Clinical research technology and CRO services

Evaluate an AI device against a reference standard, not against its own brochure.

An AI medical technology needs an intended use, a model version that does not change silently, and a reference standard independent of the model. We can structure that evaluation: representativeness, agreement, error types, subgroups and a silent-mode or prospective design if the protocol says so. This page publishes no accuracy result and no MDR certificate.

What you can use

Intended use

Who the model is for, and what decision it is allowed to inform.

Version control

The model version is part of the record. A later model is a new version.

Reference standard

Readers or a clinical standard defined before the cases are scored.

Errors that matter

False reassurance and missed urgency, not only an average score.

Subgroups

Performance is not assumed to be the same in every hospital or age band.

Human factors

How a clinician uses the output is a study question, not a slogan.

How the work moves

Lock the version

Before enrollment.

Collect against the reference

The model does not grade itself.

Analyze the named set

The snapshot is the one in the report.

Keep independence

If we develop a related device and also run the study, that conflict is written down and oversight is independent.

Related

Deliverables

Evaluation strategy
What claim is being examined, and what would count as a clinically important error.
Protocol
Population, setting, and whether the output is hidden from staff or may be seen.
Intended use and population
Who the model is for, and who is outside that use.
Reference standard
Who or what is the reference, and how disagreements are adjudicated. Independent of the model.
Sample-size justification
Assumptions written beside the number. No number is published on this page.
Data collection plan
Which fields, which model version, and how a later version is kept out.
Performance analysis plan
The estimates and intervals the study will report, before the look.
Error analysis
Missed urgency and false reassurance, not only an average score.
Subgroup evaluation
The groups named in advance. Not a search for a flattering slice afterwards.
Clinical study report
Sentences that match an approved output from the named snapshot.

People, software, and what is excluded

A qualified person writes the strategy, the reference standard and the interpretation. The platform holds the protocol version, the model version, the forms, the adjudication and the snapshot. The platform does not score its own accuracy and does not decide that the device is safe. Excluded unless the scope says otherwise: building the model, a live hospital interface, and any public accuracy claim. STARD is the reporting checklist authors use for diagnostic accuracy; this page is not that checklist.

Questions

Do you publish an accuracy number for ERTRIAGE here?

No. Enrollment, approvals and results are not claimed on this site.

Is a silent-mode study the same as clinical use?

No. The protocol defines what staff are allowed to do with the output.

Software on this page is what the product can record today. CRO work is a professional service under a separate scope. Neither is a certification, a regulator’s approval, or proof that a study is compliant because the software was used. A qualified person still signs consequential decisions.

See the workflow on your protocol.

The conversation is about your study design and what the software can and cannot do today.