How FDA Change Plans Can Inform AI Device Procurement

Hospital buyers should ask how an AI-enabled device will be changed after purchase, not only what the initial demonstration shows. Our assessment is that FDA change-control documents provide a useful starting point for those questions. Procurement teams should distinguish a manufacturer's proposed modification process from evidence that a particular update is appropriate for their own clinical setting.

Turn planned changes into specific questions

FDA guidance recommends describing planned modifications, their development and validation methodology, and their impact. (PCCP guidance). The joint PCCP principles describe plans as focused and bounded to specific changes within the original intended use or purpose. (joint change principles).

Our recommendation is to ask which changes the supplier intends to make and how the buyer will recognise them. Request a plain-language explanation of the authorised scope where applicable. Do not write a contract on the assumption that every future modification is covered by the same plan.

Related reading: health research coverage.

Design transparency around the user

The joint transparency principles emphasise human-AI team performance and clear information for users. (transparency principles). FDA's good machine learning practice page describes principles covering the total product life cycle. (machine learning practice).

For a proposed deployment, ask what clinicians will be told before and after an update. Our suggested review should include how limitations are presented and who receives change notifications. Evaluate that communication in the intended workflow rather than treating the existence of a technical document as sufficient explanation.

Related reading: medical AI coverage.

Check the device rather than a list entry

FDA states that its AI-enabled medical device list is not comprehensive. (device list limitations). FDA describes a risk-based approach that considers intended use and technological characteristics. (AI device overview).

Our proposed diligence process should inspect the particular device record and the supplier's explanation of its intended use. Ask how the proposed hospital workflow compares with that description. Neither presence nor absence on a general list should replace a device-specific review by appropriately qualified clinical and regulatory specialists.

Related reading: hospital technology analysis.

Keep classifications and workflow distinct

FDA's clinical decision support guidance discusses software functions excluded from the device definition and functions that remain devices. (decision support guidance). CDRH's transparency announcement highlights understanding users, environments and workflow. (CDRH announcement).

Our recommendation is to ask the supplier to explain the classification relevant to the exact functionality being bought. Keep that explanation separate from a local assessment of usability and clinical integration. If the proposed use changes, seek a fresh qualified review rather than extrapolating from a different configuration.

Related reading: digital health coverage.

Separate draft recommendations from final documents

FDA labels its AI device lifecycle management document as draft guidance. (draft lifecycle guidance). The Digital Health Center of Excellence lists sharing knowledge and innovating regulatory approaches among its objectives. (digital health objectives).

For the next procurement review, request a version-specific evidence package, a change-notification process and an agreed monitoring plan. Our assessment is to keep draft proposals clearly labelled and avoid reading guidance as a guarantee of clinical benefit. Local acceptance should depend on the evidence relevant to the device and setting. We would also agree who can pause deployment and how a concern reaches the supplier. Ask the clinical team to define the circumstances that should trigger reassessment, then review whether the proposed contract gives the institution a practical way to act.

Related reading: AI biology coverage.