AI·СПБЭПР медицинского ИИ
ENВ коалицию
Меню

Saint Petersburg · working policy concept

Evidence before autonomy.

A proposed health AI sandbox where a system earns permission for one bounded clinical action — never for “medicine in general”. Humans validate the system, handle uncertainty and can stop every version.

Open evidence packInternational peer review
STATUS / 13 AUG 2026Concept. Not operational.

Not an official programme of the Government of Saint Petersburg. Not a medical service, trial invitation or regulatory approval.

No. But removing case-level confirmation should require stronger system-level controls: a narrow intended purpose, independent local validation, an abstention envelope, a closed human escalation path, version control and continuous outcome monitoring.

Human control ≠ human click

The proposal moves attention from a ceremonial signature to demonstrable governance before, during and after each decision. Whether Russian law can permit A3 remains a federal legal question.

Autonomy is a permission layer

A0

Digitisation

Scheduling, search and data transfer.

A1

Decision support

A clinician makes the decision.

A2

Delegated protocol

A clinician performs meaningful case-level confirmation.

A3

Bounded autonomy

A proposed special regime for one validated action, with mandatory abstention.

A4

Legal act

Prescription or sick leave only through a separate federal legal gateway.

  1. Diabetic retinopathy screeningThe closest candidate for bounded autonomous classification, with image-quality abstention and closed-loop referral.
  2. ECG signal tasksSplit into urgent triage, selected rhythm classification and long-term screening — not one broad “ECG diagnosis”.
  3. Hybrid closed-loop diabetes systemsA reference case for autonomy inside a registered device configuration and a clinician-set prescription.
  4. Acute respiratory symptomsShadow mode and A1–A2 first. Safety is measured by missed red flags and repeat care, not completion rate.
  5. Localised urinary tract infectionNarrow cohort, mandatory exclusions and antimicrobial stewardship. No autonomous antibiotic prescription.

The concept learns from the EU AI Act sandbox framework, the UK MHRA AI Airlock, Singapore’s AI-SaMD sandbox, FDA-authorised autonomous screening systems and WHO guidance. It does not claim equivalence or foreign endorsement.

  • machine-readable intended purpose and exclusion envelope;
  • locked external validation and meaningful subgroup analysis;
  • prospective clinical utility, workflow and health-economic endpoints;
  • public stop events, negative results and model-change history;
  • four evidence labels: measured, assumption, modelled, target.

Review primary sources