162
Chapter 11: Artificial Intelligence
model-server
worker-a worker-b worker-c
Step 1
central server chooses
a statistical model
to be trained
model-server
worker-a worker-b worker-c
model sync
Step 2
central server transmits
the initial model to
several nodes
model-server
worker-a worker-b worker-c
Step 3
nodes train the model
locally with their own data
model-server
average
upload
worker-a worker-b worker-c
Step 4
central server pools model results
and generates one global model
without accessing any data
Pre-Determined Change
Local change occurs in machine learning devices
that adapt, refine their performance, or calibrate
themselves to a specific patient or healthcare
setting’s characteristics. An AI that learns and
changes itself during clinical use requires manu-
facturers to determine the limits for such change
to occur safely while assuring performance for
the intended purpose. Manufacturers should
establish those boundaries of change before
placing the product on the market. Set boundar-
ies determine the framework in which regulatory
approval allows for changes.
Pre-determined change is not unique to
machine learning devices. Many medical devices
adapt to the patient or their environment or
compensate for wear and tear by reconfiguring
or self-healing their design. For example, a CT
machine undergoes regular calibration cycles to
adjust for wear and tear. The calibration recon-
figures the CT software to compensate for, and
adapt to, hardware changes.
Pre-determined changes through machine
learning personalized healthcare include a mod-
ification of the AI’s ‘working point’ based on the
local/patient environment it allows AI to maxi-
mize performance for a given patient or situation
and enables the device to adapt to the patient
rather than force them to adapt to the device. For
example, AI used in the joints of a bionic foot
allows the kinetics to be adapted to the patient,
rather than letting the patient adapt their gait to
the prosthesis’s kinetics.
Change Dynamics
Depending on the device, the manufacturer, the
health institution, the caregiver, the patient, or
a combination of these, can control the change.
The actor responsible for the change and whether
the change occurs before or after the device was
placed on the market brings forth different regu-
latory implications (see Figure 11-4).
In the first scenario, the AI is locked, and
the manufacturer controls the learning. ‘Locked
AI’ does not change its design during runtime.
Examples of locked AI are static lookup tables,
decision trees, and complex classifiers. Locked
AI generally provides the same result each time
the same input is applied. However, there are
some exemptions to this, e.g., if the AI contains
non-deterministic20 algorithms, or the user can
change the working point on the operating curve.
The performance of a ‘locked’ AI can therefore
still change. Consider an algorithm to screen for
tuberculosis at airports. While the algorithm’s
design is locked, the user may still choose a point
Figure 11-3. Visualization of Federated Learning
Chapter 11: Artificial Intelligence
model-server
worker-a worker-b worker-c
Step 1
central server chooses
a statistical model
to be trained
model-server
worker-a worker-b worker-c
model sync
Step 2
central server transmits
the initial model to
several nodes
model-server
worker-a worker-b worker-c
Step 3
nodes train the model
locally with their own data
model-server
average
upload
worker-a worker-b worker-c
Step 4
central server pools model results
and generates one global model
without accessing any data
Pre-Determined Change
Local change occurs in machine learning devices
that adapt, refine their performance, or calibrate
themselves to a specific patient or healthcare
setting’s characteristics. An AI that learns and
changes itself during clinical use requires manu-
facturers to determine the limits for such change
to occur safely while assuring performance for
the intended purpose. Manufacturers should
establish those boundaries of change before
placing the product on the market. Set boundar-
ies determine the framework in which regulatory
approval allows for changes.
Pre-determined change is not unique to
machine learning devices. Many medical devices
adapt to the patient or their environment or
compensate for wear and tear by reconfiguring
or self-healing their design. For example, a CT
machine undergoes regular calibration cycles to
adjust for wear and tear. The calibration recon-
figures the CT software to compensate for, and
adapt to, hardware changes.
Pre-determined changes through machine
learning personalized healthcare include a mod-
ification of the AI’s ‘working point’ based on the
local/patient environment it allows AI to maxi-
mize performance for a given patient or situation
and enables the device to adapt to the patient
rather than force them to adapt to the device. For
example, AI used in the joints of a bionic foot
allows the kinetics to be adapted to the patient,
rather than letting the patient adapt their gait to
the prosthesis’s kinetics.
Change Dynamics
Depending on the device, the manufacturer, the
health institution, the caregiver, the patient, or
a combination of these, can control the change.
The actor responsible for the change and whether
the change occurs before or after the device was
placed on the market brings forth different regu-
latory implications (see Figure 11-4).
In the first scenario, the AI is locked, and
the manufacturer controls the learning. ‘Locked
AI’ does not change its design during runtime.
Examples of locked AI are static lookup tables,
decision trees, and complex classifiers. Locked
AI generally provides the same result each time
the same input is applied. However, there are
some exemptions to this, e.g., if the AI contains
non-deterministic20 algorithms, or the user can
change the working point on the operating curve.
The performance of a ‘locked’ AI can therefore
still change. Consider an algorithm to screen for
tuberculosis at airports. While the algorithm’s
design is locked, the user may still choose a point
Figure 11-3. Visualization of Federated Learning