How Should Surgeons Evaluate an AI Tool?

Surgeons are increasingly asked to try, pilot or comment on AI tools, from image analysis and risk prediction to video review and documentation. Marketing material rarely gives the full picture. The questions below draw on published reporting standards and NHS evaluation frameworks, and are intended as a structured starting point rather than a checklist that settles whether a tool is suitable.

What problem is it solving, and for whom?

A clear clinical question should come first. It helps to ask who the intended users and patients are, where in the care pathway the tool sits, and what currently happens without it. A tool that performs well on a problem that does not matter locally is unlikely to add value [1].

How was it developed and validated?

For prediction models, the TRIPOD+AI statement sets out what a developer should report, including data sources, handling of missing data, and how performance was measured [2]. Useful questions include: was the model validated on data from outside the development site? Does the validation population resemble your own patients? Are calibration and clinical utility reported, not only accuracy or area under the curve?

Has it been tested in real clinical use?

Retrospective performance often falls when a tool is used prospectively. The DECIDE-AI guideline describes how early live clinical evaluation should be reported, with attention to human factors and how clinicians interact with the output [3]. Where randomised evidence exists, CONSORT-AI and SPIRIT-AI describe the expected reporting standard [4,5].

What is its regulatory status?

Software that is intended to inform diagnosis or treatment is likely to be a medical device. In Great Britain this falls under the Medical Devices Regulations 2002 and the MHRA's Software and AI as a Medical Device programme [6]. Asking for the device classification and conformity marking (UKCA or CE) is a reasonable early step.

Does it meet NHS standards?

The NICE Evidence Standards Framework for digital health technologies describes the level of evidence expected for different types of product [7]. The NHS Digital Technology Assessment Criteria (DTAC) covers clinical safety, data protection, technical security, interoperability and usability [8]. Local clinical safety officers will usually want to see DCB0129 and DCB0160 documentation.

Who is monitoring it after deployment?

Performance can drift as patient populations, equipment and practice change. It is worth asking how performance will be monitored locally, how errors will be reported, and who is responsible for switching the tool off if concerns arise.

Summary

No single study or certificate answers every question. Reading the evidence against recognised reporting standards, confirming regulatory status and agreeing local monitoring gives surgeons a more informed basis for discussion with colleagues, managers and developers.