1. What exactly is the intended use?
Define the patient population, clinical setting, user, decision and expected action. Determine whether the tool informs professional judgment or is expected to replace part of it.
2. What evidence supports the claimed benefit?
Ask for the study design, comparator, validation setting, outcome measures and subgroup results. A technically accurate model may still fail to improve workflow, decisions or patient outcomes.
3. How does it perform in patients like ours?
Examine age, sex, race and ethnicity, language, comorbidities, geography, care setting and disease prevalence. Identify data gaps and groups for whom performance is uncertain.
4. What are the foreseeable harms?
Consider false reassurance, alert fatigue, overdiagnosis, delayed care, automation bias, inequity, privacy loss and new workload. Define who detects and responds to these harms.
5. What human oversight is required?
Specify who reviews outputs, what training they receive, how uncertainty is communicated and how users document disagreement. Ensure staff can override or escalate without penalty.
6. What happens when the system changes or fails?
Clarify software updates, model changes, downtime, data drift and fallback processes. Establish monitoring thresholds and criteria for pausing use.
7. How will patients and staff be informed?
Transparency should be proportionate to the role and risk of the system. Explain what it does, what it does not do, how it affects decisions and whom to contact with concerns.
If the organization cannot answer these seven questions clearly, the next step is not deployment. It is further evaluation.