Intended Use Statement

The statement must unambiguously define the exact function the model performs and the target domain where it operates. Vague language creates compliance gaps.

• Good: "Classifies chest X-rays for the presence of pneumothorax in adult patients."
• Bad: "Analyzes medical images."
• Mechanism: This precision directly maps to the intended purpose definition in the EU AI Act, determining the system's risk classification and conformity assessment requirements.

A robust statement must explicitly enumerate prohibited applications and contexts where the model was not validated. This acts as a technical guardrail against misuse.

• Examples of Out-of-Scope Declarations:
  • "Not for use in pediatric populations."
  • "Not validated for triage or emergency diagnosis."
  • "Must not be used as a standalone diagnostic tool without physician review."
• Purpose: This section establishes the manufacturer's foreseeable misuse analysis, a critical component of risk management under ISO 14971 and the EU AI Act.

The statement defines the environmental and input prerequisites required for safe operation. It sets the boundaries for valid performance.

• Key Constraints to Define:
  • Input Data Format: e.g., "Requires DICOM-compliant, 16-bit grayscale images."
  • Acquisition Protocol: e.g., "Validated only on posterior-anterior view X-rays."
  • Operational Environment: e.g., "Intended for use in a clinical laboratory setting by trained radiologists."
• Impact: These constraints form the basis for data drift monitoring in production. Any deviation from these conditions invalidates the performance guarantees.

The statement must specify the target user and their required expertise level. This defines the human-AI interaction paradigm.

• User Categories:
  • Expert User: "For use by board-certified radiologists."
  • Lay User: "For use by patients to self-assess skin lesions."
  • Operator-in-the-Loop: "Requires a trained human operator to confirm all outputs before action."
• Rationale: This directly informs the Human Oversight Mechanisms required by Article 14 of the EU AI Act, ensuring the level of human control is appropriate for the user's expertise and the risk context.

The statement links the intended use to the specific evaluation protocols and performance benchmarks that validate it. It defines what "working" means.

• Elements to Reference:
  • Benchmark Dataset: "Validated on the MIMIC-CXR dataset."
  • Target Metrics: "Achieves a sensitivity of >95% at a specificity of 90%."
  • Geographic Validity: "Trained and validated on data from North American hospitals only."
• Connection: This creates a direct, auditable link between the Intended Use Statement and the quantitative evidence in the Model Card, establishing a chain of accountability from design to deployment.

The statement is not a static document; it must be version-controlled and explicitly linked to a specific model version and its Model Card.

• Best Practices:
  • Each statement version has a unique identifier (e.g., IUS-v2.1).
  • The statement is a mandatory field in the Model Registry.
  • Any change in intended use triggers a new Algorithmic Impact Assessment and potentially a new conformity assessment.
• Governance: This linkage ensures that the legal declaration of purpose is always synchronized with the technical artifact, preventing a disconnect between what is claimed and what is deployed.