Wearable Technolog Robert Janikowski 2017 (PDF)

Wearable Technology Challenges and Adoption in
Health Care
Robert Janikowski
June 28 2017

1

Introduction

Wearable technology has recently made a large impact on the fitness market. The question is whether
wearable technology will do the same to the health industry. A lot of research and development has gone
into new types of sensors to expand beyond positional
data. Most of these new sensors attempt to give more
insight into a user’s health which is why wearable
technology may completely overhaul the health care
industry.
This paper will look at the potential for wearable
technology adoption in health care. It will look at
each component of the basic model presented in Figure 1 and how it pertains to health care. Sensors will
be explored initially as they are the basis of all wearable data. This will be followed by an insight into
machine learning which changes raw sensor data into
something more meaningful. Finishing up, this paper will look at how data security, storage and cloud
technology may help or hinder health care. All of
these sections will mention modern challenges associated with adopting some of these technologies. Keep
in mind that this paper is a general overview of complex and expandable topics.

2

Figure 1: Simple model of typical data movement in
wearable technology
burned.
The problem with positional information is the
limit of its uses. Positional data is sufficient enough
for sports medicine application[4] but is limited in
other areas of medicine. The addition of sensors that
can measure health related data has been an increasing focus in the wearable technology sector. With
over 30 percent of the 30 billion sensor market being
composed of new sensors by 2025, we can see there
is an insatiable demand for new data types[9]. It
is projected that stretch/pressure sensors as well as
chemical sensors will experience the largest growth as
seen in Figure 2. This trend is to be expected as inertial movement unit sensors currently hold the largest
market share.

Sensors

Sensors are the basis in which wearable tech and mobile devices currently acquire data. The most current
iPhone 7 has a total of 6 sensors listed in its technical
specifications[6]. Three of these sensors are related to
The market is expected to be heavily dominated
gathering positional data. Some applications involvby chemical sensors by 2020 according to Figure 3.
ing this positional data include: fitness tracking, gps,
Most chemical sensor data would likely be used for
location services and emulated properties like calories
1

2.1

health monitoring. The issue with chemical sensors
is their potential to produce an invasive experience.
User’s would likely have to wash their sweat sensor
every once in a while to mitigate contamination. This
may seem like a trivial issue but recent sensors on
mainstream devices have really been pushing for non
invasive-experiences. A typical consumer does not
know in a definite sense what sensors their device
possesses, and instead know the functions the device should perform. Introducing chemical sensors to
a market that boasts “seemless” functionality could
prove to be problematic.

Sensor Applications

New types of sensors emerging allows more diverse
applications of wearable technology, applicable to the
health care sector. This section will go through some
of the new types of sensor players on the market.
Heart Monitors
The Apple Watch 2 can monitor heart rate through
a non-invasive technique called
photo-plethys-mography[7]. The watch shines a green
LED onto the user’s skin which then measures how
much light is absorbed by red blood cells.
Glucose Tracking
Apple is currently developing a non-invasive glucose
tracker. Not much is known at this point as to the
accuracy or effectiveness.
Sweat Sensor
Researchers developed a non-invasive sweat sensing
glucose tracker[3], which notifies a user when to take
medication as well as when to drink water.
Emotion Monitors

Figure 2: Sensor Type Growth by 2025[9]

Sence is a consumer wrist wearable that claims to
monitor up to 40 emotions through
electro-cardio-graphy[18]. Mental health may be better remedied with the ability to monitor emotions.
Brain Computer Interface (BCI)
Head wearable devices that use an electroencephalogram (EEG) to measure electrical activity in the brain.
Consumer grade devices currently lack the accuracy
of medical devices but are promising in measuring
cognitive states[16].
Gesture Camera
Shoe Sense uses gesture sensing cameras to detect
a greater variety of gestures to help improve interaction with wearable technology. However the gestures not only need to be practical but also socially
acceptable[15].

Figure 3: Sensor Market Size by 2020[9]

2

2.2

Why expand sensors?

could allow for highly usable data in a non clinical
setting.

