This document describes how to identify and address bias in AI systems for health service
organizations during procurement, implementation, and monitoring of these systems used to support
the delivery and allocation of care and preventative services. This includes both clinician-facing, and
patient-facing AI applications provided to patients by their healthcare organization.
This document is focused on bias in AI models, specifically models that exhibit decreased
accuracy or cause differential impact that may unfairly disadvantage certain demographic
groups (also referred to as “protected classes” in some jurisdictions). This includes
algorithmic bias, data representation bias, labelling bias, and deployment-related bias. This
document takes a broad view of bias, acknowledging that unfairness in AI outcomes may
arise from a variety of technical and non-technical sources throughout the AI system
lifecycle.
This document is applicable to all types and sizes of health service organizations (e.g.
hospitals, laboratories, pharmacies, community radiology clinics, and primary or ambulatory
care services), and stakeholders involved in regulating AI use within health service
organizations.
This document does not consider AI system development or manufacturing processes, as
these are expected to be addressed in separate standards (noting that some health service
organizations may be directly involved in the development or manufacture of AI systems).
Similarly, the document does not cover personal bias that may be exhibited by end users of
AI systems (e.g. cognitive or automation bias).
Registration number (WIID)
93147
Scope
This document describes how to identify and address bias in AI systems for health service
organizations during procurement, implementation, and monitoring of these systems used to support
the delivery and allocation of care and preventative services. This includes both clinician-facing, and
patient-facing AI applications provided to patients by their healthcare organization.
This document is focused on bias in AI models, specifically models that exhibit decreased
accuracy or cause differential impact that may unfairly disadvantage certain demographic
groups (also referred to as “protected classes” in some jurisdictions). This includes
algorithmic bias, data representation bias, labelling bias, and deployment-related bias. This
document takes a broad view of bias, acknowledging that unfairness in AI outcomes may
arise from a variety of technical and non-technical sources throughout the AI system
lifecycle.
This document is applicable to all types and sizes of health service organizations (e.g.
hospitals, laboratories, pharmacies, community radiology clinics, and primary or ambulatory
care services), and stakeholders involved in regulating AI use within health service
organizations.
This document does not consider AI system development or manufacturing processes, as
these are expected to be addressed in separate standards (noting that some health service
organizations may be directly involved in the development or manufacture of AI systems).
Similarly, the document does not cover personal bias that may be exhibited by end users of
AI systems (e.g. cognitive or automation bias).
