A Good Article
A Good Article
A Good Article
discussions, stats, and author profiles for this publication at: https://www.researchgate.net/publication/265212045
CITATIONS READS
4 64
5 authors, including:
Zhinan Li Wenyao Xu
XenoTech University at Buffalo, The State University o
9 PUBLICATIONS 53 CITATIONS 98 PUBLICATIONS 396 CITATIONS
All in-text references underlined in blue are linked to publications on ResearchGate, Available from: Wenyao Xu
letting you access and read them immediately. Retrieved on: 21 August 2016
IEEE ICC 2014 - Selected Areas in Communications Symposium
Abstract Mobile Health (mHealth) is expected to play a As observed above, a wearable system with mobile
special role in today and the future healthcare delivery. Based on interfaces may play a vital role to detect falls and prevent
this trend, we design a Smart Wristlet mHealth system with elderly people from the related injuries [1],[3]. In such a
mobile interface. The designed Smart Wristlet is dedicated to system, sensor units (for example, accelerometers) are placed
offer real-time alert for elderly fall, which is the most important at human bodies to monitor user activities. In this system, the
when population ageing is becoming. In the Smart Wristlet collected data streams are processed and analyzed in the
mHealth system, fall detection is the bottleneck of the system embedded CPUs. In order to detect whether the wearer is
operation. To remove this bottleneck away, we propose a fall falling, existing human activity recognition methods can be
perception solution for elderly care. In this proposal, we abstract
classified into four categories, namely, fixed threshold methods
and construct primitive-based features from raw data collected
by the Smart Wristlet mHealth system, in which the most
[7]; pattern recognition strategies [8]; conventional or fuzzy
valuable features can be selected by using a TF-IDF (Term logic [9]; and artificial neural networks [10].
FrequencyInverse Document Frequency) metric. In reality, Unfortunately, these traditional methods are confronted
these selected features are the most effective to perform fall with two main challenges. Firstly, their detection accuracy is
detection. Our system tests and clinical trials demonstrate that relatively low, which cannot meet the requirements for elderly
this proposal is eligible to turn the Smart Wristlet mHealth care and may cause serious medical negligence. Secondly, their
system into a real solution for elderly care. Results show that the high computing efficiency conflicts with terminal limitations in
recognition precision and recall can reach 93% and 88%,
embedded wearable systems. This is because a highly complex
respectively. Compared with existing solutions, the gain from our
proposal is an efficient prevention method for elderly fall, and
computing process may easily drain the embedded CPU
can save more than 800 million dollars per year at todays socio- resources and the battery energy. In turn, the system response
economic level. is also affected. This greatly restrains the service quality of
elderly care because the fall detection accuracy, response time,
KeywordsmHealth (mobile Health), Smart Wristlet, Fall and battery life are all the sensitive factors for an elderly fall
Perception, TF-IDF (Term FrequencyInverse Document prevention system.
Frequency), Clinical Trials.
In this study, we present a novel wearable sensor system
I. INTRODUCTION called Smart Wristlet, which can provide 24 hours fall
detection service. To ensure the reliability and practicability of
A fall accident is the leading cause of accidental injury and Smart Wristlet, we developed a machine-perception based fall
death for the elderly people [1]. A hospital-based study found detection algorithm in it, which is called Fall Perception for
that 68% of elderly hospitalizations for injury are fall-related short in the rest of this paper. Due to the high efficiency and
[2]. In addition, with a person getting older, he/she is at a accuracy of Fall Perception, Smart Wristlet achieves fall
higher risk to be involved in a fall injury. As reported in [2], for detection precision as high as 93%, which is 3% higher than
persons older than 85, 86% of their injuries are from fall conventional methods. The recall is improved by 9% as well.
accidents. Only in United States, elderly falls cause around 8 More importantly, by simplifying the computational
billion dollars direct medical costs each year [3]. This huge loss complexity and sensor data, the battery life is extended by
has made elderly-fall be a major socio-economic problem. more than 30%. Obviously, these gains are very important to
Thus, it is significant to prevent elderly-fall accidents. Recent ensure long-time and uninterrupted mHealth services in real
advances in wireless mobile networking technologies have applications. For instance, with pop-up airbags on the users
provided an opportunity to alleviate this problem. This concept body, Smart Wiristlet can trace users activity and popup the
is known as mobile health (mHealth) [4],[5],[6] which is airbags to protect him when a fall event is detected. Besides
changing health-care delivery today and is at the core of that, automate remote emergency alerting for elderly fall is also
responsive health systems [4]. important since the old people usually are not able to call for
help when they fall and get injured.
4271
IEEE ICC 2014 - Selected Areas in Communications Symposium
4272
IEEE ICC 2014 - Selected Areas in Communications Symposium
larger, the computing complexity is becoming higher. If N is precision can be improved. However, it is inevitable to lower
getting smaller, more information would be lost. How to down the rate of recall. This is because the precision and the
determine the size of N will be investigated in the future. In this recall are working at a conflict fashion. In reality, the threshold
way, the N-grams of primitives are constructed and they are should be customized according to the application scenario.
treated as feature candidates.
