Many healthcare AI systems generate recommendations from patient information available at a particular point in time.
However, patient circumstances are dynamic.
A clinical condition may progress, a medication may become inaccessible, or a patient's living and financial circumstances may change. These developments can affect whether an existing care plan remains appropriate.
The problem with static personalisation
Consider an AI system that supports follow-up for a patient with a chronic condition.
The system schedules appointments, sends reminders, and monitors medication refills.
If the patient relocates or loses access to transportation, the original follow-up pathway may no longer be feasible.
A system that continues sending the same reminders without recognising the change may be executing its workflow correctly while failing to support the patient.
What should adaptive healthcare AI consider?
Systems should distinguish between current and outdated information, identify meaningful changes, recognise patterns that may require reassessment, and establish when previous recommendations should be reviewed.
Relevant information may include changes in clinical status, medication access, functional ability, care preferences, and practical barriers to receiving treatment.
Such information should only be collected and used where appropriate, with privacy and consent safeguards.
Implications for agentic AI
An agentic system should not assume that an active workflow remains appropriate indefinitely.
When new information conflicts with the assumptions behind an existing care plan, the agent should be able to pause relevant actions, flag the change, request updated information, and seek authorised human review.
It should not independently modify consequential clinical decisions beyond its approved authority.
How should these systems be evaluated?
Evaluation should examine how quickly systems recognise meaningful changes, whether outdated recommendations are appropriately reassessed, and whether patients receive timely professional review.
It should also assess false alerts, unnecessary interruptions, privacy implications, and the impact of reassessment on care continuity.
The design principle
Healthcare AI should treat patient information as evolving context rather than a permanent snapshot.
Build systems that know when changing circumstances make an old recommendation worth reviewing.
