IEEE Systems, Man and Cybernetics Magazine - July 2022 - 36

and provide more accurate automated models. Thus,
access to data and codes of previous studies can pave the
way for future research. This is even more important in
medicine because not all researchers have access to real
and high-quality data on various diseases. Unfortunately,
a review of past work done on various medical topics
reveals that access to medical codes and data are almost
impossible in most cases. Most likely, the main reason for
not sharing medical data may be due to ethical commitments
made by medical centers to researchers before
sharing the data. Easy access to medical data is essential,
for which medical centers must provide a fundamental
solution. This will increase access for more active
researchers in AI to provide newer, more accurate, and
standardized AI-based solutions. Hence, we strongly
believe that more public medical databases should be
without patients' personal information.
On the other hand, not sharing codes is entirely up to
AI researchers. This means that researchers have complete
independence to share their codes with others. Nevertheless,
the strategic suggestion is to make code
sharing more common among individuals after getting
final approval for the articles. The point that should not
be overlooked is that code sharing among researchers
has increased in recent years. This may be due to many
reasons, such as requests from various conferences/
journals to publicize data and codes, and ease of access
to various websites to share their codes and data (e.g.,
GitHub and Kaggle). It should be noted, however, that
there are many more websites for this, but we have mentioned
only two. Consequently, sharing codes and medical
databases plays an important role in developing
better CDSSs.
Future Orientations
Although the challenges in using UQ for various AI
methods are significant, new research opportunities
are also equally important (possibly even more so)
and exciting. But it should be noted that reviewing
past or ongoing research in the area of UQ is beyond
the scope of this work (See [16] and [17] for comprehensive
UQ review papers.)
In this sect ion, we
describe a few important open research directions
regarding the application of UQ techniques with AI
methods for medical data analysis, which requires the
contribution of computer scientists, the data science
community, engineers, physicians, the health-care
industry, and clinical sectors.
◆ Application of generative adversarial networks
(GANs) [61]: A lack of access to sufficient medical
data is one big problem in medical research. As mentioned
earlier, researchers do not have access to sufficient
medical data, which may prevent them from
developing accurate and robust AI-based systems
using different ML methods. GANs are novel generative
techniques that create new data samples
36 IEEE SYSTEMS, MAN, & CYBERNETICS MAGAZINE July 2022
resembling the training data. These techniques have
been widely used for different purposes, such as signal
processing [62]; image processing [63], [64]; text
analysis [65]; and so on. Along with these applications,
GANs have shown outstanding performance in
medical applications [66], [67]. There is uncertainty
while GAN generates more medical data, which
needs to be examined.
◆ Application of visual attention mechanisms [68]:
Generally, attention mechanisms assist a network in
weighing features in terms of the importance of one
specified task and then use weights to improve the
performance of the model for the given task. The outstanding
performance of attention mechanisms on
NLP tasks is reported in [69]. However, the outstanding
performance of these attention mechanisms is
well proven, and the importance of their use in vision
and image processing tasks is not negligible [68]. On
the other hand, these methods performed well in
medical areas [70]-[72]. We believe that choosing the
right place of applying attention might be accompanied
by uncertainties. This is very important for medical
topics because of the location of diseases and
cancers. Hence, attention mechanisms need to be
chosen more carefully as the associated uncertainty
of these schemes negatively affects results. Therefore,
UQ methods are excellent techniques for dealing
with this challenge.
◆ Consideration of continual learning [(CL), also
named continual lifelong learning] methods [73]:
Throughout their lifespan, humans have a tremendous
capacity to acquire, fine-tune, and transfer
knowledge and skills continually. It can be said that
CL aims to fill the gap between human learning and
ML. In other words, CL tries to bring ML closer to
humans. Hence, many effective and useful CL methods
have been introduced in medicine (for example,
see [74]). However, regularization and quantification
of uncertainties in CL play a critical role in developing
more reliable results when new class data are
applied. Recently, uncertainty-regularized CL methods,
which yielded promising performance, were
introduced [74]-[76]. A few studies have been conducted
on the uncertainty of CL applied to medical
data [74], [77]. We therefore propose introducing more
uncertainty-based CL methods for medical data.
◆ Application of multimodality (multimodal) medical
data analysis [78]: In multimodal data, data come
from multiple sources, which is common in medical
data. For example, electronic health records are representative
examples of multimodal (multisource)
data collections [79]. Combining features from multiple
medical sources such as images, signals, and clinical
repor
ts
(text) helps obtain an accurate
diagnosis. The uncertainties in predictions increase
while working with different multimodal data and
IEEE Systems, Man and Cybernetics Magazine - July 2022

Table of Contents for the Digital Edition of IEEE Systems, Man and Cybernetics Magazine - July 2022
