Regulatory Atlas Methodology

Scope and admission

The Atlas asks a narrow question: what does the public FDA record establish about software used in or near neurosurgical care? A submission is eligible when its function directly supports neurosurgery or materially shapes an imaging, planning, monitoring, neuro-oncology, trauma, or acute-neurology workflow that reaches a neurosurgical team.

The submission number is the primary join key. A product name, manufacturer name, press release, or FDA AI-list row alone is not sufficient to build a public case.

1. Match the exact FDA submission and decision date.
2. Read the decision summary and database record.
3. Extract intended use, users, inputs, outputs, exclusions, human-review requirements, validation, PCCP status, and missing information.
4. Run targeted postmarket searches using exact names and relevant product-family variants.
5. Publish only after a separate editorial review of evidence and attribution boundaries.

Source hierarchy

Claims are checked in a defined order: FDA decision summary or De Novo review memorandum; FDA submission database; FDA AI-Enabled Medical Device List; FDA recall or enforcement record; and product-matched peer-reviewed evidence when available.

Company material may help locate a source or clarify a feature. It does not independently establish regulatory status, clinical validity, safety, or outcome benefit. FDA AI-list lead panel and submission review panel remain separate fields because they may differ.

AI disclosure

AI/ML explicit

The reviewed FDA material names AI, machine learning, deep learning, a CNN, or another learning architecture.

Algorithm / classifier

The material describes an algorithm, model, or classifier but does not establish a foundation-model or generative-AI architecture.

AI function not established publicly

The device appears in FDA’s AI-enabled list, but the reviewed public decision material does not identify the AI function. The uncertainty remains visible.

Every public case separately states whether an LLM, foundation model, or generative-AI function was identified. “Not identified” means only that it was not found in the reviewed public material.

Evidence classification

Nonclinical verification, retrospective clinical validation, prospective clinical validation, and patient-outcome evidence are not interchangeable. The Atlas records the strongest design actually reported and places the main limitation beside the result.

• Accuracy, Dice, sensitivity, specificity, and navigation error are performance measures, not automatic evidence of better outcomes.
• Retrospective patient data do not establish prospective workflow utility.
• Expert acceptance of an AI-generated plan is not observed procedural accuracy.
• Clearance or De Novo authorization does not itself establish comparative superiority.

Comparator, ground-truth construction, sample size, available representation, external-validation status, and exclusions are retained so superficially similar metrics are not treated as equivalent.

Change control

A Predetermined Change Control Plan is recorded only when an FDA database record or decision document establishes it. The Atlas distinguishes a PCCP authorized in the indexed submission from one inherited from an earlier submission.

An authorized PCCP permits bounded modifications under predefined development, validation, impact-assessment, and release controls. It is not evidence of unrestricted or continuously self-learning behavior, and its acceptance criteria are not observed patient outcomes.

Postmarket attribution

Recall searches use submission numbers, exact device names, manufacturers, components, software versions, and product-family variants. Findings are labeled as component matches, historical product-family matches, or no exact-name openFDA result.

A product-family or earlier-version recall is not assigned to every installation or to the newest submission. “No exact-name result” is a bounded search finding, not proof that no recall or safety event exists. Recall records are not patient-event rates.

Public selection and maintenance

The public Atlas is deliberately smaller than the private registry. Cases are selected for explanatory value across function, pathway, evidence design, change control, human review, and postmarket supervision—not to rank vendors or approximate market share.

Monthly

Refresh the FDA AI-list snapshot and targeted recall queries.

Quarterly

Recheck decision summaries, database flags, PCCP lineage, and product-matched literature.

Event driven

Review material authorizations, labeling changes, safety communications, recalls, withdrawals, and later submissions.

Dates advance only for the work actually repeated. A layout edit does not become a new evidence review, and a narrow recall scan does not silently reset the underlying clinical-validation date.

Corrections and limits

Substantive corrections should identify the affected route, original statement, corrected statement, supporting primary source, date, and implication. Send source-backed corrections to hello@neurosurgery.ai without patient-identifying information.

This method covers a selected US FDA collection. It is not a legal opinion, procurement assessment, systematic review, continuous safety-monitoring service, or substitute for current regulator records and professional judgment.