Below, we perform system tests and clinical trials to
C. Feature Selection verify effectiveness of this proposal.
In general, the number of obtained N-grams feature IV. EXPERIMENTS AND EVALUATION
candidates is large. This phenomenon leads high computing
overhead to online fall detection in the Smart Wristlet. In To evaluate the performance of the Fall Perception
addition, some N-grams may be irrelevant or even misleading framework, we conduct experiments in the Smart Wristlet
features. These misleading features are dangerous and may system. In our study, 246 aging people in Beijing are chosen as
decrease the detection accuracy, especially when the training experimental subjects. These subjects are all between 59 and
data are not clean and correct. To solve these concerns, the N- 63 years old, and each of subjects performed standing, walking
gram features should be filtered to select proper ones. and falling motions wearing the Smart Wristlet. In system
implementation, the TF-IDF metric is adopted to select K
The basic principle of filtering features is to select relevant proper N-gram features. The Smart Wristlet performance is
features while discarding the irrelevant and misleading features. tested in two major aspects fall detection accuracy and
In the Fall Perception, three metrics are taken into account for battery life. For comparison in terms of fall detection, the
evaluating the relevance of a feature N-grami. conventional pattern recognition [15] and neural networks [16]
1) Occurrence Number of N-grami: Occurrence Number are tested in the Smart Wristlet as well. The test performances
of N-grami is the number of N-grami in positive are demonstrated below.
training data sets. This metric is effective because A. Fall Detection Accurancy
common features tend to be relevant, while the rare
features are usually tail or caused by noise. Firstly, we discuss the fall detection accuracy. Generally,
the fall detection accuracy is quantified by precision and recall.
2) Frequency of N-grami: Frequency of N-grami is the (In this study, the precision and recall are the common concepts
Occurrence Number of N-grami divided by all N- in data mining area.) Here,
grams in the positive training data sets. Compared with {
}
Occurrence Number, Frequency can reduce the = { !"#}
,
influence of training data set size.
{-./01234 5678 9:;<=>}?{@ABC DEFGHI}
3) TF-IDF of N-grami: TF-IDF is a widely cited and $%&'()*+, = {JKLMNOPQ RSTU VWXYZ[}
.
technique in text mining area, and its applied to
As discussed in Sec. III, the accuracy of the Fall Perception
estimate the relevance between word terms and articles
framework is affected by K-value of feature selection. For a
[12]. In this study, we define each N-gram as a term,
larger K-value, more N-grams features will be selected and
and develop TF-IDF to evaluate the relevance between
more variations of positive data will be contained. This is
one N-gram and the positive training data set.
nature to improve fall detection accuracy because of more
Based on these metrics, best features can be searched by positive effect accumulated. But if the K-value is too large,
fixed number. For example, we sort the N-grams ascendingly some of unexpected features may be also involved into the
by their metric values, and define the first Kth ones as relevant classifier training, which affects the classifier performance. We
features. The algorithm performance with different K values 100
will be discussed in Section IV.
90
D. Classifier Training
80
polynomial. 0
0 10 20 30 40 50 60 70 80 90 100
Then, a discrimination threshold is needed to distinguish K-value in Feathre Selection step (% Primitive Vocabulary size)
4273
IEEE ICC 2014 - Selected Areas in Communications Symposium
evaluated the algorithm accuracy with different K-values and into real applications with required terminal performances, for
the result is demonstrated in Figure 5. To consider the effect of example, longer battery life. More valuably, this proposal can
N-gram feature set size, the K-value is quantized by be deployed to any similar mHealth systems if they need the
comparison with the number of all N-gram features. As shown fall detection.
in Fig. 5, when K-value is about 60% of the N-gram feature set
size, the precision and the recall reach their maximum value.
Results show that the recognition precision and the recall can REFERENCES
be improved 3% and 9%, respectively, compared with exiting
[1] T.Tamura, T. Yoshimura, M. Sekine, M. Uchida and O. Tanaka, A
methods. Wearable Airbag to Prevent Fall Injuries, IEEE Transactions on
B. Terminal Battery Life Information Technology in Biomedicine, vol. 13, pp. 910 - 914, Nov.
2009.