The benefits of adding more varied types of sensors
onto wearables can have huge implications on the
health industry. One of the fastest growing epidemics
in the United States is diabetes. It is estimated that
14.3% of the population has diabetes, with 9.3% undiagnosed and 38% in prediabetic status[11]. Apple
is currently working on a non-invasive glucose tracker
which could have a major impact on the diabetes epidemic. Mass adoption of glucose tracking capability
could decrease the large amount of undiagnosed diabetics. Those in prediabetic status would have the
ability to better monitor and be more aware of their
glucose levels in a non-invasive fashion.

3.1

Medical Applications of
Machine Learning

Early Diagnosis of Fatal Conditions
Many people (33% in the United States[2]) avoid regular medical care checkups for many reasons which
can be detrimental to their health. The second largest
reason for avoiding the doctor’s office is low perceived
need at 12.2%[2]. The solution to this problem could
be wearables that use machine learning to accurately
identify and notify users about their condition. If a
user can be notified that there may be a problem,
then that would increase the perceived need for medical attention. Almost 40% of the American population is in prediabetic status[11]. Having the opportunity to be proactive and keep their diabetic conditions
in check, could save billions of dollars for the American health care industry[11].
One study was performed at the University of California that tested the accuracy of detecting atrial
fibrillation with the Apple Watch 2. It utilized machine learning to train and recognize certain heart
rate patterns as seen in Figure 5. The study concluded to find it was 97% accurate in detecting an
paroxysmal atrial fibrillation[17]. If such a high accuracy can be expected, it is likely health care professionals will adopt this preventative technology.

Figure 4: What can consumer wearables do?[1]

3

Machine Learning

Sensor data has almost no intrinsic value without the
aid of machine learning. Wearables especially will experience a lot of noise in data collection. Machine
learning will often employ a training set and test Figure 5: Atrial Fibrillation compared with normal
set and use cross validation (CV) to test prediction heart rhythm[17].
accuracy[10]. This method goes beyond traditional
algorithms as the program is able to train itself to
recognize patterns much like a human would[12]. Filtering out data intelligently much like a doctor would,

3

In Home Clinical Data

in-home falls of Parkinsons patients[21]. These methods are the two simplest forms of machine learning
Clinical trials take a lot of space, time and are usually
used in the study, yet they were most accurate. Simlimited to small sample sizes[12]. Allowing patients
pler methods of machine learning also better adapt
like those suffering from Parkinson’s Disease to go
to data drift[12].Spending time to determine the best
home with wearables provides a larger sample size
machine learning method may simplify the work rewhen testing for drug titration and timing. Having
quired and in the end produce more accurate results.
access to a larger volume of data that has been filGoing forward, researchers may have to test multiple
tered through machine learning can greatly benefit
machine learning methods in an ordered list fashion
the medical research sector.
to ensure reproducible results.
Mental health research could also benefit greatly
from wearable data. One of the issues with mental
health research is the lack of data and clinical trials.
Not only is it currently hard to diagnose someone
with a mental health disorder, it can also be very
time consuming[5]. Instead of a patient attending
multiple appointments with a mental health professional, emotional data could be collected to aid in a
proper diagnosis.

3.2

Misuse and Reproducible Data
More tools are being released to allow machine learning to become more and more accessible. Large companies like Apple, Facebook and Google have released
frameworks where a user can use machine learning
with ”a few lines of code”. Although this is a step
in the right direction it opens up avenues for misuse
where the data produced is not accurate. Mainstream
machine learning can make data become a black box
in which the untrained user may incorrectly believe
their trained algorithm is accurate enough for industry
standards[10].

Challenges in Machine Learning

Reducing Dimensionality
Wearables tend to have more than one sensor, often
times with different types of data. The dimension
of a wearable represents the degree of its data complexity. As wearable technology progresses and more
sensors are added, dimensions will need to be simplified as much as possible. Using the least amount
of dimensions possible could produce higher fidelity
data[14].

Many studies use one of two cross validation(CV)
techniques to test the accuracy of their algorithms,
record-wise CV or subject-wise CV. However, recordwise CV tends to overestimate the prediction of algorithms and is employed in 45% of studies using wearable sensors, smart phones and accelerometers[10].
This may result in many studies obtaining data that
is not reproducible which can be problematic to reFitting Data
searchers and data scientists. Figure 6 shows a deOverfitted data can be a problem in machine learn- piction of how subject-wise CV is much more reliable
ing. Machine learning often deals with filtering out over record-wise CV in displaying classification error.
noise and using an overfitted model with more parameters than necessary can pick up extra artifacts.
Cloud and Security
An underfitted model may be too simple to predict 4
with accurate results[12].

