Surgical video for medical device development: scoping the evidence

Which device questions can video help investigate?

A team may want to understand when an instrument enters a workflow, which structures are visible during a step, or how frequently an event is observable from a particular camera view. Those are concrete questions that can guide selection and annotation. “Understand surgery” is too broad to produce a useful data specification.

Separate exploratory learning from formal evidence generation. An archive may help a team refine a hypothesis or design a protocol without being suitable for a later performance claim. Make that distinction explicit in the project brief, license, and interpretation of results.

How should intended use shape the dataset?

Write down who will use the proposed system, what information it will provide, and in what setting. Even at an early stage, those decisions help identify the relevant procedures, recording systems, users, and difficult conditions. A dataset selected only for easy access may leave the important conditions unexamined.

The FDA: good machine learning practice points to the 2025 IMDRF principles and a lifecycle approach to AI-enabled medical devices. That is a useful context for planning; acquiring a commercial dataset does not itself satisfy a regulatory pathway or establish clinical evidence.

Bring engineering, clinical, quality, and regulatory perspectives into the brief early enough to change it. A mismatch discovered after annotation can require substantial rework.

What can a recording fail to show?

A camera may not capture a force applied outside the view, an action before recording began, or an event occurring after the procedure. A visible instrument movement does not automatically reveal the operator’s intention. Document these limits before turning observations into outcome labels.

Device identity can also be ambiguous. Similar-looking instruments may have different versions or configurations. If a study needs exact device information, specify a reliable source for that information rather than asking annotators to guess from appearance.

Differences in sites, case selection, and recording practices may affect comparisons. A video collection assembled for one purpose should not be assumed suitable for a head-to-head product claim without a separate study design.

What should an evaluation protocol include?

• A precise target task and the conditions in which it will be assessed.
• A justified reference standard and a plan for ambiguous examples.
• Independent evaluation data appropriate to the claim.
• A breakdown of important error types and relevant conditions.
• Traceability from each result to the data and annotation version.
• A clear statement of what the evaluation does not establish.

If experts assess model outputs, keep the scoring instructions separate from model development decisions where appropriate. Record disagreements and adjudication. An evaluation should expose weaknesses that help the team decide what to do next.

How can a device team start a data conversation?

Share the procedure families, intended research task, required views, relevant device context, and desired annotation. State whether the request is exploratory or part of a defined development program. Describe the timing and review requirements without including proprietary or patient details in a public form.

Operavue can discuss existing surgical data, custom capture, and clinician expertise around that brief. Availability, permitted uses, exact specifications, and delivery criteria are confirmed through project scoping rather than implied by a general website description.