Page 21 - AIH-1-3

Artificial Intelligence in Health Optimizing EHRs to support AI

• Do you have appropriate performance measures and and classifications to be able to provide content related
targets? to a range of concepts and how those concepts impact
• Do you have a way to monitor and calibrate the the requirements of the terminology. One example is
performance of your AI system? the categorial structure of nursing practice, which
53
• How will sensitive data be handled? may also be applicable to represent all types of clinical
54
“…….development and implementation of AI practice requiring the use of their own terminology.
technologies must be undertaken with appropriate The representation of concepts and characteristics need
consultation, transparency, accountability, and regular, to especially be described in this manner for use in
ongoing review to determine its clinical and social formal computer-based concept representation systems.
impact and ensure it continues to benefit, and not Categorial structures also show relationships between
harm, patients, health-care professionals, and the wider categories and sub-categories.
community.” 43 The SNOMED CT terminology is an ontology-based

A number of standards have been developed or are in comprehensive medical terminology used by many
development by the ISO/IEC/JTC1 Standards Committee international members for standardizing the storage,
55
42, to assist all of us to responsibly develop, and make retrieval, and exchange of electronic health data. It is able
use of AI technologies including one for data quality for to represent each data element and identify it together with a
analytics and machine learning, data visualization to assess code. The WHO develops and updates a family of health-care
data quality and an AI data framework. The World Health terminologies including the ICD. Version 11 has an updated
56
Organization (WHO) has recently published its regulatory structure based on the use of ontology-driven tools as one
considerations on AI for health. 44 strategy designed to improve semantic interoperability.
There are numerous relevant standards for data sharing Interoperability among systems requires the
and use, which cover data in general; these standards are harmonization of such models; a project was undertaken
42
not specific for health or clinical care data. A number of by the Office of the National Coordinator for Health
57
guidelines, 45-47 as well as a data governance framework, Information Technology between 2017 and 2019 to
48
have also been developed. Similar initiatives are being advance the utility of observational data for Patient-Centered
undertaken in other jurisdictions. 49,50 All of these measures Outcomes Research (PCOR) and its interoperability across
are designed to improve data quality, streamline our use of multiple networks. The PCOR project resulted in four
data, and support AI development. clinical data models: (1) Sentinel, (2) PCOR Network
(PCORnet), (3) Informatics for Integrating Biology and the
3.1. Data, information, and knowledge management Bedside (i2b2), and (4) Observational Medical Outcomes
requirements Partnership (OMOP). 57
Every known health-related terminology standard is based 3.2. AI data quality objectives
on an agreed categorial structure. Many of these do not
identify as a formal ontology that represents a specific This review has demonstrated that the design of these
knowledge domain, thus resulting in ambiguities. A formal next-generation systems needs to be able to address the
ontology consists of classes, instances, relations, functions, following key requirements to optimize data availability
and axioms to reflect meaning by providing context. for AI development and applications:
This allows for a clear digital understanding of concepts • Ensure we have access and are able to make use of, the
representing a defined knowledge domain. Terminologies maximum number of data points at any required level
were originally structured and developed to suit paper-based of granularity as required to develop reliable accurate
systems. An ontological design determines the knowledge algorithms to suit AI application development.
domain’s structure that enables semantic interoperability. 51 Accessing a maximum possible number of multiple
desired data points needs to be achieved through
In this digital era, it is important for standard
terminologies in use to comply with the ISO standard linking and aggregating health-care data at scale
52
and safely, across different tiers of care and multiple
that specifies how categorial structures of terminologies organizations, using interoperability standards and
need to be represented. The purpose of this ISO standard vendor-neutral data infrastructures.
includes the need to support the development of specific • Maximize automation of routine reporting.
standards of categorial structures for particular health-
care subject fields with the minimum requirements to • Safeguard patient safety and ethical data use.
support meaningful exchange of information. The categorial Every data point represents a single unit of information.
structure approach recognizes the need for terminologies For AI purposes, it is necessary to make use of a defined

Volume 1 Issue 3 (2024) 15 doi: 10.36922/aih.3056