Veraltete Daten sind das grosse Risiko bei KI-Agenten im Gesundheitswesen
Ein Healthcare-AI-Agent ruft Patientendaten ab, verarbeitet sie und leitet Aktionen ein. Aendert sich der Datensatz zwischen Abruf und Ausfuehrung, entsteht ein Time-of-Check-zu-Time-of-Use-Problem mit moeglich folgenschweren Entscheidungen.
A healthcare AI agent may retrieve a record, reason over it and initiate an action. The problem is that the record can change between retrieval and execution.
This creates a time-of-check to time-of-use problem. The information was valid when checked, but may no longer be valid when the system acts.
Imagine an agent reads a referral status as pending. Before it submits another referral, a clinician accepts the original request. If the agent does not refresh the status, it may create a duplicate.
Similar problems can arise with medication reconciliation, laboratory corrections, appointment changes and discharge workflows.
Define freshness requirements by data type and action. Different decisions need different freshness thresholds. Avoid applying one universal expiry rule to every clinical record.
Track update times and versions. Timestamps help, but version identifiers and authoritative source information can improve detection of newer or conflicting records.
Revalidate before consequential actions. When a workflow takes time, refresh critical information immediately before execution rather than relying indefinitely on an earlier snapshot.
Handle delayed updates explicitly. Distinguish a confirmed failure from a delayed acknowledgement. If the current state cannot be established, avoid blindly repeating an action.
Escalate unresolved conflicts. If two systems disagree and the difference could affect patient care, the agent should seek verification or human review.
An old diagnosis can remain clinically relevant. A recently entered value can still be incorrect. Freshness assessment should therefore consider provenance, clinical context and whether the information remains appropriate for the decision.
For agentic AI, these checks should influence permissions. A system may be allowed to prepare a recommendation from available data but required to verify current information before executing a high-impact action.
The key principle is simple: validate the information at the point where it matters, not only when it is first retrieved.
If you are building healthcare AI, test delayed updates and changing records as part of your deployment readiness assessment.