FDA denial, surgical robot clearance and AI change control
The FDA denied Novocure's premarket approval for its Tumor Treating Fields technology in metastatic brain cancer the same week it cleared Virtuoso Surgical's robotic bladder tumour resection system. These decisions, paired with an emerging framework for predetermined change control plans in adaptive AI medical devices, reveal where regulatory tolerance sits right now. For device manufacturers, the pattern matters more than the individual outcomes. Efficacy thresholds are rising, technical innovation alone does not carry the submission, and adaptive systems need documented change boundaries before they reach the post-market phase.
Novocure's denial and what uncertainty looks like at review
Novocure disclosed that the FDA expressed uncertainty about the degree of clinical benefit for its Tumor Treating Fields device targeting brain metastases from non-small cell lung cancer. The agency did not dispute the mechanism. It questioned the magnitude and consistency of the effect. That distinction is tactical. When clinical endpoints show signal but reviewers cannot pin down the size of the benefit reliably enough to support a broad indication, the submission stalls. Novocure can resubmit, but the path forward now involves additional data, tighter endpoint definitions, or a narrower claim. The company has not yet disclosed which route it will take.
This is not a device failure. It is a submission architecture failure. The FDA's comfort with uncertainty has contracted over the past eighteen months, especially for devices with complex mechanisms of action where the clinical effect is distributed across heterogeneous patient populations. The lesson for manufacturers is operational, not scientific. If your pivotal trial shows a statistically significant result but leaves reviewers unable to quantify benefit at the subgroup level, expect the question to surface in review and plan your response strategy before submission. That means pre-specified subgroup analyses, clearly defined responder criteria, and a clinical strategy that isolates the population most likely to benefit. Retrospective analysis will not close the gap.
Virtuoso's clearance and the robotics threshold
The same week, Virtuoso Surgical secured 510(k) clearance for its robotic system designed to remove bladder lesions in one piece during minimally invasive procedures. The FDA categorised the technology as substantially equivalent to existing electrosurgical devices and robotic platforms, but the clinical value proposition is procedural, not just technical. By resecting tumours intact rather than fragmenting them, the system aims to improve staging accuracy and reduce recurrence risk. That outcome depends on surgical precision, which is where the robotic assistance claim becomes regulatory.
For Virtuoso, the 510(k) pathway worked because the device performs a well-understood task using a mechanism with documented predicates. The innovation lies in execution, not in the fundamental method. That distinction is why the clearance moved through without requiring a de novo or PMA route. But it also exposes the limits of equivalence-based submissions for surgical robotics. As robotic platforms integrate adaptive assistance, real-time imaging feedback, or autonomous gesture refinement, the predicate logic weakens. At that point, manufacturers need to demonstrate not just equivalence but safety and effectiveness under conditions that predicates do not cover. The transition from assistive to adaptive robotics will force more de novo and PMA submissions, and teams should plan regulatory timelines accordingly.
| Device characteristic | Assistive robotics (predicate-based) | Adaptive robotics (learning or autonomous) |
|---|---|---|
| Primary regulatory route | 510(k) if predicates exist | De novo or PMA, depending on risk class |
| Clinical evidence requirement | Performance data, usability, procedure equivalence | Benefit-risk analysis, real-world validation, model behaviour documentation |
| Change control expectation | Standard design controls, version-locked at clearance | Predetermined change control plan, lifecycle monitoring |
| Post-market surveillance focus | Adverse events, device malfunction reports | Model drift, performance decay, unexpected adaptation |
Predetermined change control for adaptive AI and the logic behind it
MedTech Intelligence published a detailed exploration of what a predetermined change control plan would look like for an AI system that continues to adapt post-market. The article uses a speculative governance structure called a 'contocracy' as a framing device, where an AI stand-in for a deceased leader must be updated over time without violating the original mandate. The regulatory parallel is precise. If a medical device uses a machine learning model that retrains on new data after deployment, the manufacturer must define in advance which changes are acceptable, how performance will be measured, and when revalidation is required. That framework is the predetermined change control plan, and it is becoming non-negotiable for adaptive AI submissions.
The core challenge is boundary definition. A manufacturer cannot simply state that the model will improve over time. The submission must specify which performance metrics will trigger a change, what retraining data will be used, how model versions will be tracked, and when the FDA expects notification versus prior approval. The article outlines a control rule where a rebuilt model must match the original leader's decision record at least as well as the previous version. In regulatory terms, that translates to a requirement that any updated model performs no worse than the cleared baseline on a locked validation dataset. If performance degrades on that set, even if real-world accuracy improves, the update cannot proceed without new clearance. That logic protects against distributional drift, where a model optimises for new data at the expense of the original indication.
