Evaluation of Threshold-based Fall Detection on Android Smartphones
Tobias Gimpel, Simon Kiertscher, Alexander Lindemann, Bettina Schnor and Petra Vogel
Department of Computer Science, University of Potsdam, August-Bebel Str. 89, Potsdam, Germany
Keywords:
Fall Detection, Development of Assistive Technology, Sensors-based Applications.
Abstract:
This paper evaluates threshold-based fall detection algorithms which use data from acceleration sensors that
are part of the current smartphone technology. Different detection algorithms are published in the literature
with different threshold values. This paper presents the evaluation of 5 different algorithms which are suited
for Android smartphones. In contradiction to prior work, our experiments indicate that the Free Fall detection
Phase is necessary for a low False Positive Rate. Further, we present an empirical evaluation of currently
available fall detection apps in the Google Play store.
1 INTRODUCTION
The probability of falls increases for older people.
Approximately, about 33 % of older persons fall un-
intentionally each year (Mellone et al., 2012). Falls
with the loss of consciousness are of special harm.
Therefore, several groups are working on automatic
fall detection, for example (Bourke et al., 2007;
Sposaro and Tyson, 2009; Dai et al., 2010; Fudickar
et al., 2014).
In this paper, we will only consider fall detec-
tion systems using Android based smartphones due to
their widespread use and availability. Current smart-
phones are equipped with a tri-axial accelerometer
sensor which periodically collects a vector of axis-
specific acceleration data. Typically, this is used to
re-orient the screen as a user moves the device. But
this sensor has also potential to be used for fall de-
tection. Smartphone operating systems such as An-
droid aim to save energy and typically sample with
low rates (between 20 and 100 Hz). Prior studies have
shown that these sampling-rates are suited for fall de-
tection (Mehner et al., 2013; Fudickar et al., 2014).
An optimal fall detection algorithm should com-
bine a high sensitivity with a high specificity. The
sensitivity is a measure for the number of correctly
detected falls:
Sensitivity =
TruePositives
Number of all falls
where TruePositives is the number of correctly de-
tected falls.
The specificity measures the rate of
TrueNegatives, i.e. the percentage of correctly
classified non-fall situations:
Speci f icity =
TrueNegatives
Number of all non falls
A high specificity means a low number of false
alarms. This is also a very important metric for the
usability of the system.
We analyzed the already published algorithms for
fall detection and discovered that there exist differ-
ent variations of the algorithms which differ for ex-
ample in the calculation of the orientation change
and in their threshold settings. Since these algo-
rithms were not compared before, we implemented
and tested these different algorithms. As a result from
these tests, we delivered new threshold values which
perform better than the ones published before.
In September 2014, there are already about 13
apps available in the Google play store which claim to
perform fall detection. The second part of this work
is a survey about the quality of those applications.
The remainder of the article is structured as fol-
lows: In Section 2 related work is discussed. The ba-
sic algorithm of threshold-based fall detection is in-
troduced in Section 3. Section 4 presents the evalu-
ated algorithms and parameter settings. The results
of our evaluation are shown in Section 5. Further,
we tested current available fall detection applications
from the Google App Store. The results are presented
in Section 6. The article ends with a conclusion.
598
Gimpel T., Kiertscher S., Lindemann A., Schnor B. and Vogel P..
Evaluation of Threshold-based Fall Detection on Android Smartphones.
DOI: 10.5220/0005280805980604
In Proceedings of the International Conference on Health Informatics (HEALTHINF-2015), pages 598-604
ISBN: 978-989-758-068-0
Copyright
c
2015 SCITEPRESS (Science and Technology Publications, Lda.)
2 RELATED WORK
Several groups have already investigated fall detec-
tion on smartphones.
In 2010, Dai et al. proposed the PerFallD proto-
type for the first available Android smartphone, the
G1 phone (Dai et al., 2010). The detection algorithm
uses four thresholds to detect an impact and a sta-
ble phase and the change of orientation. The chosen
thresholds are not applicable for current Android ver-
sions anymore, but the study already shows the fea-
sibility of smartphone based fall detection. For the
evaluation, data from 450 falls were collected with
different directions (forward, lateral and backward),
different speeds (fast and slow) and in different en-
vironment (living room, bedroom, kitchen and out-
door garden). Further, data of activities of daily living
(ADL) including walking, jogging, standing and sit-
ting were gathered. Fall and ADL data were provided
by graduate students from 20 to 30 years old. The au-
thors present evaluations for wearing the smartphone
in a shirt pocket, on the belt, or in the pant pocket. The
presented evaluation shows that PerFallD achieves the
best performance when the device is attached with
the belt with an average False Negative value being
2.67 % and the False Positive value being 8.7 %.
