February 2020 | ISE Magazine 33
Working in healthcare can cause major men-
tal, physical and organizational stress, which
in turn can lead to clinical hazards. Artificial
intelligence may help, though recent research
on such advanced technology is largely devoted
to patient safety; regulatory bodies conducting
such studies, such as the U.S. Food and Drug Administration
(FDA), focus on improving an AI systems performance purely
from a statistical viewpoint. What’s missing is a need to in-
tegrate AI and cognitive ergonomics to make it more user-
friendly for medical workers.
Cognitive ergonomics – the gathering of knowledge about
human perception, memory and mental processes – has been
neglected in the healthcare domain and the evaluation of AI
from a cognitive ergonomics perspective is not well-estab-
lished. Though the effect of clinicians’ heavy mental load on
patient safety is known, there is no framework to guide the
graphical user interface (GUI) of complex AI systems and their
influence on clinicians’ and patients’ thoughts and feelings.
Here we seek to adopt a systems approach to develop and pro-
pose a conceptual framework across AI, cognitive ergonomics
and patient safety.
The gap between cognitive ergonomics, AI
Cognitive ergonomics is a component of human factors and
ergonomics aimed at ensuring effective interaction between
technology and humans. In this article, we discuss human-
system interaction from a cognitive ergonomics perspective
in healthcare. In this interaction, cognitive ergonomics con-
centrates on mental processes such as thinking, reasoning and
problem-solving, as well as psychological or behavioral inter-
actions. In cognitive ergonomics, these aspects are studied in
the context of work and other systems.
Due to the increasing complexity of healthcare, researchers
are focused on the application of AI primarily in improving
its ability to provide accurate diagnoses. AI systems have long
held the promise of rening the prediction of diseases, such as
guiding imaging for pulmonary embolisms.
A healthcare AI system has two dimensions that impact
cognitive workload: Complexity of the algorithm and user in-
terface. Many researchers have tried to simplify its underlying
algorithm, yet no significant steps have been taken to improve
the AI systems interface. Thus, its impact on human cognition
remains uncertain.
A well-designed graphical interface can help an AI user lo-
cate relevant information, then interpret and prioritize it. The
signicance of cognitive ergonomics and human factors in the
intersection of healthcare and AI has not yet been studied and
cognitive load has been assumed rather than measured. There
is therefore a need to use the methods of cognitive ergonom-
ics in healthcare articial intelligence systems. We propose a
framework to simplify a GUI of healthcare AI systems and
understand its impact on clinicians’ cognitive load.
Meaningful outcome and cognitive load
Those who use advanced AI, such as clinicians, might face
difficulties in understanding and interpreting the outcomes of
the technology. This may be due to the users’ inability to form
a conceptual model of the information presented on a com-
puter screen or a device, and their lack of understanding of
the systems working principle. Addressing confusion with AI
technology and its interface will not only aid in safer applica-
tion but can foster better satisfaction.
Nevertheless, a benchmark for examining the cognitive
load generated by a healthcare AI system is not dened in the
FDAs premarket clearance program. In an early-phase study
AI in healthcare:
Improving human
interface for
patient safety
Better standards needed to make
artificial intelligence user-friendly
for clinicians
By Avishek Choudhury
34 ISE Magazine | www.iise.org/ISEmagazine
AI in healthcare: Improving human interface for patient safety
of 326 hospitalized patients, the FDA approved a pre-
dictive algorithm, WAVE, which indicates vital signs
abnormalities and has led to a reduction in the average
duration of patient instability. Though this was judicious
under current regulatory standards, the approved system
was not tested for the usability and complexity of its GUI.
Is the systems interface simple or intuitive enough for an
inexperienced clinician to implement and understand in a
chaotic environment? It is uncertain whether WAVE has
reduced clinicians’ cognitive load.
The FDA should rigorously conrm and test surrogate
endpoints of new technologies to prohibit the introduc-
tion of AI systems with questionable GUI into a chaotic
healthcare environment where human life is at risk. The
agency must ensure that implementing an AI system not
only improves diagnostic capabilities but also minimizes
a clinicians time spent analyzing and interpreting com-
plex outcomes; such necessary measures could be valu-
able for premarketing authorization. Unfortunately, no
AI algorithms or systems that received regulatory clear-
ance have been tested for their impact on cognitive load.
With the increase of technology in healthcare, clini-
cians and patients are encountering AI devices that involve
complex and unfamiliar GUIs. The industry needs to con-
duct comprehensive and informative research to develop user-
friendly AI devices and systems by considering such design
needs as human memory limitation, perception and attention.