In terms of battery life, it is the basis for stable service of a [2] C. DL, M. JG and C. TV, Hospital resources used to treat theinjured
wearable system. Clearly, to reduce power consumption, it is elderly at North Carolina trauma centers, JAm Geriatr Soc. vol. 41, pp.
significant to minimize online computing complexity for 847-852, 1993.
extending battery life. Fall Perception achieves this goal by [3] N. V. Carroll, P. W. Slattum, and F. M. Cox, The Cost o f Falls Among
using sparse primitive features, selecting only top K relevant the Community-Dwelling Elderly, Journal of Managed Care
Pharmacy, JMCP, vol. 11, no. 4, May 2005.
features, and utilizing simple linear SVM classifier. Figure 6
shows the battery life of Smart Wristlet when it is employed [4] Global survey report, mHealth: New horizons for health through mobile
technologies, www.who.int/goe/publications/goemhealthweb.pdf, World
with different fall detection algorithms. In this figure, the size Health Organization, 2011.
of feature set is 600. The experimental results show a lower [5] A. Huang, W. Xu, Z. Li, L. Xie and M. Sarrafzadeh, System Light-
power consumption of our proposal than those of traditional loading Technology for mHealth: Manifold-Learning based Medical
methods. To achieve the highest accuracy, K-value is set at 60% Data Cleansing and Clinical Trials in WE-CARE Project, IEEE Journal
of the N-gram feature set size, which is equal to 360. Then our of Biomedical and Health Informatics, to be published.
proposal is able to extend the Smart Wristlet battery life to [6] A. Huang, C. Chen, K. Bian, X. Duan, M. Chen, H. Gao, C. Meng, Q.
more than 24 hours, which is 6 hours longer than neural Zheng, Y. Zhang, B. Jiao, and L. Xie, WE-CARE: An intelligent
mobile telecardiology system to enable mHealth applications, IEEE J.
network method and 9 hours longer than pattern recognition Biomedical Health Informatics, to be published.
methods. This feature of our proposal not only improves user [7] J. Fahrenberg, F. Foerster, M. Smeja and W. Muller, Assessment of
experience of the Smart Wristlet, it also lowers the possibility postureandmotion bymultichannel piezoresistive accelerometer
of unexpected system breakdown and related elderly hurt. recordings, Psychophysiol., vol. 34, pp. 607612, 1997.
[8] P. H. Veltink, H. B. Bussmann, W. de Vries, W. L. Martens and R.C.
30 van Lummel, Detection of static and dynamic activities using
Fall Perception uniaxialaccelerometers, IEEE Trans. Rehabil. Eng., vol. 4, no. 4, pp.
28
Pattern Recognition 375385,Dec. 1996.
Neural Network
26 [9] J. M. Winters, Y. Wang and J. M. Winters, Wearable sensors and
telerehabilitation, IEEE Eng. Med. Biol. Mag., vol. 22, no. 3, pp. 56
24 65,May/Jun. 2003.
Battery life (Hour)
22
[10] K. Aminian, P. Robert, E. Jequier and Y. Schutz, Incline, speed, and
distanceassessment during unconstrained walking, Med. Sci. Sports
20 Exerc.,vol. 27, pp. 226234, 1995.
[11] M. Zhang and A. A. Sawchuk, Motion Primitive-Based Human
18
Activity Recognition Using a Bag-of-Feature Approach, ACM SIGHIT
16 International Health Informatics Symposium (IHI), Miami, Florida,
USA, Jan. 2012.
14
[12] K. W. Church, W. A. Gale, Inverse Document Frequency (IDF): A
12
Measure of Deviations from Poisson, Natural language processing
using very large corpora, Kluwer Academic Press, pp. 283-295, Boston,
10
50 100 150 200 250 300 350 400 450 500
1999.
K-value in Feathre Selection step [13] A. Brakensiek and G. Rigoll, A comparison of character n-grams and
dictionaries used for script recognition, Sixth International Conference
Fig. 6.Battery life of Smart Wristlet. on Document Analysis and Recognition, 2001. Proceedings., Seattle,
WA, USA, 10-13 Sep. 2001.
[14] G.-X. Yuan and K.-L. Ma, Scalable Training of Sparse Linear SVMs,
V. CONCLUSION IEEE 12th International Conference on Data Mining (ICDM), 2012, pp.
775-784, Brussels, Belgium, Dec. 10 - 13, 2012.
In this study, we propose a new fall detection solution
[15] C.W. Chang, and G.A.Bekey,Pattern recognition applied to the
based on machine-perception, called Fall Perception, which is a prediction of post-operative gait in man,1978 IEEE Conference on
combination of efficiency and accuracy requirements. We Decision and Control including the 17th Symposium on Adaptive
perform system experiments and clinical trials in the developed Processes, pp. 787 788, San Diego, US, Jan. 1978.
Smart Wristlet mHealth system. Test results verify that this [16] S.Zein-Sabatto, K.Kuah, R.Hajeomar and M.Bodruzzaman, An
proposal is an effective mHealth tool to prevent elderly falls. intelligent vision system for motion detection using a neural network,
Specifically, the fall detection accuracy can be improved, in Proceedings of the 1994 IEEE Southeastcon '94. Creative Technology
Transfer - A Global Affair, pp. 362 366, 10-13 Apr. 1994.
which the performance gain can save more than 800 million
dollars direct medical costs per year in the United States.
Meanwhile, this proposal is helpful to turn the Smart Wristlet
4274