While this is a promising result, no evaluation with
ADLs of older people is given. Further, the authors
report that the system keeps running about 33.5 h un-
til the battery is exhausted which also confirms the
usability of these devices for fall detection.
Dai et al. compare their system with a commercial
product provided by Brickhouse (Brickhouse, ) . This
fall detector consists of a teleassist base and a portable
sensor. Since the base device needs to be installed
indoors and the signal transmission distance between
the sensor and the base is limited, the system is only
useful in this environment. The experiments reported
in (Dai et al., 2010) show that this system has a high
false negative rate in backward falls (29.9 %) and a
high false positive rate (21.9 %).
In (Hwang et al., 2012), a smartphone running
Android 2.3.3 was used to record 100 Hz bandwidth
signals from the three-axis acceleration sensor. The
saved data were processed to detect falls using Mat-
lab 7.0 not on the device itself, but on a PC. Based
on data from 200 experimental falls, obtained by fas-
tening a smartphone to a belt worn around the waist,
an overall detection rate of 95% was achieved, cor-
responding to direction-specific rates of 94% for for-
ward falls, 100% for backward falls, 94% for leftward
falls and 92% for rightward falls. Based on the exper-
imental results of 6 ADL scenarios, the threshold for
acceleration was established at 2.2 g and the threshold
for angular displacement was set at 50
.
Since Android based smartphones use sampling-
rates of maximum 100 Hz, different sampling rates
were investigated in a trace-driven simulation ((Fu-
dickar et al., 2014)). The results confirm that low
sampling rates of at least 50 Hz can be used and have
a sensitivity of 99% for the collected fall records.
Further, the specificity of the threshold-based fall-
detection algorithm was tested with ADLs of elderly
people. In the simulation, no false positive occurred.
The threshold settings were taken from (Karth, 2012).
Mehner et al. (Mehner et al., 2013) also published
results from experiments with threshold based fall-
detection on smart phones which confirm that the
lower sampling rates (such as 50 Hz) that are sup-
ported by Android are uncritical. Their results indi-
cate that the exclusion of the free-fall detection phase
may increase the detection accuracy by 27 % from
56% with free-fall detection to 83% without free-fall
detection. Overall the proposed algorithm achieved a
maximal sensitivity of 83% and a specificity of 100%
in the presented experiments. Results from real falls
and ADLs from elderly people are not presented.
3 THRESHOLD BASED FALL
DETECTION
For a tri-axial accelerometer, fall situations are char-
acterized by multiple sequential phases. The ac-
celerometer collects a vector (x,y,z) of the axis-
specific acceleration forming together the accelera-
tion. The length of the acceleration vector corre-
sponds to the power of the acceleration. Since we use
gravity as our metric, we have to divide the length of
the acceleration vector by the power of the gravity on
earth which is 1g = 9, 81
m
s
2
1
. This leads to equation
(1):
|
v | =
p
x
2
+ y
2
+ z
2
9, 81
(1)
This value is the so-called G-Value used in the thresh-
old based fall detection algorithm.
The different phases of the fall detection algorithm
are shown in Figure 1. The fall detection algorithm
starts with the FreeFall Test Phase. Within a fall
situation, the falling body experiences zero-gravity.
1
The exact gravity appeals on a body (or mass in gen-
eral) depends on where you are on earth. This is due to
gravity anomalies and the fact that the earth is not a perfect
sphere. One can say in general, that gravity is stronger near
the poles. The differences in gravity is still small enough to
set a general value for gravity in a fall detection application
EvaluationofThreshold-basedFallDetectiononAndroidSmartphones
599
Figure 1: The different phases of the fall detection algo-
rithm.
