ideasIdea
Rules for retiring cancer AI when performance drops or the standard of care moves
Just as drugs are withdrawn when they prove unsafe, AI tools should have clear triggers for being switched off, and someone responsible for pulling the switch.
No framework exists for taking a deployed model out of service: models trained on outdated staging or treatment eras continue to run. The proposal defines decommissioning triggers (performance below threshold on monitoring, guideline change affecting the task, vendor withdrawal, unaddressed red-team findings), assigns responsibility (site clinical AI officer, vendor, regulator), and requires notification of affected patients where results may have been wrong, mirroring device recall processes.
Hypothesis
Explicit decommissioning rules will lead to retirement of a measurable share of currently deployed cancer AI tools that are obsolete or under-performing, and will shorten the time between trigger and action.
Rationale
Software in other safety-critical domains has defined end-of-life processes; healthcare AI has accumulated a decade of deployments with no retirement mechanism.
What would test it
Apply the rules to the AI inventory of one health system; count tools meeting decommissioning triggers; measure time to action.
Maturity
speculative
Who has to act
clinic
Cost to try
Small (under $1M)
Years to first evidence
1
Bottlenecks it attacks
- AI that is built but not validated or deployed · Thousands of cancer AI models are published; a handful are in clinical use, and fewer have shown they help patients.