GUI design considerations
Here are aspects of GUI design that should be considered
when approving new AI technology for healthcare use:
Human memory limitation. This partly involves reten-
tion theory and cognitive load theory. The human memory
allows an average person to retain seven (plus or minus two)
sets of information at a time. Thus, to ensure better retention,
health information should be divided into smaller units that
do not exceed the limits of 7 ± 2 per AI system display. Dis-
playing fewer sets of information per screen can reduce clini-
cians’ need to memorize data, which in turn helps minimize
the cognitive load on them and their patients. In addition,
using suitable colors in designing GUI is shown to enhance
information retention by 50%.
Perception. This aspect employs the schema theory and
Gestalt law, illustrated in Figure 1. A schema signifies the con-
ceptual representation of how an individual decodes informa-
tion and extracts contextual knowledge. For instance, when
we see an image showing sun, blue sky and birds flying, we
interpret it as daytime and associate a feeling of happiness. In
addition, users’ prior experience and knowledge plays a cru-
cial role in enhancing their perceiving ability (transference and
mental imagery).
Transference denotes the users’ anticipation of the behavior
of an AI system based on past experience with other computer
interfaces, such as text placement and the location, appearance
and functionality of buttons and icons. Mental imagery refers
to the conceptual representation of how things look.
Gestalt law is explained in the table below; some law of Ge-
stalt complements schema theory as shown in Figure 1.
Attention. In the placement of texts and images, the left-
to-right theory can be applied in designing an effective AI
interface. This suggests allocating critical information on the
top left corner of the screen. This is due to the responsiveness
Laws of Gestalt Meaning
Balance/symmetry Asymmetrical images are perceived as incomplete
Continuation Eyes typically follow a direction as shown in GUI
Closure Shapes that are not fully closed are perceived as
Figure-ground Different colors are perceived as different entities
Focal point Each AI system’s display should have a focal point
Different users perceive a same image differently
Prägnanz Simple GUI design are considered as good forms
Proximity Objects that are clustered are perceived as a group
Similarity Similar objects grouped together attract more
Simplicity Users understand faster if the information is
arrangement of the
Objects are perceived as one group if there is a
similar arrangement of the objects
Schema and Gestalt
A look at the fundamental functioning of human perspective.
February 2020 | ISE Magazine 35
where people tend to interpret data from upper left to lower
right. Additionally, the placement of texts and images should
be applied considering the human visual field, which is di-
vided into two different segments, as shown in Figure 2. Thus,
the placement of texts or images according to humans’ visual
field minimizes the users’ cognitive load.
Clinicians may experience difficulties whenever there are
multiple parallel tasks overwhelming their senses. For exam-
ple, an alarm noise in an emergency floor or a patient talking
in an outpatient setting can be distracting if auditory infor-
mation is also running on the AI system. Such multitasking
should be avoided to reduce their cognitive load. Their sense
of sight might be impaired when background images or texts
are embedded under pieces of informative texts. Thus, AI-en-
abled applications should avoid such embedded content. And
the need to process information simultaneously by two motor
systems increases cognitive load, such as listening to a patients
distress while using a complex AI system.
Trust and meaningful use
As governing bodies decide which downstream features mat-
ter for AI systems, they should also keep in mind these sys-
tems will fail as a technology without acceptance and trust
from clinicians despite having good analytical performance.
Studies have shown that trust is built in a continuous manner,
demanding two-way interactions between the user and the
Initial trust is essential for ensuring the adoption of new
technology. Trust is influenced by a user’s first impression and
is built based on that persons personality and institutional cues.
Once trust is developed, it must be nourished to be sustained.
In our context, continuous trust depends on the functioning
of the AI system in reducing users’ cognitive load and yielding
clinically meaningful outcomes.
It has been believed that trust in technology is determined
by human characteristics (personality and ability), environ-
mental characteristics (culture, task and institutional factors)
and technology characteristics (performance, process and pur-
pose). However, the impact of meaningful use on trust has
been neglected. Meaningful use,
just as it is imposed on the applica-
tion of EHR technology, should be
imposed on AI systems to improve
the quality of care.
The meaningful use of AI sys-
tems means that specific AI should
be implemented and interpreted in a
specic manner. Depending on the
functioning of the AI algorithm, not
all systems can be generalized across
a healthcare system.
For instance, the WAVE platform
algorithm is based on five vital signs: heart rate, respiration,
oxygen saturation, temperature and blood pressure. Since such
measures are common across health systems, the system could
be employed by multiple diverse health systems. However,
other AI-based platforms, especially those based on institu-
tion-specific EHR or image datasets, may not translate across
other EHRs. Moreover, Al trained on specific datasets, such
as patients from a specific institution, may not be generally ap-
plied across broader populations.