The FDA has signalled this expectation in its draft guidance on AI and machine learning, and stakeholder comments reveal fundamental disagreement on who owns the risk when models change. Device manufacturers argue that predetermined change control plans should allow continuous learning within defined bounds. Provider groups and patient advocates want tighter oversight, including mandatory reporting for every material model update. The gap has not closed, and until it does, manufacturers face a choice. Lock the model at clearance and accept static performance, or build a predetermined change control plan robust enough to satisfy both the FDA and downstream liability concerns. Most will choose the latter, but few have the internal infrastructure to execute it without external regulatory and technical support.
Building a predetermined change control plan for adaptive AI medical devices
- Define the locked validation datasetEstablish a representative, version-controlled dataset that reflects the cleared indication. Every model update must perform at least as well on this set as the baseline version.
- Specify retraining triggers and data sourcesDocument the conditions under which the model will retrain (for example, performance drift thresholds, new clinical data availability) and the sources from which training data will be drawn.
- Set performance bounds and acceptance criteriaEstablish quantitative thresholds for sensitivity, specificity, positive predictive value or other relevant metrics. Updates that fall outside these bounds require new clearance.
- Implement version tracking and traceabilityAssign unique identifiers to every model version, log all retraining events, and maintain a record of which version was deployed when and where.
- Define notification versus prior approval thresholdsClarify which changes can proceed with post-market notification (for example, minor retraining within bounds) and which require prior FDA approval (for example, algorithm architecture changes, new indications).
- Establish post-market surveillance and monitoringDeploy real-world performance monitoring to detect drift, bias, or unexpected behaviour, and build escalation pathways for intervention when thresholds are breached.
What this means for your team
For regulatory affairs and quality teams, these three stories clarify where the FDA's attention sits. Clinical uncertainty is no longer something you can resolve in post-market studies if it exists at submission. If your pivotal data leaves reviewers unable to quantify benefit reliably, the submission will stall, and the path forward will involve more data, not better storytelling. That means earlier engagement with clinical strategy, tighter endpoint definitions, and pre-specified subgroup analyses that isolate the populations most likely to respond. Waiting until the review cycle to address these questions is too late.
For surgical robotics, the 510(k) pathway remains viable as long as the device performs a well-understood task using documented predicates. But as platforms integrate adaptive assistance or autonomous features, the predicate logic weakens and the evidence threshold rises. Teams working on next-generation robotics should assume a de novo or PMA route and plan timelines accordingly. That includes early designation requests, pre-submission meetings to align on endpoints, and clinical validation that demonstrates benefit under real-world variability, not just controlled lab conditions.
For adaptive AI, the predetermined change control plan is no longer optional. If your device uses a model that will retrain or adapt post-market, you need to define the change boundaries, performance thresholds, and revalidation triggers before submission. That requires cross-functional alignment between regulatory, clinical, data science, and quality. It also requires infrastructure. You need version control, locked validation datasets, real-world performance monitoring, and traceability systems that can survive an FDA post-approval inspection. Most organisations do not have that infrastructure in place yet, and building it takes longer than writing the plan itself. Start now.
Key takeaways
- The FDA's tolerance for clinical uncertainty has contracted. If reviewers cannot quantify benefit reliably at submission, the pathway forward involves more data, not revised narratives.
- Surgical robotics can still clear via 510(k) if the device performs a well-understood task with documented predicates, but adaptive features push the submission toward de novo or PMA routes.
- Predetermined change control plans for adaptive AI are operational frameworks, not compliance documents. They require locked validation datasets, performance bounds, version tracking, and post-market monitoring infrastructure.
- Regulatory strategy for adaptive systems must align clinical, data science, quality, and regulatory functions before submission. The technical work takes longer than the regulatory writing, and both must be complete before the clock starts.
The regulatory environment is not tightening arbitrarily. It is responding to devices that introduce new sources of variability, whether through adaptive algorithms, complex mechanisms of action, or procedural innovation that outpaces predicate logic. For manufacturers, the task is not to fight that shift but to build submission strategies and internal systems that meet the standard before it becomes explicit. That means earlier clinical engagement, tighter change control architecture, and infrastructure that can demonstrate performance under scrutiny. The organisations that build those capabilities now will move faster when the next wave of guidance arrives. The ones that wait will face longer timelines, more deficiencies, and harder conversations with reviewers who have already seen the pattern.