患者身份識別:醫療 AI 的基本安全要求
Patient Identity Resolution: A Fundamental Safety Requirement for Healthcare AI
醫療 AI 整合不同來源的患者紀錄時,必須先確認紀錄是否屬於同一人;身份配對錯誤即使在模型準確處理資料的情況下,也可能導致不安全的建議。可靠流程應採用經核准的識別資料及經驗證的紀錄連結方法,並把模糊個案交由獲授權人員核實。評估應涵蓋錯誤配對、漏配、不同羣體的表現及身份判定錯誤的後果。
Healthcare AI systems depend on accurate patient records. However, integrating information from multiple sources creates an important challenge: establishing that the records being analysed belong to the same individual.
A model can process data accurately and still produce an unsafe recommendation if patient identity is incorrect.
The patient matching problem
Healthcare environments may contain duplicate records, inconsistent identifiers, spelling variations, incomplete demographic information, and different patient identification systems across institutions.
Consider a patient receiving care from two hospitals. Each organisation maintains a separate identifier, and some demographic details differ.
An AI system attempting to integrate these records must determine whether they represent the same person.
An incorrect match can combine unrelated medication histories, allergies, diagnoses, and laboratory results.
What should a reliable system do?
A robust patient identity workflow should use approved identifiers, validated record-linkage methods, appropriate matching thresholds, and processes for resolving ambiguous cases.
It should also retain information about the source of each record and distinguish confirmed identity matches from uncertain associations.
Importantly, probabilistic record matching should not be treated as definitive identity verification in every context.
Implications for agentic AI
An agentic system should verify patient identity before initiating consequential actions.
If identifiers conflict or confidence is insufficient, it should pause the workflow, avoid merging uncertain records, and request authorised human verification.
The system should also maintain an audit trail of the records used and the identity checks performed.
How should healthcare AI be evaluated?
Testing should include duplicate records, conflicting identifiers, demographic variations, cross-institutional matching, and deliberately ambiguous cases.
Evaluation should consider false matches, missed matches, subgroup performance, and the consequences of incorrect identity resolution.
The design principle
Patient identity is not a minor data-cleaning step. It is a foundation for safe clinical intelligence.
Build healthcare AI that verifies whose data it is using before deciding what that data means.
來源:DEV Community · dev.to