Increasing AI interoperability may necessitate developers to
deliver more specified data to conrm that predictive algo-
rithms will achieve reliable, replicable and valid results. In-
deed, regulators should focus on balancing the clarity of pre-
dictive models without impeding the proprietary interests and
intellectual property of algorithm developers.
Healthcare AI systems are just getting evaluated and being
made available for clinical use, so the influence of the exist-
ing regulatory framework on patient outcomes is yet to be
determined. It is also uncertain what impact the 21st Cen-
tury Cures Act passed to relax regulatory standards for low-
risk health technology will have on the value and quality of
predictive algorithms. The FDAs Digital Health Innovation
Action Plan, issued in 2017, launched a precertification pro-
gram to analyze clinical outcomes of AI-based algorithms.
Such efforts should be acclaimed but improved based on our
recommended norms.
Some developers may disparage the overregulation and
standardization of a vaguely understood field. Indeed, a pledge
to regulate healthcare AI systems will emerge over time and
impose financial costs to stakeholders. Policymakers should
also be sensitive to the stability between regulation and in-
novation in this evolving field.
Avishek Choudhury is a Ph.D. student and research assistant in the
School of Systems and Enterprises at Stevens Institute of Technology
in Hoboken, New Jersey. He received his masters degree in industrial
and systems engineering from Texas Tech University. His research in-
terests include healthcare systems, artificial intelligence, clinical decision-
making, patient safety and network science. He is an IISE member.
Human visual field
Where the eyes and the brain connect to process information.
36 ISE Magazine | www.iise.org/ISEmagazine
AI in healthcare: Improving human interface for patient safety
5 examples of how technology is changing healthcare
Artificial intelligence, machine learning, wearable sensors and
virtual reality have taken root in various aspects of industrial and
systems engineering, nowhere more influential than in healthcare.
Here is a sample of the new technological processes in the works:
Seeing through a robot’s eyes helps those
with motor impairments live independently
An interface system using augmented reality technology could help
individuals with profound motor impairments operate a robot to
feed themselves and perform routine personal care tasks such as
scratching an itch and applying skin lotion. The interface displays
a “robot’s eye view” of surroundings to help users interact with the
world through the machine.
The system, described in the journal PLOS ONE, could help
make robots useful to people by using standard assistive computer
access technologies such as eye trackers and head trackers.
“Our results suggest that people with profound motor deficits
can improve their quality of life using robotic body surrogates,”
said Phillip Grice, a recent Georgia Tech Ph.D. graduate and first
author of the paper.
Researchers used a PR2 mobile manipulator manufactured by
Willow Garage, a wheeled robot with two arms and a “head” and
the ability to manipulate objects such as water bottles, washcloths,
hairbrushes and even an electric shaver.
The study made the PR2 available across the internet to 15
participants with severe motor impairments. They learned to
control the robot remotely using their own assistive equipment to
operate a mouse cursor to perform a personal care task. Eighty
percent of the participants were able to manipulate the robot to pick
up a water bottle and bring it to the mouth of a mannequin.
In a second study, researchers provided the PR2 and interface
system to Henry Evans, a California man with limited control of
his body who tested the robot in his home for seven days. He not
only completed tasks, but also devised novel uses combining the
operation of both robot arms at the same time – one to control a
washcloth and the other to use a brush.
“The system was very liberating to me, in that it enabled me to
independently manipulate my environment for the first time since
my stroke,” Evans said.
Source: Georgia Tech
AI can detect 26 skin conditions as accurately
as dermatologists, better than primary care doctors
Skin conditions are among the most common kind of ailment,
resulting in an estimated 25% of all patient treatments worldwide
and up to 37% of clinical patients having at least one skin
With dermatologists facing enormous caseloads and general
practitioners tending to be less accurate in identifying skin
conditions, researchers at Google investigated an AI system
capable of spotting the most common dermatological disorders
seen in primary care. Their paper, “A Deep Learning System for
Differential Diagnosis of Skin Diseases,” reports accuracy across
26 skin conditions when presented with images and metadata
about a patient’s case, which they say is on par with U.S. board-
certified dermatologists.
Google software engineer Yuan Liu and Google Health technical
program manager Dr. Peggy Bui explained that dermatologists
don’t give just one diagnosis for any skin condition but instead
generate a ranked list of possible diagnoses narrowed by lab tests,
imaging, procedures and consultations. The system does the same
by processing inputs that include one or more clinical images of
the skin abnormality and up to 45 types of metadata.
