Signal Processing - July 2016 - 45

function in near real-time and under computational, memory,
and energy consumption constraints. We provide an overview
of the goals, methodologies, and smartphone apps for health
monitoring to validate whether different areas in this field have
reached the level of maturity required for general public use,
the target population.

Categorical assessment
At the very first step, it is important to determine the legacy
algorithms and open-source codes, such as those from the sensor fusion, signal processing, image processing, and machinelearning fields. But as discussed in the "Mobile Image
Processing Algorithms for Assisted Living" section, many of
these algorithms cannot be used in for assisted-living apps
unless they address the existing challenges, including computational, memory, and energy constraints. Algorithms such
as segmentation, 3-D reconstruction, and machine learning
should be used in many advanced assisted-living apps, but as
they have the highest complexity and cost factors, they will
require major changes before being incorporated into smartphone apps. There has been a movement toward migrating
commonly used algorithms to smartphone platforms (e.g.,
the OpenCV package for smartphones, light-weight neural
networks, and midlevel image classification algorithms), but
more specific algorithms, such as specialized features, may
still require significant adaptation for smartphone use. It is
important to note that this compatibility issue is different from
algorithm performances in terms of accuracy and robustness.
In addition to the underlying algorithms used, it appears that
different categories of health-monitoring apps are at different
development levels, most likely due to differences in the challenges faced. Lifestyle-tracking apps, such as pedometers and
heart-rate-monitoring apps, can be readily used by the public
because they address easier problems and, at the same time,
can tolerate lower accuracies. For example, inconsistencies
in heart rate readings (e.g., due to different finger placement
or motion artifacts) do not substantially affect the usability of
an app. The collected user information has no high-level privacy concerns and can be easily transmitted to and stored on
a remote server using almost any available data link. Generally, these apps can be improved to provide better results in an
uncontrolled environment, but, even at their current level, they
can be used as a component for assisted living.
The challenges escalate as we move toward more advanced
and critical apps, such as chronic wound assessment. These
apps incorporate some level of artificial intelligence to help the
user either gain some degree of expert capability (e.g., generic
assessment of wound condition) or obtain an expert opinion
without the need for and office visit (e.g., through image or
statistics transmission). They perform a deeper analysis of the
captured image or signal, extract relevant features, and then
infer higher-level information through statistical analysis.
Users are empowered with more advanced capabilities compared with the first category of assisted-living apps. To achieve
this, however, the apps need to address the same problems as
the first category (e.g., noise reduction and normalization) but

with a higher level of scrutiny. In addition, they face problems
that are not present or are negligible for the first category,
including estimating 2-D or even 3-D reconstruction, feature
extraction in the presence of noise, and robust and customized
classification using both local and contextual information.
In the case of wound assessment, for example, the color and
illumination normalization is itself an unresolved challenge.
Obtaining basic and yet vital features such as wound dimensions can be obtained only by investing substantial computational, memory, and energy; external markers (of specific size
and color) still need to be placed near the wound for camera
calibration and illumination normalization.
These issues prevent many apps of this type from using
more advanced features such as 3-D structure or video processing of the wound. In addition, contextual features [e.g.,
the periwound condition, exudation, and history information
(e.g., history of diabetes)] that are routinely used by experts
for wound grading have not been implemented in any existing
system, regardless of platform. These challenges are the reason
that the existing state of the art is used only by medical experts
to assist them (and not replace them) in low-level tasks, e.g.,
wound measurements. Functionalities such as tissue classification have severe shortcomings that can mislead even experts.
We conclude that many of the basic challenges facing this
second category of health-monitoring apps and other assistedliving apps at the same level have been solved. However, these
apps still have more advanced issues that need to be addressed
before use by a larger target population.
Finally, the third and most advanced smartphone assistedliving app category targets problems that still exist without
smartphone platform restrictions. It aims to address challenges that no definitive approach has yet resolved. The
underlying problem for this category of apps is particularly
sensitive and demands a solution with high accuracy. Due to
the sensitivity (e.g., privacy or security issues) of this problem, it faces additional issues, such as a lack of proper and
standardized test beds or availability of a robust and discriminative feature. Example of these problems are visible in the
smartphone-based skin cancer diagnosis systems, reviewed
in the "Chronic Wound Assessment" section. An acceptable
skin cancer diagnosis system requires near-expert accuracy to
avoid endangering the user or misleading the expert. To reach
that level of accuracy, different modalities of data should be
combined and evaluated, and the system should be extensively
validated on medically proven cases. But our review shows
the contrary. There is a significant variation in the quality,
modality, and number of images used in different proposed
works. The numbers of images range from fewer than 100 to
more than 12,000, making the results incomparable. Moreover, many apps use dermatoscopic images in their process
when they are targeted for smartphone. This divergence without visible and verified model transfer is one of the limitations
that undermine the existing works.
In addition, processes used for preprocessing, feature extraction, and classification are not still accurate enough to address
the problem at hand. There seem to be no substantial color and

IEEE SIgnal ProcESSIng MagazInE

July 2016

Table of Contents for the Digital Edition of Signal Processing - July 2016