Governed AI in
healthcare.
Patient decisions. Clinical evidence. Human authority.
AI may support diagnosis, stratification, monitoring and clinical reasoning — but confidence is not clinical authority.
What is at stake
in this domain
In this domain the consequence of an AI decision reaches a patient. That is what makes the difference between a fluent answer and a sound one worth engineering for.
A patient
[Decision: clinical]
The questions are the same whether the decision concerns a patient, a portfolio or a pumping station. What changes is the consequence, not the test.
One clinical decision,
step by step
The same path every governed decision takes, shown here in a clinical context.
Domain data may include clinical records, imaging, monitoring signals and other evidence used in clinical decision-making.
A model proposes — a stratification, a flag, a reading of monitoring data — and states what the proposal rests on.
The governed decision layer establishes what the proposal rests on, tests it against the relevant clinical mechanism rather than plausibility alone, and records what evidence would defeat it.
An identified clinical authority decides whether the proposal may become action. Model confidence does not confer that authority.
In this reference path, action may proceed only within the authorised boundary.
The outcome is evaluated against criteria agreed beforehand — including the cases where nothing changed.
This is a reference path, not a description of a deployed clinical system.
Four questions before
anything acts
Every governed decision passes the same test, in this domain as in the other two.
Verified premises
Are the premises verified or assumed?
In healthcare: are the clinical data and patient history complete, current and verified — or are missing facts being treated as known?
Adequate mechanism
Does it hold against the domain’s own mechanism?
In healthcare: does the recommendation hold against the relevant clinical mechanism, not merely against a fluent explanation?
Counter-evidence
What evidence would defeat it?
In healthcare: what finding would overturn the recommendation — and has that counter-evidence been actively sought?
Acceptance authority
Who signs before anything acts?
In healthcare: who is the identified clinical authority that accepts the decision — and does the record show what was known at the time?
Six disciplines,
one governance model
Each one exists so that a decision can be made, defended and measured — not merely produced.
Architecture & Infrastructure
The foundation beneath every governed decision
AI Platform
Connects models, agents, context, provenance and authority
Algorithms
Reasoning, prediction, optimisation, anomaly detection and decision-support engines
Governance
Human authority, provenance, bounded action, evaluation and safety
Research
The scientific and methodological foundation informing the architecture
Economics
Value measured against the organisation’s own baseline before scale
What we can design
and build here
In healthcare, we design and build the governed decision layer on top of the systems and data infrastructure an organisation already trusts — with provenance, acceptance authority and evaluation — without asking it to replace them.
Within that layer, a model's proposal is grounded, tested against counter-evidence and presented to an identified clinical authority before any high-stakes action is taken.
Our role is to design and build the architecture and governance around the decision — not to diagnose, treat or assume clinical authority. Clinical decisions remain with the authorised clinician and institution.
Research-led.
Transparently documented.
Four peer-reviewed journal articles co-authored by our CTO were produced within the authors’ own institutions at the time, before and independently of SINAIGROUP-AI. They cover brain-volume change in primary-progressive multiple sclerosis, a finger-dexterity biosensing application, and two studies of COVID-19 vaccination in patients with multiple sclerosis.
Two preprints on medical digital twins and one on human digital twins develop the governance principles used here. Across them, model confidence is separated from authority to act, and inferred attributes are required to remain contestable.
Preprints document active research. They are not peer-reviewed validation of a product, and none of these records reports a clinical deployment.
Before SINAIGROUP-AI, CTO Emanuel Shirbint worked in data analytics and data-warehouse architecture across healthcare and other sectors, and led research projects in neuroimmunology and medical imaging. This is prior founder experience, not a SINAIGROUP-AI delivery.
SINAIGROUP-AI currently has no healthcare deployments and no healthcare clients. The evidence presented here consists of prior founder experience and published research, not a SINAIGROUP-AI clinical delivery record.
Vendor-neutral
by design
We architect around an organisation’s existing stack, integrate third-party technologies, and design and build the governed decision layer — provenance, acceptance authority and evaluation — on top of it. Technology is selected for the problem, not for a vendor relationship.
Governed AI must produce
measurable value
Baseline
What happens today? What the current process costs or produces, established before anything is built.
Target
What measurable outcome should change? Agreed in advance, in the organisation’s own units, not in generic percentages.
Evaluation
Did it change, and what was it worth? Measured against the baseline after the pilot, including the cases where it did not.
What a first conversation covers
A first conversation begins with one decision: what clinical decision matters, what evidence it depends on, who holds authority over it today, and what happens if it is wrong.
We set out what we would design, what we would measure and what we would need from you.
The domain changes.
The governance problem does not.
Discuss a high-stakes
healthcare decision
Tell us what decision matters, what evidence it depends on, and what happens if it is wrong.