Page 20 - AIH-1-3

Artificial Intelligence in Health Optimizing EHRs to support AI

information models and standard terminology value sets digital healthcare across Australia. The consortium made
38
to work together to create a coherent data ecosystem. use of universal computable clinical models (Archetypes)
39
40
The most accurate and adaptable method for mapped to SNOMED CT or LOINC, etc., which are utilized
representing computable clinical knowledge is through a in these HL7 FHIR artifacts. The resultant AU core data set
dual information architecture model, which enables the does not specify how and to what extent its elements are
development of clinical information models built from included in FHIR or other exchange standards. SPARKED
common reference components. Some existing strategies represents another small evolutionary step toward
include data sharing through the use of cloud technologies improving data quality. While continuing to make use of
and federated clinical data repositories (CDRs) to provide legacy systems, these new initiatives need to be viewed as
access to large amounts of data. CDRs need to enable reuse transitional arrangements.
of data while preserving the data’s original meaning and 3. Clinical data asset use
context.
This review has identified a number of risk factors to be
Effective data sharing requires a strong data considered when extracting and collating data/information
management strategy and framework including the creation for the purpose of AI use from EHR/EMR systems. The
of standardized, centralized processes around ingesting, New South Wales Government has identified these within
classifying, storing, organizing, linking, and maintaining their comprehensive AI Assurance framework 41,42 informed
data. Centralization and linkage of health data on the cloud by groups of standards developed by the International
raises many security and privacy concerns as well. The use Electrotechnical Commission (IEC)/ISO/and Joint TC
of cloud technologies to store data has the advantage of the (JTC1) family of SDOs. The New South Wales Government
ability to retrieve data using any type of device anytime. strategy includes the following key risks that need to be
A major cultural shift is required to move to externally mitigated. These risks include:
hosted services and the adoption of one set of compatible • The use of incomplete or inaccurate data
standards. CDRs need to be able to support timely health- • Having poorly defined descriptions and indicators of
care delivery, research, and public health initiatives as well “fairness”
as facilitate the creation and efficient implementation of • Not ensuring ongoing monitoring of “Fairness
decision-support tools. Many beneficial advances made Indicators”
to date are not necessarily visible to those providing • Decisions made to exclude outlier data
frontline care. 36
• Using informal or inconsistent data cleansing and
2.4. Continuing use of legacy systems repair protocols and processes
• Using informal bias detection methods
There is a desire to make the best possible use of our legacy • The likelihood that re-running scenarios could
systems to sustain existing profitable business models, produce different results (reproducibility)
to make the best possible use of significant investments • The inadvertent creation of new associations when
made, and to maintain access to historical data. The market linking data and/or metadata
continues to be dominated by a few mega-EMR providers • Differences between the data used for training compared
and numerous other legacy systems who are making their to actual data
own data sharing arrangements, such as the HL7 Argonaut • Missing from this list was not ensuring that scenarios
project, a private-sector initiative designed to advance can be explained, which is a requirement for the
37
industry adoption of open interoperability standards. This generation of trustworthiness (explainability).
represents a small step toward a digital transformation but • Some of the questions to be answered by AI developers
is limited to users of the same enterprise-wide EMR system include:
and its proprietary platform. • Is the data needed for the project in question available
Recent collaboration managed by the Commonwealth and of appropriate quality given the potential harms
Scientific and Industrial Research Organization’s (CSIRO) identified?
Australian e-Health Research Center has resulted in the • Does your data reflect the population that will be
first release of the Australian Core Data for Interoperability impacted by your project or service?
(AUCDI) release for community comment. This collaborative • Have you considered how your AI system will address
consortium set out to build robust HL7 FHIR profiles, issues of diversity and inclusion (including geographic
27
extensions, and terminology value sets and bindings. This diversity)?
consortium’s initiative (SPARKED) has launched a national • Have you considered the impact regarding minority
FHIR Accelerator program to reinforce the move toward and disadvantaged groups?

Volume 1 Issue 3 (2024) 14 doi: 10.36922/aih.3056