The free fall is the initial event of each fall situa-
tion and therefore has to be recognized first. To de-
tect a free fall, the gravity value must drop below the
free fall threshold LOWER THRESHOLD for a min-
imal duration of TIMER-FREEFALL. If this is not
the case and the threshold is exceeded again before
TIMER-FREEFALL has expired, the algorithm stays
in the FreeFall Test Phase. Other algorithms (such as
(Bourke et al., 2007; Mehner et al., 2013; Fudickar
et al., 2014)) exclude a minimal duration for free fall
detection since without the free fall detection phase
they observed a higher fall detection rate in their ex-
periments.
Once a free fall was detected, the Impact Test
Phase starts in which the algorithm tries to detect an
impact, which is given if the G-Value exceeds the so-
called UPPER THRESHOLD. The maximal duration
between free fall and impact recognition may not ex-
ceed the duration defined in TIMER-IMPACT. If this
period expires and no impact was detected, the algo-
rithm resets and starts again with the FreeFall Test
Phase.
In a fall situation the impact is followed by a sta-
ble phase, thus the algorithm switches into the Stable
Test Phase. During this Phase, the G-Values has to
drop below the VARIATION threshold (1st value of
the Stable A row of Table 2), for a minimal duration
of TIMER-NOMOVEMENT (2nd value of the Sta-
ble A row of Table 2). All this must happen within
another time window called TIMER-STABLE (3rd
value of the Stable A row of Table 2) to pass this
phase. Mehner et al. (Mehner et al., 2013) divide this
phase into two sub-phases (depicted by the Stable B
row of Table 1 and 2). The first sub-phase is only
for omitting fluctuations in the G-Value which may
occur right after an impact and its duration is equal
to TIMER-NOMOVEMENT. The second sub-phase
tests, if the VARIATION threshold is exceeded by the
G-Value within the time of that sub-phase which is
equal to the TIMER-STABLE parameter. If it is ex-
ceeded the Stable Test Phase has failed and the algo-
rithm resets.
The correct distinction between a fall situation and
an Activity of Daily Life (ADL) is one of the most im-
portant features of a fall detection algorithm, since
even a small false positive rate will make the algo-
rithm useless for real world scenarios. To distinguish
a fall situation from an ADL, some researcher pro-
pose an Orientation Test Phase (Karth, 2012; Mehner
et al., 2013). During this phase it is tested, if the de-
vice, and thus its carrier has changed its orientation
significantly. This can be done in different ways as
depicted in Table 1 and 2 by the A, B and C in the
Orientation row.
Version A, from (Karth, 2012), takes the last ac-
celeration vector prior to a possible fall situation
and compares it with the acceleration vector given
after the Stable Test Phase. If the difference is 45
or more, the fall is verified.
HEALTHINF2015-InternationalConferenceonHealthInformatics
600
Version B, from (Mehner et al., 2013), compares
the mean of the last 100 values before the fall sit-
uation with the mean of the first 100 values after
the detected fall. This is done separately for every
axis. If one of the axis has a deviation of 0,4g or
greater, the fall is verified.
Version C which is proposed in this paper com-
bines Version A and B and compares the mean ac-
celeration vector
v of the last 100 acceleration
vectors given before the fall situation occurred
with the mean acceleration vector
w of the last
100 acceleration vectors given after Stable Test
Phase is passed. The angle is calculated as:
α = arccos
v
w
|
v | |
w |
(2)
If the angle is greater than 66
, the fall is verified.
After a fall was detected by passing all previous
phases, the Alert Phase begins. In this phase, a timer
is started and the device informs the user that a fall
was detected. Typically the user can confirming or
aborting the alert, so an emergency call can be either
initiated or prevented. In case the timer expires and
no action was chosen, an emergency call is initiated.
It often happens to elder people that they will not lose
its consciousness after a fall, but will still be unable to
interact with the device. If this is the case, the timer
approach will work fine due to an activation of the
hands-free phone system. Alternatively or in addition
to an emergency call, a SMS message, TCP message
or other message systems can sent alerts to a config-
ured phone number, IP address etc.
4 EVALUATION ENVIRONMENT
While fall detection with smartphones seems to be a
promising approach and different groups have pub-
lished work in this direction, an analysis with real falls
of elderly people is unrealistic and still missing.
In (Fudickar et al., 2014), results from a simula-
tion with ADLs of elderly people was presented. It
was shown that fall detection is also possible with the
low sampling rates used on smartphones. But the sim-
ulated fall detection algorithm was adapted for a spe-
cial sensor, the ADXL345, which is not used in cur-
rent smartphone technologies.
