From X-Rays to AI: Navigating US Regulations in Radiological Health
131
Recent Developments and Future
Perspectives
FDA guiding principles for Good Machine
Learning Practice
As AI/ML is becoming increasingly popu-
lar in medical device software development,
it is important to have a framework of key
considerations or guiding principles to help
manufacturers produce safe, effective, and
high-quality SaMDs.
As previously discussed in the Key Design
Considerations for Successful AI/ML SaMD
Development section of this chapter, such a
framework has been proposed and is outlined in
Table 8-4.68
The 10 guiding principles proposed cover
the development and deployment of SaMD AI/
ML models into the clinical workflow while
considering the potential risks and benefits to
patients. Pointers are given to manufactures
regarding data collection, management, and
quality assurance practices. The importance of
considering the best available methods while
selecting a reference standard are reiterated, as
well as the need to focus on the AI/ML model’s
impact on both the global and local performance.
The good machine learning practice pointers also
suggest that manufacturers consider performance
degradation over time and encourage companies
to obtain feedback from real-world performance.
Development of the FDA Digital Health
Center of Excellence and Ongoing FDA
Activities
The FDA has recognized the vital role that AI/
ML can play in software-based medical devices
and has developed strategies for a regulatory
framework to advance the development of digital
health products.81 Table 8-6 provides a sum-
mary of key FDA activities in the digital health
space from 2013 onwards, including the incep-
tion of the Digital Health Center of Excellence
(DHCoE) in 2020. 6,17-20,32,36,68,74,81-95
In December 2013, the International
Medical Device Regulatory Forum (IMDRF)
formed the Software as a Medical Device
Working Group to develop guidance for
manufacturers to support innovation and timely
access to safe and effective SaMD globally.17-19
The working group developed guidance around
the definitions to use, framework for risk catego-
rization, quality management system, and clinical
evaluation of AI/ML technologies.
The Association for the Advancement of
Medical Instrumentation (AAMI) has published
a consensus report for identifying, evaluating, and
managing the risk of healthcare technology that
incorporates AI or ML.92 This report responds
to an urgent, immediate need to understand how
to apply risk management principles to SaMD
development and its life cycle.92
The Future – Adaptive Learning, Global
Harmonization
The full potential of AI/ML SaMD technologies
has yet to be uncovered as the use of AI contin-
ues to grow in medical devices. The recent White
House AI Bill of Rights91 suggests that in the
coming years, health authorities will focus on the
development of safe and effective systems algo-
rithmic discrimination protections data privacy
notice and explanation and human alternatives,
consideration, and fallback.90 Moreover, the release
of recent FDA guidance documents on clinical
decision support software,76 recommendations on
predetermined change control plans,81 software
device functions,94 and the creation of the NIST
Public Working Group on Generative AI by White
House95 suggest a significant federal investment in
AI/ML SaMD regulation and governance.
To date, the majority of AI/ML algorithms
that have received FDA marketing authorization
are locked algorithms. As more developers and
manufacturers attempt to optimize algorithm
performance over time by continuously training
models on real world data, the FDA appears to be
focusing on streamlining reviews related to regu-
lating adaptive algorithms. The direction towards
adaptive algorithms is clear from the recent De
Novo marketing authorization DEN220063
of the Caption Interpretation Automated
Ejection Fraction software (Caption Health,
Inc., San Mateo, CA)33 and the creation of the
associated new regulation (21 CFR §892.2055)
for Radiological Machine Learning Based
Quantitative Imaging Software with Change
131
Recent Developments and Future
Perspectives
FDA guiding principles for Good Machine
Learning Practice
As AI/ML is becoming increasingly popu-
lar in medical device software development,
it is important to have a framework of key
considerations or guiding principles to help
manufacturers produce safe, effective, and
high-quality SaMDs.
As previously discussed in the Key Design
Considerations for Successful AI/ML SaMD
Development section of this chapter, such a
framework has been proposed and is outlined in
Table 8-4.68
The 10 guiding principles proposed cover
the development and deployment of SaMD AI/
ML models into the clinical workflow while
considering the potential risks and benefits to
patients. Pointers are given to manufactures
regarding data collection, management, and
quality assurance practices. The importance of
considering the best available methods while
selecting a reference standard are reiterated, as
well as the need to focus on the AI/ML model’s
impact on both the global and local performance.
The good machine learning practice pointers also
suggest that manufacturers consider performance
degradation over time and encourage companies
to obtain feedback from real-world performance.
Development of the FDA Digital Health
Center of Excellence and Ongoing FDA
Activities
The FDA has recognized the vital role that AI/
ML can play in software-based medical devices
and has developed strategies for a regulatory
framework to advance the development of digital
health products.81 Table 8-6 provides a sum-
mary of key FDA activities in the digital health
space from 2013 onwards, including the incep-
tion of the Digital Health Center of Excellence
(DHCoE) in 2020. 6,17-20,32,36,68,74,81-95
In December 2013, the International
Medical Device Regulatory Forum (IMDRF)
formed the Software as a Medical Device
Working Group to develop guidance for
manufacturers to support innovation and timely
access to safe and effective SaMD globally.17-19
The working group developed guidance around
the definitions to use, framework for risk catego-
rization, quality management system, and clinical
evaluation of AI/ML technologies.
The Association for the Advancement of
Medical Instrumentation (AAMI) has published
a consensus report for identifying, evaluating, and
managing the risk of healthcare technology that
incorporates AI or ML.92 This report responds
to an urgent, immediate need to understand how
to apply risk management principles to SaMD
development and its life cycle.92
The Future – Adaptive Learning, Global
Harmonization
The full potential of AI/ML SaMD technologies
has yet to be uncovered as the use of AI contin-
ues to grow in medical devices. The recent White
House AI Bill of Rights91 suggests that in the
coming years, health authorities will focus on the
development of safe and effective systems algo-
rithmic discrimination protections data privacy
notice and explanation and human alternatives,
consideration, and fallback.90 Moreover, the release
of recent FDA guidance documents on clinical
decision support software,76 recommendations on
predetermined change control plans,81 software
device functions,94 and the creation of the NIST
Public Working Group on Generative AI by White
House95 suggest a significant federal investment in
AI/ML SaMD regulation and governance.
To date, the majority of AI/ML algorithms
that have received FDA marketing authorization
are locked algorithms. As more developers and
manufacturers attempt to optimize algorithm
performance over time by continuously training
models on real world data, the FDA appears to be
focusing on streamlining reviews related to regu-
lating adaptive algorithms. The direction towards
adaptive algorithms is clear from the recent De
Novo marketing authorization DEN220063
of the Caption Interpretation Automated
Ejection Fraction software (Caption Health,
Inc., San Mateo, CA)33 and the creation of the
associated new regulation (21 CFR §892.2055)
for Radiological Machine Learning Based
Quantitative Imaging Software with Change