Chapter 8. Artificial Intelligence-based Software
126 Regulatory Affairs Professionals Society (RAPS)
Design and Development
Similarly, when designing a software product,
companies should consider a systematic and
methodical approach. Models, methods, archi-
tecture, and design-modelling techniques should
be used that are appropriate for product develop-
ment based on the intended use and the patient
population to which the device is being applied.67
Successful AI/ML medical device development
includes establishing a process governing data
curation, data quality control, data annotation,
algorithm training, validation, and final product
deployment.
The first step in development of AI/ML
based SaMD is the data curation process that
involves defining the scope or use case of the
project, identifying the data requirements estab-
lishing a baseline, and data collection methods.
Curated data must be properly labelled and
organized. Structured and unstructured data
should be organized in a manner that allows for
traceability to the data collected. When organiz-
ing and labeling, the data developers should work
closely with physicians and other clinical end
users to ensure that the data annotation reflects
clinical knowledge. A critical step in the initial
phase of model development is ensuring that the
quality and quantity of data are sufficiently high
to train the algorithm successfully. This is the
most time-consuming step in AI/ ML develop-
ment and should follow a governance process.
The curated data should be separated into inde-
pendent training and testing datasets.
Training, tuning, and testing data
Training, tuning, and independent testing of data
are crucial steps in any AI software development
lifecycle. Each step has a specific purpose and
requires collaboration of interdisciplinary teams
to work closely with different methods and avail-
able scientific tools. Images from a single medical
center may be insufficient to train a model for
a given task or may be biased because of the
sampled population. Multicenter datasets help to
address this issue but could also introduce other
challenges related to standardization.
Training is the process of feeding the algo-
rithm with data so it can learn specific patterns
and behavior. The accuracy and performance of
the algorithm depends primarily on the qual-
ity and quantity of data used in the training
(put colloquially, garbage in, garbage out). To
demonstrate the generalizability of an algorithm
a diverse representation of real-world scenarios
should be considered so that the algorithm can
learn and perform better.
Tuning or fine tuning is the process of
optimizing a model to improve its accuracy and
reliability. This step includes adjusting the parame-
ters, such as hidden layers, within a neural network.
Note that the performance of a deep learning
model on training data does not predict its ability
to generalize to cases it has never seen before.
Training is a highly iterative process
requiring repeated experiments to identify
optimal settings for the device. Training is the
process of evaluating the performance of the
algorithm. Training should always be performed
Table 8-4. Good Machine Learning Practice for Medical Device Development: Guiding Principles
Multi-disciplinary expertise is leveraged throughout the
TPLC
Good software engineering and security practices are
implemented
Clinical study participants and data sets are
representative of the intended patient population
Training data sets are independent of test sets
Selected reference datasets are based upon best
available methods
Model design is tailored to the available data and
reflects the intended use of the device
Focus is placed on the performance of the human-
artificial intelligence team
Testing demonstrates device performance during
clinically relevant conditions
Users are provided clear, essential information Deployed models are monitored for performance and
re-training risks are managed
TPLC, Total Product Life Cycle
126 Regulatory Affairs Professionals Society (RAPS)
Design and Development
Similarly, when designing a software product,
companies should consider a systematic and
methodical approach. Models, methods, archi-
tecture, and design-modelling techniques should
be used that are appropriate for product develop-
ment based on the intended use and the patient
population to which the device is being applied.67
Successful AI/ML medical device development
includes establishing a process governing data
curation, data quality control, data annotation,
algorithm training, validation, and final product
deployment.
The first step in development of AI/ML
based SaMD is the data curation process that
involves defining the scope or use case of the
project, identifying the data requirements estab-
lishing a baseline, and data collection methods.
Curated data must be properly labelled and
organized. Structured and unstructured data
should be organized in a manner that allows for
traceability to the data collected. When organiz-
ing and labeling, the data developers should work
closely with physicians and other clinical end
users to ensure that the data annotation reflects
clinical knowledge. A critical step in the initial
phase of model development is ensuring that the
quality and quantity of data are sufficiently high
to train the algorithm successfully. This is the
most time-consuming step in AI/ ML develop-
ment and should follow a governance process.
The curated data should be separated into inde-
pendent training and testing datasets.
Training, tuning, and testing data
Training, tuning, and independent testing of data
are crucial steps in any AI software development
lifecycle. Each step has a specific purpose and
requires collaboration of interdisciplinary teams
to work closely with different methods and avail-
able scientific tools. Images from a single medical
center may be insufficient to train a model for
a given task or may be biased because of the
sampled population. Multicenter datasets help to
address this issue but could also introduce other
challenges related to standardization.
Training is the process of feeding the algo-
rithm with data so it can learn specific patterns
and behavior. The accuracy and performance of
the algorithm depends primarily on the qual-
ity and quantity of data used in the training
(put colloquially, garbage in, garbage out). To
demonstrate the generalizability of an algorithm
a diverse representation of real-world scenarios
should be considered so that the algorithm can
learn and perform better.
Tuning or fine tuning is the process of
optimizing a model to improve its accuracy and
reliability. This step includes adjusting the parame-
ters, such as hidden layers, within a neural network.
Note that the performance of a deep learning
model on training data does not predict its ability
to generalize to cases it has never seen before.
Training is a highly iterative process
requiring repeated experiments to identify
optimal settings for the device. Training is the
process of evaluating the performance of the
algorithm. Training should always be performed
Table 8-4. Good Machine Learning Practice for Medical Device Development: Guiding Principles
Multi-disciplinary expertise is leveraged throughout the
TPLC
Good software engineering and security practices are
implemented
Clinical study participants and data sets are
representative of the intended patient population
Training data sets are independent of test sets
Selected reference datasets are based upon best
available methods
Model design is tailored to the available data and
reflects the intended use of the device
Focus is placed on the performance of the human-
artificial intelligence team
Testing demonstrates device performance during
clinically relevant conditions
Users are provided clear, essential information Deployed models are monitored for performance and
re-training risks are managed
TPLC, Total Product Life Cycle