifications (SPS)

The SPS provides a section for the potential changes that the manufacturer expects the device to achieve while in use as the algorithm learns from real-world data.

Algorithm Change Protocol

The ACP relates to the methods the manufacturer intends to use to control the risks resulting from the anticipated modifications identified in the SPS. It outlines the mechanism through which the modification will achieve its goal while ensuring device safety and effectiveness.

Scope and Shortcomings of The SPS and ACP

The FDA acknowledges that not all changes can be pre-specified in the SPS and managed through the ACP. They concede that some devices may require a customized premarket review of the benefits and risks to patients. For cases where the risks change significantly after the modification, the FDA will explore an alternative to the SPS and ACP approach. However, several applications render themselves to the general principle of establishing adequate SPS and corresponding ACP. This include cases where:

  • The modifications involve improvements in performance of changes in the input but do not affect the intended use of the device. In such cases, the modification can be achieved by developing appropriate pre-specifications and ACP that offer realistic assurances that the initial level of performance will be maintained or improved. The ACP can outline the grounds for validation, methods of monitoring, and controlling for significant decline in performance or new risks to patients.
  • Changes related to the intended use of the device, such as the increased appropriation of the data available to the user in the same healthcare situation. The SPS can include modifications related to the intended use by describing how the new change will impact healthcare practice. The FDA and can collaborate with the manufacturer in developing an appropriate ACP that will describe how the device will improve performance to an acceptable level.
  • Changes related to the intended use, such as expanded use where the manufacturer adds new applications of the device or new patient population to the indications for use. In this case, the ACP can include a plan for the reference standard used for a medical condition to ensure that it provides a suitable representation of the condition. Alternatively, the ACP can provide a demonstration of the clinical relationship between the input data and the disease. It may also include strategies for data collection and testing the algorithm in the patient population.

Modification Methods After Initial Review

The FDA proposes a modified version of the learning, adaptation, and optimization requirements for SaMD to be applied to AI/ML-based SaMD after they receive market authorization. The proposed approach is expected to adequately manage risks resulting from the modifications while allowing developers to improve device performance and advance patient care.

In this model, the manufacturer will be expected to evaluate the risks posed to patients by complying with the software modifications guidance. The guide utilizes a risk-based approach and expects the developer to perform risk assessment and ensure reasonable risk mitigation. Depending on the type of modification the device may require:

  • The submission of a new 510k for premarket review
  • Documentation and analysis in the risk management and 510k files.

For devices with an approved SPS and ACP, and the modifications are within the provisions of the SPS and ACP, the manufacturer will document the changes in a similar way to the documentation approach outlined in the software modifications guidance. For modifications that are beyond the intended use, the manufacturer will not be expected to submit a new premarket submission.

Instead, the FDA proposes a model where the SPS or ACP will be refined based on real-world data and training for the intended use. The FDA may perform focused reviews of the proposed SPS and ACP for a particular device.

Transparency and Monitoring

The FDA proposes a TPLC approach that ensures transparency about the functions and modifications made to devices as a way of ensuring safety. Meanwhile, real-world data collection and monitoring will enable manufacturers to mitigate the risks involved in the use of AI/ML-based SaMDs. The FDA will expect developers to provide periodic reports on the updates that are part of the approved SPS and ACP. They will also be expected to provide performance metrics for the devices. The FDA has proposed several mechanisms of achieving this objective.

  • Transparency measures might include providing updates to the FDA, the public device manufacturers and supply chain partners, patients, clinicians, and general users.
  • Manufacturers will be required to ensure that the modifications in the SPS and the ACP are labeled accurately and that the labels completely describe the changes, including the rationale and an update on the performance.
  • Manufacturers will need to provide updates on the changes in the specifications or compatibility with supporting devices, components, and accessories.
  • Manufacturers will need to establish communication procedures outlining how they will notify users about updates and the information they will share in these updates.

The FDA has also suggested ways in which manufacturers can monitor real-world performance using its pilot programs, such as making filings in annual reports, engaging in case for quality activities, and extracting real-world performance analytics using the Pre-Cert Program. The type of report and frequency of reporting will depend on the risks associated with the device, the maturity of the algorithm, and the types and amount of modifications. Participation in the pilot programs will provide another opportunity for manufacturers to support continued safety assurance and the effectiveness of the devices.

Penrod’s Final Take

The FDA’s new framework is a response to the opportunities, challenges, and risks posed by AL/ML-based devices. Fortunately, it’s clear that the FDA will support adaptive technologies, one of the most promising applications of AI and machine learning to improve patient outcomes. We predict this mainly because the new proposal recognizes the benefits of adding real-world data to the algorithms iteratively without making premarket submissions. As the framework becomes more defined, we’re confident they will allow device manufacturers to continually improve devices and provide more benefits to patients.

Excited about medical device innovation?

We help innovate your sales and revenue strategies with the power of Salesforce.

Let’s talk! Fill out the form and we'll be in touch.