AI Transparency

1. AI System Overview

Neuridion uses artificial intelligence to assist with screening Field Safety Notices (FSNs) and other supported public safety records as part of a manufacturer's post-market surveillance (PMS) process. The system assesses potential relevance to a customer-defined device profile and presents its output for human review.

AI does not make clinical decisions, diagnose patients, recommend treatments, or replace the judgment of a Person Responsible for Regulatory Compliance (PRRC). It does not determine reportability or regulatory action. The manufacturer remains responsible for the final disposition and any downstream action.

2. Models Used

Anthropic's API usage policy states that inputs sent via the API are not used to train their models.

Alternative model providers may be evaluated in shadow benchmarks. Their output is not used as the production regulatory disposition unless it passes the controlled release process for the stated dataset and configuration.

3. How AI Classification Works

NEURIDION separates auditable safety rules from advisory AI ranking:

1. Deterministic scope and vigilance rules: Structured, checkable grounds may mark a record outside scope without deleting it. Death, serious deterioration, serious-incident, FSCA, recall, and field-safety-action signals bypass model ranking and are routed to human review.
2. AI ranking: Residual records receive a high, medium, or low presentation rank with a written rationale. AI output may prioritize the review queue, but it never deletes a record or creates a regulatory exclusion.

4. Human Oversight Measures

• Every source record remains in the evidence record. Low-ranked records are a presentation category, not silently discarded evidence.
• If the AI filter fails for a specific FSN, that item is marked as "requires manual review" rather than being silently dropped.
• Users can cancel any running search at any time.
• Report release is subject to the configured human-review gate. The customer is responsible for assigning qualified reviewers and defining the meaning of approval within its own procedure.
• All filter decisions are stored in an immutable, append-only audit trail — decisions cannot be edited or deleted after the fact.
• Each AI decision records the exact provider and model identifier, prompt and ruleset versions, input snapshot hash, output hash, decision time, and whether a compatible cached decision was reused. Re-runs can reveal drift; identical output is not promised.

5. Data Handling

What is sent to AI

• FSN or source-record title, manufacturer name, date, and content text
• Device profile context (device name, manufacturer, intended use, device class)
• Version-controlled passages from customer-supplied controlled evidence, such as an IFU, only when that evidence is enabled for the assessment

Public regulatory evidence is separated from customer-supplied controlled evidence. When controlled evidence is enabled, the assessment records the document version and authorized passage references used. Customers must confirm that their provider terms, confidentiality requirements, and approved procedure permit this processing.

What is NOT sent to AI

• Your email address, password, or authentication credentials
• Customer patient records or clinical records intentionally uploaded for screening
• Payment information or billing details
• Controlled documents that have not been explicitly enabled for the assessment

6. Known Limitations

• All AI classifications are advisory only and must be reviewed by a qualified person before being relied upon for regulatory decisions.
• Confidence scores are model estimates, not calibrated statistical probabilities.
• Model APIs may return different wording or rankings on a later re-run even with the same temperature setting. Neuridion records inputs, configuration, and outputs to detect and investigate that drift rather than claiming bit-level determinism.
• Novel, rare, multilingual, or ambiguously described device records may be harder to assess reliably.
• No universal sensitivity, specificity, or accuracy claim is made. Verification evidence is bounded to the stated dataset, sources, release, and configuration.
• Controlled documents affect an assessment only where the active release explicitly extracts, versions, and supplies that evidence to the screening pipeline. A stored or uploaded path alone does not establish that its contents were used.
• AI-generated output must be reviewed under the manufacturer's approved procedure before it is relied upon in official PMS documentation.

7. Prohibited Practices Statement (EU AI Act Art. 5)

NEURIDION does not engage in any practices prohibited under Article 5 of the EU AI Act. Specifically, the system does not:

• Deploy subliminal, manipulative, or deceptive techniques
• Exploit vulnerabilities related to age, disability, or social/economic situation
• Evaluate or classify individuals based on social behaviour (social scoring)
• Perform real-time remote biometric identification

The system's sole function is classifying publicly available government-published Field Safety Notices against user-defined device profiles.

8. EU AI Act Risk Classification

Neuridion's current working position is that the bounded screening function is not a high-risk AI system under Article 6: it is not intended as a medical device or a safety component and does not match an Annex III use case. This is a draft position, not a certification or legal opinion, and it must be reassessed if the intended use, decision authority, integrations, models, data, or applicable law changes. Neuridion labels AI-assisted output and documents limitations as governance controls. The exact Article 50 duties for a released configuration require confirmation by qualified counsel and quality/regulatory review.

9. Contact

For questions about our use of AI, data handling, or this transparency disclosure, contact us at info@neuridion.eu.