Therefore, we decided to repeat the experiment for
Android smartphones. The following steps were exe-
cuted:
1. implementation of the fall detection algorithm (as
described in Section 3) for Android smartphones,
2. integration of the new implemented fall detection
algorithm into the simulation testbed,
3. collecting falls with an Android smartphone
4. collecting ADLs from elderly people with an An-
droid smartphone.
During the implementation phase, we considered
the already published algorithms and discovered the
different variations presented in Section 3. Since it
was still unclear which algorithm and parameter set-
ting will give the best results, we implemented and
tested the following configurations for Android smart-
phones:
1. Karth FF: The parameter values for thresholds
and timer settings were taken from (Karth, 2012).
2. Karth: The same as Karth FF, but the Free Fall
Phase is omitted.
3. Mehner FF: The parameter values for thresholds
and timer settings were taken from (Mehner et al.,
2013).
4. Mehner: The same as Mehner FF, but the Free
Fall Phase is omitted.
5. Gimpel: The fall detection algorithms
from (Karth, 2012) and (Mehner et al., 2013)
were implemented by Tobias Gimpel in (Gimpel,
2014). He also developed a combination of the
former two settings.
An overview of the different variants and the relat-
ing parameter settings are given in the Tables 1 and 2.
5 RESULTS
The sensitivity and specificity of the threshold based
fall detection algorithms is evaluated with the fall sets
and ADLs that are described in the following subsec-
tions.
5.1 Fall Set and Sensitivity
Since, the risk of injuries in case of older test persons
is obviously much too high, the fall set was recorded
by three test persons between 23 and 55 years old.
The fall set covered frontal falls, falls to the left
and to the right. The test persons used two types of
smartphones: HTC Desire 816 and Sony Xperia V.
The device was put into a funny bag at the hip in front.
The first test person (29 years old) simulated 30
falls: 10 frontal, 10 to the left side, and 10 to the right
side (prob29). The second test person (23 years old)
simulated 12 falls: 4 frontal, 5 to the left side, and 3
to the right side (prob23). The third test person (55
EvaluationofThreshold-basedFallDetectiononAndroidSmartphones
601
Table 1: Different versions of the fall detection algorithm.
Karth FF Karth Mehner FF Mehner Gimpel
Free Fall X X X
Impact X X X X X
Stable A X X X
Stable B X X
Orientation A X X
Orientation B X X
Orientation C X
Table 2: Parameter settings of the different algorithms.
Karth Mehner Gimpel
Free Fall 0,75g/30ms 0,5625g/30ms 0,75g/30ms
Impact 2g/500ms 2,3g/300ms 2g/500ms
Stable A 0,4375g/1sec/3,5sec - 0,2g/1sec/3,5sec
Stable B - 0,4g/1sec/1sec -
Orientation A 45
- -
Orientation B - 0,4g -
Orientation C - - 66
years old) simulated 10 falls: 4 frontal, 3 frontal left,
and 3 frontal right (prob55). The different fall char-
acteristics and number of falls are due to the different
fall capabilities of the test persons. All falls recorded
were critical falls, with loss of consciousness (where
the test person did not move after the falls).
We used the data from our simulated falls as input
to test the algorithms and parameter settings described
in Section 4. The results are shown in Table 3. In the
column NOF the number of falls is given. The detec-
tion rate using the Karth FF, Karth, and Gimpel
settings are very good. The Mehner settings show an
inconsistent behavior. For all three test persons, the
algorithm without Free Fall Phase has a higher sen-
sitivity, which correlates with prior published results
in (Mehner et al., 2013). But for the falls from prob29,
the sensitivity for both variants is only about 50 %.
While these variants have a good detection rate for
falls ahead, they have difficulties to detect falls to the
left or right side.
5.2 ADLs of Elderly People
To evaluate the specificity of the algorithms, we
recorded the acceleration characteristics of ADLs of
two test persons, 55 resp. 72 years old. As before, the
test persons were equipped with a fanny pack, worn at
the test persons’ hip in which the device was carried.
The considered recording duration per test person
ranged from 11 (prob72) to 286 (prob55) hours. The
older test person carried the device during one day,
while the younger test person carried it several days
during her daily activities covering typical daily activ-
ities including eating lunch, walks, activities at work
and also homework.