To test the system’s accuracy, researchers compiled diagnoses
from three U.S. board-certified dermatologists and the AI system’s
ranked list of skin conditions achieved 71% and 93% top-1 and
top-3 accuracies, respectively. When the system was compared
with clinicians, the team reported its top three predictions
demonstrated a top-3 diagnostic accuracy of 90%, or comparable
to dermatologists (75%) and “substantially higher” than primary
care physicians (60%) and nurse practitioners (55%).
Source: Google
Photo by Henry Clever/Phillip Grice, Georgia Tech
Henry Evans, a California man who helped Georgia Tech
researchers with improvements to a web-based interface,
uses a robot to shave himself. The team worked to develop
robotic body surrogates to help people with profound motor
deficits interact with the world.
February 2020 | ISE Magazine 37
Photo by Sarah Freeman, University of Georgia
AI is learning to read mammograms
Researchers in the United States and Britain have found that
artificial intelligence can help doctors do a better job of finding
breast cancer on mammograms.
According to an article in the journal Nature, the new system
for reading mammograms is not yet available for widespread
use. Google paid for the study and worked with researchers from
Northwestern University in Chicago and two British medical
centers, Cancer Research Imperial Centre and Royal Surrey County
The system performed better than radiologists in diagnosing
known cases from images. Reviewing scans of 3,000 women in
the U.S., the system produced a 9.4 percent reduction in false
negatives, when cancer is missed, and a 5.7 percent reduction
in false positives, where there is no cancer. A study of 25,000
mammograms from Britain showed that AI reduced false negatives
by 2.7 percent and false positives by 1.2 percent.
About 33 million screening mammograms are performed each
year in the United States. The test misses about 20 percent of
breast cancers, according to the American Cancer Society, and
false positives are common.
“There are many radiologists who are reading mammograms
who make mistakes, some well outside the acceptable margins
of normal human error,” said Dr. Constance Lehman, director of
breast imaging at the Massachusetts General Hospital in Boston.
To train computers to read the mammograms, the authors used
scans from about 76,000 women in Britain and 15,000 in the
United States whose diagnoses were already known.
Sources: New York Times, www.nature.com
Virtual reality could boost flu shot rates
Researchers at the University of Georgia and the Oak Ridge
Associated Universities in Oak Ridge, Tennessee, have conducted
a study using virtual reality simulation to show how flu spreads
and its impact on others as a way to encourage more people to get
vaccinated. It is mostly aimed as a communication tool at the “flu
vaccine avoidant” 18- to 49-year-old adults.
The research, “Using Immersive Virtual Reality to Improve
the Beliefs and Intentions of Influenza Vaccine Avoidant 18- to
49-year-olds,” was published by the journal Vaccine.
The Centers for Disease Control and Prevention reports only
26.9% of 18- to 49-year-olds in the U.S. received a recommended
annual influenza vaccination during the 2017-18 flu season. The
low current acceptance of flu vaccination makes it important to
identify more persuasive ways to educate them.
The 171 participants in the study were divided into four groups,
each exposed to a different form of persuasion: A virtual reality
experience, a video based on the VR platform, an e-pamphlet with
text and images from the video, and a control group provided only
with the CDC recommendations.
The virtual reality group wore headsets that enabled them
to experience events and controllers enabling them to actively
participate in the story. It resulted in participants showing
greater confidence that a vaccination would protect others, more
positive beliefs about flu vaccine and increased intention to get a
Source: University of Georgia
Wireless sensors on patients’ skin tracks health
Stanford engineers have developed wireless sensors that can
detect physiological signals emanating from the skin and beam
wireless readings to a receiver clipped onto clothing.
The sensors can be applied like Band-Aids on various areas. To
demonstrate the technology, researchers stuck sensors on the wrist
and abdomen of one test subject to monitor pulse and respiration
by detecting how the skin stretched and contracted. Sensors on
the elbows and knees tracked arm and leg motions by gauging
tightening or relaxation of the skin when muscles flexed.
Zhenan Bao, the chemical engineering professor whose lab
described the system in an Aug. 15 article in Nature Electronics,
thinks this wearable technology, called BodyNet, will first be
used in medical settings such as monitoring patients with sleep
disorders or heart conditions. The goal is to create an array of
wireless sensors on the skin that work with smart clothing to
accurately track a wide variety of health indicators.
“We think one day it will be possible to create a full-body skin-
sensor array to collect physiological data without interfering with a
person’s normal behavior,” said Bao.
Sources: Stanford University, Nature Electronics
University of Georgia faculty members tested methods of
delivering effective flu vaccination messages through print,
video and virtual reality.