Medical artificial intelligence systems are increasingly capable of self-modification and continuous learning—but current clinical validation practices assume interventions remain static after approval. A new governance framework published in Discover Artificial Intelligence proposes a systematic method for determining when AI system updates require fresh clinical evidence, addressing what researchers call a fundamental mismatch between AI's pace of change and the speed of clinical validation.

A Two-Axis Approach to AI Evolution

The VERSION-AI framework, developed by Mohammad Hossein Imani and Amir Hossein Imani, separates system changes from their potential clinical consequences. The authors classify modifications into nine categories (C0 through C8) that describe what has changed in the AI system itself—whether that's the foundation model, prompt engineering, retrieval layer, generation settings, available tools, memory systems, user interface, or clinical workflow integration.

This classification is then mapped against clinical consequence factors: the system's intended use, how much clinical decisions depend on it, time sensitivity of those decisions, and the likelihood and severity of harm if the modified system produces incorrect, omitted, or delayed outputs. By separating "what changed" from "what could go wrong," the framework aims to provide a structured decision pathway that neither blocks beneficial updates nor allows potentially harmful modifications to slip through unvalidated.

From Simple Audits to Full Clinical Reassessment

The framework proposes a spectrum of evidence responses tailored to the nature and risk of each change. At the low end, minor modifications might require only documentation audits. Mid-range changes could necessitate usability testing, repeated-run bridging studies that compare old and new versions on the same cases, or cross-version validation exercises.

Higher-risk modifications might demand local validation in the deployment environment, shadow evaluation where the updated system runs alongside the current version without affecting care, continuous monitoring infrastructure, or mandatory rollback capabilities. At the highest tier, substantial changes to high-consequence systems would trigger renewed prospective clinical assessment—essentially starting the validation process over.

The authors emphasize that generative and agentic AI systems present unique challenges because they can change through multiple pathways: developers may update the underlying model, modify prompts or retrieval mechanisms, or alter the tools available to the system. Meanwhile, the systems themselves may evolve through accumulated memory or learned patterns, even without explicit developer intervention.

An Unvalidated Proposal Seeking Empirical Testing

The researchers are explicit about the framework's limitations: VERSION-AI has not been prospectively validated and should not be interpreted as a regulatory standard or implementation-ready instrument. No datasets were generated or analyzed in developing the framework, which the authors present as a research proposal requiring empirical testing across diverse clinical AI applications.

The paper acknowledges that determining appropriate evidence thresholds involves judgment calls that balance innovation speed against patient safety—a tension that will require input from clinicians, developers, regulators, and patients themselves. The framework's ultimate utility will depend on whether it proves practical in real-world deployment scenarios and whether it can adapt to AI architectures that don't yet exist.

For healthcare institutions already deploying AI systems, the framework highlights a critical operational question: Who is tracking what changes in your AI tools, and who decides whether those changes require revalidation before continued clinical use? As medical AI systems become more dynamic, answering those questions may become as important as the initial validation itself.

Sources

  1. Keeping clinical evidence current as medical artificial intelligence systems change — Discover Artificial Intelligence
  2. New framework asks when medical AI updates require fresh clinical evidence — bioengineer.org