The maximal recorded acceleration in the cropped
recordings was at 3.17 g.
Table 4 shows the results for the different
versions and parameter settings. The most ro-
bust versions are Mehner with the Free Fall
Detection Phase (no False Positives) and Gimpel
(only 2 False Positives). Further, the Karth FF ver-
sion including the Free Fall Phase also performs better
than the simple Karth version. But both Karth ver-
sion show much higher False Positives than the other
settings. Obviously, the Fall detection Phase is impor-
tant for a low False Positive Rate.
5.3 Interpretation of the Results
While (Mehner et al., 2013) proposed the exclusion
of the Free-Fall Detection step to enhance the sensi-
tivity, our experiments with ADLs show that the free-
fall detection phase is important for a low False Pos-
itive Rate. Therefore, we consider in the following
only those algorithms which use this phase: Karth
FF, Mehner FF, and Gimpel. From these, Mehner
FF has a significantly worse sensitivity (see Table 3).
The sensitivity of Kart FF and Gimpel is compara-
ble. Since regarding the specificity, Gimpel is much
better than Kart FF in our experiments, we conclude
that the settings from Gimpel are a good compromise.
HEALTHINF2015-InternationalConferenceonHealthInformatics
602
Table 3: Sensitivity: Detected Falls.
fall direction NOF Karth FF Karth Mehner FF Mehner Gimpel
prob29 overall 30 30 28 13 15 26
ahead right 5 5 4 5 5 4
ahead left 5 5 5 5 5 4
left side 10 10 10 3 4 10
right side 10 10 9 0 1 8
prob23 overall 12 10 9 3 10 9
left side 5 4 4 1 5 3
right side 3 2 2 2 2 3
ahead 4 4 3 0 3 3
prob55 overall 10 9 10 2 8 9
ahead right 3 3 3 1 2 3
ahead left 3 3 3 0 3 3
ahead 4 3 4 1 3 3
Table 4: Specificity: Results from ADLs.
duration Karth FF Karth Mehner FF Mehner Gimpel max g min g
prob55 286 h 24 57 0 5 2 3.179 0.061
prob72 11 h 0 3 0 1 0 2.555 0.041
6 EVALUATION OF APPs FROM
THE GOOGLE APP STORE
The second part of this work is a survey about useful
fall detection applications in the Google play store. In
September 2014, the search for fall detection, within
the Google play store, results in 22 hits for applica-
tions in different categories (e.g. medicine, health &
fitness, lifestyle and social). 13 out of these 22 appli-
cations are related to real fall detection and thus were
considered in the following. 2 out of the 13 remaining
applications were commercial applications and avail-
able for 4 C and 120 C respectively. Due to the high
price, the 120 C application was not tested any fur-
ther.
First, we tested their user interface and the usabil-
ity. The following characteristics of an application
resulted in an exclusion for further tests:
a failed or impossible installation
no reaction of the application after installation
the need to register for a phone call in a foreign
country
the phone call destination is not obvious
Only 8 applications passed these initial tests and
were further tested regarding their sensitivity and
specificity.
The specificity was tested with a fixed set of ac-
tivities of daily living (ADL) like walking around,
climbing stairs or sitting down on a chair. All this was
done at varying speeds in a time window of about 10
minutes per application. During these tests, the smart-
phone was in a trousers’ pocket.
The sensitivity of the applications was tested with
10 falls in forward direction performed by two test
persons (23 resp. 55 years old). During the fall tests,
the smartphone was worn in a belt bag in front.
The results of the tests are shown in Table 5. The
FP column indicates observed false positives within
the ADL tests. Columns prob23 and prob55 refer to
the 23 and 55 years old test persons. 50 % of the appli-
cations did not pass the ADL test and indicated falls
while executing. Regarding the sensitivity, the best
application was iCare Personal Emergency Alert with
a detection rate of 70 %, followed by Fade: fall de-
tector and SecureMe Active with 60% detection rate
each. All other applications had a very low sensitivity
even for the easiest category of falls in forward direc-
tion, or low specificity.
7 CONCLUSIONS
Automatic fall detection is one of the most important
assistive technologies for elderly to lead a self deter-
mined life. An old person who lives alone is sup-
ported by the automatic fall detection system and feels
safer.
