1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
159
11 Artificial Intelligence
By Koen Cobbaert, MSc
Artificial General Intelligence (AGI) can
learn incrementally, reason abstractly, and act
effectively over a wide range of domains. As the
author does not anticipate AGI to appear on
the market in the near- or medium-term future,
this chapter focuses on so-called ‘narrow AI,’ i.e.,
artificial intelligence with a specific, rather than a
general, purpose.
Current medical device legislation applies
to machine learning devices. It appears fit to
assure safety and reliable performance, including
AI that continues to change after it is placed
on the market.14 What is lacking is guidance
on how to comply with regulations practically.
Despite a flurry of AI standardization activity,
practical guidance for medical devices is scarce.
This chapter provides an introduction to AI, its
characteristics, and how these impact regulatory
compliance. The chapter concludes with a hori-
zon scan of legislative initiatives that may affect
AI in medical devices.
Definition
Having one common internationally accepted
definition for AI would be helpful when com-
paring AI investments and regulations across
the world. Several definitions of AI exist,15 each
with different flavors. At the time of writing, the
International Medical Device Regulators Forum
(IMDRF) is drafting a definition for machine
learning medical devices.16
In trying to capture AI’s essence, defini-
tions tend to focus on AI’s learning ability or its
intelligent nature, two aspects that pose significant
Introduction
Although different people may understand
artificial intelligence (AI) differently, it has been
a reality in healthcare for decades. Healthcare
has adopted AI technology in medical devices,
workflows, and decision-making processes.
Rather than replacing the human component
of healthcare delivery, artificial intelligence has
become a vital tool or a companion to improve
patient outcomes.
Artificial intelligence refers to a wide variety
of techniques.1 While neural networks are in the
spotlight today, this chapter covers all forms of
AI, including classical AI (e.g., search and opti-
mization techniques), expert systems,2 Hidden
Markov Models,3 and older forms of computer
vision), symbolic AI4 (e.g., logical reasoning5
and decision making), Abstract Syntax Tree6
(AST) modifying code, probabilistic reasoning,7
machine learning,8 knowledge representation,9
planning and navigation, natural language pro-
cessing,10 and perception.11 Hybrid approaches
also exist, which use a combination of techniques,
e.g., neural networks and symbolic AI. The con-
nectionist AI takes care of the messiness and cor-
relations of the real world, for example, to detect
patient position and anatomy, and help convert
those into symbols that a symbolic AI can use to
interact with the patient during physiotherapy.
The influence of the latter will likely increase in
the future.12
See Figure 11-1 for a vastly simplified dia-
gram of the different types of AI.13
CHAPTER
2
3
4
5
6
7
8
9
10
11
12
13
14
15
159
11 Artificial Intelligence
By Koen Cobbaert, MSc
Artificial General Intelligence (AGI) can
learn incrementally, reason abstractly, and act
effectively over a wide range of domains. As the
author does not anticipate AGI to appear on
the market in the near- or medium-term future,
this chapter focuses on so-called ‘narrow AI,’ i.e.,
artificial intelligence with a specific, rather than a
general, purpose.
Current medical device legislation applies
to machine learning devices. It appears fit to
assure safety and reliable performance, including
AI that continues to change after it is placed
on the market.14 What is lacking is guidance
on how to comply with regulations practically.
Despite a flurry of AI standardization activity,
practical guidance for medical devices is scarce.
This chapter provides an introduction to AI, its
characteristics, and how these impact regulatory
compliance. The chapter concludes with a hori-
zon scan of legislative initiatives that may affect
AI in medical devices.
Definition
Having one common internationally accepted
definition for AI would be helpful when com-
paring AI investments and regulations across
the world. Several definitions of AI exist,15 each
with different flavors. At the time of writing, the
International Medical Device Regulators Forum
(IMDRF) is drafting a definition for machine
learning medical devices.16
In trying to capture AI’s essence, defini-
tions tend to focus on AI’s learning ability or its
intelligent nature, two aspects that pose significant
Introduction
Although different people may understand
artificial intelligence (AI) differently, it has been
a reality in healthcare for decades. Healthcare
has adopted AI technology in medical devices,
workflows, and decision-making processes.
Rather than replacing the human component
of healthcare delivery, artificial intelligence has
become a vital tool or a companion to improve
patient outcomes.
Artificial intelligence refers to a wide variety
of techniques.1 While neural networks are in the
spotlight today, this chapter covers all forms of
AI, including classical AI (e.g., search and opti-
mization techniques), expert systems,2 Hidden
Markov Models,3 and older forms of computer
vision), symbolic AI4 (e.g., logical reasoning5
and decision making), Abstract Syntax Tree6
(AST) modifying code, probabilistic reasoning,7
machine learning,8 knowledge representation,9
planning and navigation, natural language pro-
cessing,10 and perception.11 Hybrid approaches
also exist, which use a combination of techniques,
e.g., neural networks and symbolic AI. The con-
nectionist AI takes care of the messiness and cor-
relations of the real world, for example, to detect
patient position and anatomy, and help convert
those into symbols that a symbolic AI can use to
interact with the patient during physiotherapy.
The influence of the latter will likely increase in
the future.12
See Figure 11-1 for a vastly simplified dia-
gram of the different types of AI.13
CHAPTER