4.1

The complexity of a problem often needs to be
considered and correctly determined, this is referred
to as complexity control. From six various machine
learning methods, nearest neighbour classifier and
least squares method were most accurate in detecting

Vulnerability

The expanding market of wireless wearables is resulting in more wireless communication. A higher volume
of data will be vulnerable and exposed to attacks and
therefore it is important to look into any counter measures in place. Currently the most common method
of wireless data transfer is Bluetooth. Its low energy

4

Research has been employed to improve Bluetooth
security. An adaptive layer protocol has been proposed to compartmentalize the Bluetooth data transfer process. The idea is to “prevent adversaries from
achieving ‘breakthrough’ security measures by just
defeating one particular security mechanism in the
system”[8]. This method seems promising but attackers are sometimes able to retrieve data before it
is even encrypted[13].
For ethical reasons, further research into data protection should be continued. However we must consider the data producers and how much interest they
have in maintaining high data integrity. A study explored how concerned people are with the disclosure
of their behaviours through wearable sensor data. It
found the main concerns to be disclosure of converFigure 6: Subject-wise CV vs Record-wise CV Classations and reporting stress/psychological state[14].
sification Error
Nonetheless if the user did not have a personal stake
in the compromised data, they would not understand
the probable threat. People understand the significance of losing their credit card number and how it
may hinder them. Sensor data on the other hand
resembles more of a black box to the average consumer. The ramifications of losing sensor data are
not as clear as losing one’s credit card.

Figure 7: Main security concern for single-person
entertainment[8].
Figure 8: Results Summary of Bluetooth Data Hacking attempts[13]

utilization is especially valuable to wearable technology due to limited battery power. However it has
inherent security risks that are surprisingly alarming(shown in Figure 7). Tests on popular consumer
wearables like the Fitbit, Jawbone and the Pebble
found that Bluetooth data could be hijacked relatively easily. In Figure 8 we can see that data could
be stolen on the Fitbit 70% of the time. The Jawbone and Pebble are not much further behind at 40%
and 60% respectively[13]. Evidently, the Bluetooth
platform which all of these consumer devices share,
is not resistant to attacks.

4.2

Data Migration + Storage

A lot of data from modern technology is either stored
in the cloud or calculated there. Wearables often
don’t have their own computing capability and must
send their sensor data to either a mobile phone or the
cloud. An example of a cloud in health care would be
the electronic health record (EHR). Outpatient care
has the potential to become a greater industry with
the aid of consumer wearables that communicate with
5

Figure 9: Proposed Health Care Cloud Architecture[19].
less beneficial and in this case detrimental. One study
suggested a standardized Cloud architecture shown
The adoption rate for wearables in the health care in Figure 9. This is the correct approach in solving
environment seems to be extremely slow. Many bar- the problem. Medical staff would prefer dealing with
riers are currently holding back the wide adoption of architectures which are not constantly changing, eswearable outpatient care but one of the greatest chal- pecially if it leads to data inaccuracy[20].
lenges is standardization. Wearables often have their
own formats in storing data. This creates a problem when you try to bring high dimension data toConclusion
gether to create intrinsic value for the user or health 5
care professional. Companies tend to prefer keeping
This paper explored wearable technology and how it
all aspects regarding their product within their own
pertains to health care today. Specifically, the chalecosystem. Apple tried to remedy this by introduclenges of adopting wearable technology in outpatient
ing Health Kit which was suppose to bridge the gap
care. Three major aspects of wearable technology
between apps and data collection. However FitBit a
were assessed: sensors, machine learning and cloud
major fitness wearable brand, decided not to partictechnology. Sensor expansion research looks promisipate and not allowing their data to be compatible
ing in improving the monitoring of physical and menwith Apple’s Health Kit. This can be problematic as
tal health capacities. Machine learning can be very
it segregates consumers.
accurate in diagnosing medical problems from an inhome setting, but requires a good understanding of
Businesses often struggle with data migration. The
complexity control. Cloud technology can provide a
challenges businesses face would likely be similar to
standardized model that both patients and medical
what health care might encounter with wearable data.
staff can rely on. Security measures protecting BlueBoth users and health care professionals would have
tooth communication is still very much a problem.
decreased satisfaction when dealing with constantly
changing data sets[20]. Decreased collaboration is
the EHR.

