FDA AI/ML Action Plan: De-Risking Cardiac AI Investments

Listen to this article · 7 min listen

Companies developing AI health tools without a clear FDA Software as a Medical Device (SaMD) pathway face increasing enforcement actions and exclusion from health plan coverage. This risk is particularly acute for adaptive products, where the absence of a defined regulatory strategy can lead to significant market and operational challenges.

The FDA AI/ML Action Plan 2021 fundamentally reshaped review expectations for these adaptive products. It moved the focus from mere promises of innovation to documented commitments regarding a product’s post-clearance lifecycle. These commitments are not abstract. They are checkable, forming a critical line between a vendor’s claims and a reviewable regulatory record.

The Shift in Regulatory Expectations

Before the 2021 action plan, the regulatory field for AI/ML-driven medical devices was still coalescing. While the FDA Center for Devices and Radiological Health (CDRH) had begun to address AI, the specifics of managing continuously learning algorithms remained a significant challenge. The action plan provided a clearer framework, emphasizing transparency and a lifecycle approach to regulation. This framework acknowledges that AI/ML models, particularly those intended to adapt over time, require a different oversight model than static software.

The plan introduced the concept of a Predetermined Change Control Plan (PCCP), a critical component for adaptive AI/ML devices. A PCCP allows manufacturers to define modifications that an AI/ML device can make to its algorithm without requiring a new premarket submission for every change. This represents a significant departure from traditional device regulation, where even minor modifications often necessitate a new 510(k) clearance. Without a PCCP, every time an AI model retrains on new data, a new submission could be required, rendering the product unscalable from a regulatory perspective.

Commitments Beyond Clearance: Change Control and Evidence Generation

A submission for an AI/ML SaMD product now carries explicit commitments about how the product will be followed after clearance. These commitments typically fall into two main categories: change control and follow-up evidence generation. For change control, the PCCP outlines the types of modifications that can be implemented, the methods used to implement them, and the performance boundaries within which the modified device must operate. This provides both the FDA and the manufacturer with a roadmap for managing algorithmic evolution.

Beyond change control, manufacturers are expected to generate real-world evidence (RWE) to monitor the device’s performance and safety in actual clinical use. This RWE is important for detecting issues like algorithmic drift, where the performance of an AI model degrades over time as real-world data distributions shift away from the training data. The FDA AI/ML Action Plan 2021 underscored the importance of continuous monitoring and the submission of post-market performance data. This ensures that the benefits of adaptive AI are realized while mitigating potential risks. For instance, the recall rate for FDA-cleared AI/ML medical devices, which stands at approximately 5%, is significantly higher than for non-AI devices, highlighting the critical need for strong post-market surveillance. JAMA study on AI/ML medical device recall rates

SaMD-Informed Architecture at Scale: Hello Heart as an Example

The regulatory materials related to Hello Heart, Tempus AI, and Butterfly Network illustrate this shift in review expectations. These companies operate within the FDA AI Device Regulation frame, where the 2021 action plan set new benchmarks. Hello Heart, for example, demonstrates how a SaMD-informed architecture can be implemented at scale within the cardiac AI space. Their approach to managing hypertension and other cardiovascular conditions through a digital health platform requires continuous data intake and algorithmic refinement. The company’s regulatory posture, by most accounts, reflects an understanding that ongoing engagement with the FDA’s evolving expectations is not merely a compliance burden but a foundational element for market access and sustained operation.

Hello Heart’s engagement with partners such as Navitus for medication adherence and Amwell for telehealth integration further shows the complexity of their ecosystem. Each integration point and data flow must align with the SaMD framework, particularly concerning data privacy (HIPAA) and security certifications like HITRUST or SOC 2 Type II. The absence of such certifications can be a red flag during due diligence for investors and can complicate partnerships with health plans. FDA guidance on SaMD framework and cybersecurity

Similarly, Tempus AI and Butterfly Network, operating in diagnostics and imaging respectively, navigate comparable regulatory pathways. Tempus AI, with its focus on precision medicine and genomic data, and Butterfly Network, with its portable ultrasound technology, both represent AI-native companies whose core products rely heavily on continuously learning algorithms. Their regulatory submissions and subsequent post-market activities are subject to the same heightened scrutiny regarding change control and real-world performance monitoring that the 2021 action plan brought into sharper focus.

The Checkable Paper Trail

The instructive read of the FDA AI/ML Action Plan 2021 is that it changed the paper trail rather than merely the promise of innovation. For regulatory affairs leads, this means that a submission now carries specific, documented commitments. These commitments include detailed plans for how an AI/ML product will manage algorithmic changes, how its performance will be monitored in the real world, and what evidence will be collected and reported post-clearance. The FDA CDRH publishes expectations and guidance that allow interested parties to follow this trail without direct vendor conversations.

For instance, the FDA’s publicly available database of AI-enabled medical devices, alongside various guidance documents, provides a window into the regulatory commitments made by cleared products. This allows for an independent assessment of whether a company’s claims about its innovative AI tools are backed by a reviewable regulatory record. The emphasis is now on demonstrable adherence to a lifecycle approach, where initial clearance is just one step in an ongoing regulatory journey. This checkable trail is the distinction between a mere claim of innovation and a verifiable, reviewable record of regulatory compliance and ongoing safety. FDA AI/ML medical device list

Frequently Asked Questions

What is the primary change introduced by the FDA AI/ML Action Plan 2021 for adaptive AI products?

The FDA AI/ML Action Plan 2021 shifted review expectations for adaptive products from promises of innovation to documented commitments regarding a product’s post-clearance lifecycle. These commitments are checkable and form a critical line between vendor claims and a reviewable regulatory record. It also introduced the concept of a Predetermined Change Control Plan (PCCP) for adaptive AI/ML devices.

What is a Predetermined Change Control Plan (PCCP) and why is it important for adaptive AI/ML devices?

A PCCP allows manufacturers to define modifications an AI/ML device can make to its algorithm without requiring a new premarket submission for every change. This is crucial because without a PCCP, every time an AI model retrains on new data, a new submission could be required, making the product unscalable from a regulatory perspective. It provides a roadmap for managing algorithmic evolution.

What are the key post-clearance commitments expected from manufacturers of AI/ML SaMD products?

Manufacturers are expected to make explicit commitments about change control and follow-up evidence generation after clearance. Change control involves outlining modifications, methods, and performance boundaries within a PCCP. Evidence generation requires generating real-world evidence (RWE) to monitor device performance and safety, crucial for detecting issues like algorithmic drift and ensuring continuous monitoring.

Why is real-world evidence (RWE) important for AI/ML medical devices post-clearance?

RWE is crucial for monitoring the device’s performance and safety in actual clinical use. It helps detect issues like algorithmic drift, where the AI model’s performance degrades as real-world data shifts from training data. This ensures the benefits of adaptive AI are realized while mitigating potential risks, especially given the higher recall rates for AI/ML medical devices.

Editorial Team

The Editorial Team at Regulated AI Health is a group of writers and editors who cover health, workplace wellness, and everyday care. They research each story and publish under a shared desk byline — experts in their fields, working as one team.