EvaluationofThreshold-basedFallDetectiononAndroidSmartphones
603
Table 5: Test results for applications from the Google play store.
name published in app store by FP prob23 prob55
T3LAB Fall Detector T3LAB-Technology Transfer Team no 1/5 3/5
iCare Personal Emergency Alert Fansoft Labs no 5/5 2/5
Smart Fall Detection Hamideh Kerdegari no 0/5 0/5
Emergency Fall Detector Socaplaya21 no 0/5 0/5
Fall Detector Spantec GmbH yes 0/5 0/5
Fade: fall detector ITER S.A. yes 3/5 3/5
iFall: Fall Monitoring System FSU Mobile Solution yes 0/5 2/5
SecureMe Active Orion Systems yes 2/5 4/5
Threshold-based fall detection on smartphones
has shown to be a promising approach. In this study,
we evaluated different parameter settings and also
tested fall detection applications from the Google play
store.
The results of our experiments are:
1. We propose a variant of the known threshold-
based fall detection algorithms which performs
better than the algorithms published before in the
literature. The so-called Gimpel variant has a
high sensitivity (44 of 52 falls detected) and a
high specificity (only 2 false positives in 297 h
ADLs). Regarding sensitivity, the Karth FF al-
gorithm performed better (49 of 52 falls detected),
but suffers from a lower specificity (24 false posi-
tives).
2. While (Mehner et al., 2013) proposed the exclu-
sion of the Free-Fall Detection step to enhance the
sensitivity, our experiments with ADLs show that
the free-fall detection phase is important for a low
false positive rate.
3. The tests with the applications from the Google
App Store show that the currently available appli-
cations have to be considered with care. Only one
application performed satisfying in our tests (no
false positive and 4 of 10 falls detected).
4. From the experiences with the applications from
the Google App Store, we conclude that collect-
ing fall data is a critical factor for the evaluation:
different test persons tend to have different simu-
lated fall characteristics.
In the next step, we will investigate the usability
of our system in cooperation with a nursing home.
While younger people are used to the smartphone
technology, this may be a burden for elderly people.
REFERENCES
Bourke, A., O’Brien, J., and Lyons, G. (2007). Evaluation
of a threshold-based tri-axial accelerometer fall detec-
tion algorithm. Gait & Posture, 26:194–199.
Brickhouse. Brickhouse alert. http://
www.brickhousealert.com/personal-emergency-
medical-alarm.html; accessed October, 2014.
Dai, J., Bai, X., Yang, Z., Shen, Z., and Xuan, D. (2010).
PerFallD: A pervasive fall detection system using mo-
bile phones. In PerCom Workshops’10, pages 292–
297.
Fudickar, S., Lindemann, A., and Schnor, B. (2014).
Threshold-based fall detection on smart phones. In
HEALTHINF 2014, 7th International Conference on
Health Informatics, Angers, France.
Gimpel, T. (2014). Anpassung eines Algorithmus zur
aktiven Sturzerkennung f
¨
ur Android-Smartphones.
Bachelor thesis, University of Potsdam.
Hwang, S.-Y., Ryu, M.-H., Yang, Y.-S., and Lee, N.-B.
(2012). Fall Detection with Three-Axis Accelerom-
eter and Magnetometer in a Smartphone. In Interna-
tional Conference on Computer Science and Technol-
ogy, pages 65–70.
Karth, C. (2012). Fusion von Sensordaten zur robusten
Sturzdetektion im Bereich Assisted Living. Master’s
thesis, University of Potsdam.
Mehner, S., Klauck, R., and Koenig, H. (2013). Location-
independent fall detection with smartphone. In Pro-
ceedings of the 6th International Conference on PEr-
vasive Technologies Related to Assistive Environ-
ments, PETRA ’13, New York, NY, USA. ACM.
Mellone, S., Tacconi, C., Schwickert, L., Klenk, J., Becker,
C., and Chiari, L. (2012). Smartphone-based solu-
tions for fall detection and prevention: the FARSEE-
ING approach. Zeitschrift f
¨
ur Gerontologie und Geri-
atrie, 8:722–727.
Sposaro, F. and Tyson, G. (2009). iFall: an Android appli-
cation for fall monitoring and response. Conference
proceedings: Annual International Conference of the
IEEE Engineering in Medicine and Biology Society,
2009:6119–6122.
HEALTHINF2015-InternationalConferenceonHealthInformatics
604