6

References

futures. Movement Disorders, 31(9):1314–1326,
2016.

[1] Piwek L. Ellis D. Andrews S. Joinson A. The
rise of consumer health wearables: Promises and [13] Lofy K. Hale M. Assessing pairing and data exchange mechanism security in the wearable inbarriers. PLOS, 2016.
ternet of things. 2016 IEEE International Con[2] Taber J. Leyva B. Persoskie A. Why do people
ference on Mobile Services (MS), pages 25–32,
avoid medical care? a qualitative study using
2016.
national data. J Gen Intern Med, 30(3):290–297,
[14] Raij A. Ghosh A. Kumar S. Srivastava M. Pri2015.
vacy risks emerging from the adoption of innocu[3] Lee H. Song C. Hong Y. Kim M. Cho H. Kang
ous wearable sensors in the mobile environment.
T. Et Al. Wearable/disposable sweat-based gluProceedings of the SIGCHI Conference on Hucose monitoring device with multistage transderman Factors in Computing Systems, pages 11–
mal drug delivery module. Science Advances,
20, 2011.
3(3), 2017.
[15] Bailly G. M¨
uller J. Rohs M. Wigdor D. Kratz
[4] Li R. Et Al. Wearable performance devices in
S. Shoesense: a new perspective on gestural insports medicine. Sports Health, 8(1):74–78, 2016.
teraction and wearable applications. Proceedings
of the SIGCHI Conference on Human Factors in
[5] Spitzer R. Et Al. Utility of a new procedure
Computing Systems, pages 1239–1248, 2012.
for diagnosing mental disorders in primary care.
JAMA, 272(22):1749–1756, 1994.
[16] Bashivan P. Rish I. Heisig S. Mental state recognition via wearable eeg. Proceedings of 5th NIPS
[6] Apple. iphone 7 - technical specs. Apple iPhone
workshop on Machine Learning and Interpreta7 Site, 2016.
tion in Neuroimaging, 5(1), 2015.
[7] Apple. Your heart rate. what it means, and
[17] Geart Rhythm Society. Artificial intelligence auwhere on apple watch you’ll find it. Apple Watch
tomatically detects atrial fibrillation using apple
Site, 2016.
watch’s heart rate sensor. Heart Rhythm Society
Site, 2016.
[8] Wang S. Bie R. Zhao F. Zhang N. Cheng X. Choi
H. Security in wearable communications. IEEE
[18] Wearable Tech. Emotion tracker sence tracks 64
Network, 30(5):61–67, 2016.
emotions using ‘heart rate variability’. Wearable
Tech Site, 2016.
[9] Chansin G. Hayward J. Wearable sensors 20162026: Market forecasts, technologies, players. ID
[19] C. Vuppalapati, A. Ilapakurti, and S. Kedari.
Tech Site, 2016.
The role of big data in creating sense ehr, an
integrated approach to create next generation
[10] Saeb S. Lonini L. Jayaraman A. Mohr D. Kordmobile sensor and wearable data driven elecing K. The need to approximate the use-case in
tronic health record (ehr). In 2016 IEEE Second
clinical machine learning. Gigascience, 6(5):1–9,
International
Conference on Big Data Comput2017.
ing Service and Applications (BigDataService),
[11] Menk A. Casagrande S. Geiss L. Prevalence
pages 293–296, March 2016.
of and trends in diabetes among adults in the
united states, 1988-2012. JAMA, 314(10):1021– [20] Ikhlas Fuad Zamzami, H. A. A. Fatani, and
N. A. H. Zammarah. Data migration challenges:
1029, 2015.
The impact of data quality x2014; case study of
[12] Kubota J. Chen J. Little M. Machine learning
university putra malaysia upm. In 2011 Interfor large-scale wearable sensor data in parkinnational Conference on Research and Innovation
son’s disease: Concepts, promises, pitfalls, and
in Information Systems, pages 1–5, Nov 2011.
7

¨
[21] Ozdemir
A. Barshan B. Detecting falls with
wearable sensors using machine learning techniques. Sensors, 14(6):10691–10708, 2014.

8