Dynamic Regulatory Frameworks for Adaptive Clinical AI

This direction is split into other directions: runtime verification, continuous learning regulation. Its last version stays readable and citable here.

Version 1 of 1309 papers427 conceptsfrozen 2026-10-04

Traditional medical regulation is built on a "build-and-freeze" paradigm. When a company develops a new stethoscope or a heart valve, the device is tested, validated, and then "locked." Any significant change to its design requires a new round of regulatory approval. This static approach ensures safety for hardware, but as Amelie Echle and Jakob Nikolas Kather highlight in Deep learning in cancer pathology: a new generation of clinical biomarkers, it creates a fundamental mismatch for artificial intelligence.

Machine learning models are inherently dynamic. They are prone to "algorithmic drift"—a phenomenon where a model’s performance degrades because the real-world data it encounters (patient demographics, hospital equipment, or clinical protocols) shifts away from the data it was trained on. To remain safe and effective, these models must evolve. Researchers are now proposing a shift from regulating AI as a static "medical device" to managing it through a continuous, adaptive lifecycle.

The Failure of the Static Snapshot

The core problem is that a "pre-market snapshot"—the performance of a model at the moment of its first approval—is a poor predictor of its long-term safety. Michael Da Silva and Colleen M. Flood argue in Regulating the Safety of Health-Related Artificial Intelligence that existing laws are ill-equipped for the "dynamic" nature of ML.

When a model is "locked" to satisfy regulatory requirements, it becomes a clinical liability. It cannot adapt to new biological data or changing clinical environments. Holger Fröhlich and Thomas Lengauer note that this "build-and-freeze" mindset is a primary bottleneck for personalized medicine, where systems must iteratively learn to be effective (From hype to reality: data science enabling personalized medicine).

From Device to Provider

If AI shouldn't be regulated like a piece of hardware, how should it be treated? Edward Raff and Shannon Lantzy propose a conceptual "flip": treat the AI like a medical provider, such as a doctor, rather than a medical device (Dr. AI, Where Did You Get Your Degree?).

Instead of approving the "black-box" internals of the algorithm once, regulators would focus on:

  1. Accreditation: Validating the data collection and development process.
  2. Continuous Monitoring: Tracking clinical outcomes in real-time.

This shift allows the model to learn and update post-deployment, provided it stays within "predefined boundaries" or an approved "learning trajectory."

Mechanisms for Continuous Validation

To make this transition practical, researchers have introduced several technical and procedural frameworks:

Predetermined Change Control Plans (PCCPs)

Rather than approving a single version of a model, regulators can approve a modification path. Bouchra Derraz describes this as approving the "learning logic" of the algorithm (New regulatory thinking is needed for AI-based personalised drug and cell therapies in precision oncology). If the model updates itself using the approved method and stays within performance guardrails, it doesn't need a full re-certification.

Recurring Local Validation (REVAFT)

Because "all models are local," a model that works in a Boston hospital might fail in a rural clinic. Alex Youssef and Tingting Zhu advocate for REVAFT (REpeating VAlidation and Fine-Tuning), where models are continuously tested and recalibrated at the specific site of deployment to account for local data drift (All models are local: time to replace external validation with recurrent local validation).

Online Hypothesis Testing and "Bio-Creep"

Updating a model carries the risk of "bio-creep"—a gradual degradation where each new version is slightly worse than the original, but acceptable compared to the version immediately before it. Jean Feng and Noah Simon propose a statistical framework using online hypothesis testing to ensure that cumulative updates never fall below a rigorous safety baseline (Approval policies for modifications to Machine Learning-Based Software as a Medical Device).

The Emergence of Regulatory Science

This research direction suggests that regulation itself must become an active scientific discipline. Jochen K. Lennerz and Drew F. K. Williamson argue that "Regulatory Science" is the essential framework to bridge the gap between AI developers and clinicians (A unifying force for the realization of medical AI).

By treating governance as a dynamic process—using tools like "regulatory sandboxes" and "adaptive scope monitoring" (Johannes Schneider)—the field is moving toward a future where AI systems are not just "cleared for takeoff," but are continuously monitored and refined throughout their entire operational life.

Papers in this version

Showing 60 of 309 papers.

History & citationVersion 1 · 2026-10-04

Split into other directions · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_dynamic_regulatory_frameworks_for_adapti_c1525473,
  title = {Dynamic Regulatory Frameworks for Adaptive Clinical AI},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/dynamic-regulatory-frameworks-for-adaptive-clinical-ai/rdr_c1525473201a415e8890273be24752de},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_dynamic_regulatory_frameworks_for_adapti_c1525473_v1,
  title = {Dynamic Regulatory Frameworks for Adaptive Clinical AI},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/dynamic-regulatory-frameworks-for-adaptive-clinical-ai/rdr_c1525473201a415e8890273be24752de/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}