Optimizing Interval Training Protocols Using Data Mining Decision Trees

Abstract

Interval training consists of interleaving high intensity exercises with rest periods. This training method is a well known exercise protocol which helps strengthen and improve one's cardiovascular fitness. However, there is no known method for formulating and tailoring an optimized interval training protocol for a specific individual which maximizes the amount of work done while limiting fatigue. But by using data mining schemes with various attributes, conditions, and data gathered from an individual's exercise session, we are able to efficiently formulate an optimized interval training method for an individual. Recent advances in wireless wearable sensors and smart phones have made available a new generation of fitness monitoring systems. With accelerometers embedded in an iPhone, a Bluetooth pulse oximeter, and the Weka data mining tool, we are able to formulate the optimized interval training protocols, which can increase the amount of calorie burned up to 29.54%, compared with the modified Tabata interval training protocol.

Full text

Page 1

Abstract— Interval training consists of interl
intensity exercises with rest periods. This training
well known exercise protocol which helps stre
improve one’s cardiovascular fitness. However,
known method for formulating
interval training protocol for a specific indiv
maximizes the amount of work done while limiting
by using data mining schemes with variou
conditions, and data gathered from an individu
session, we are able to efficiently formulate a
interval training method for an individual.
{dmksuh,mahsa
eavin
meth
ngth
ther
and tailoring an op
idual
fatig
s attr
al’s e
n op
Recent advances in wireless wearable sensors and
phones have made available a new genera
monitoring systems. With accelerometers
tion of
embedded
iPhone, a Bluetooth pulse oximeter, and the Weka data
tool, we are able to formulate the optimized interval t
protocols, which can increase the amount of calorie bur
to 29.54%, compared with the modified Tab
training protocol.

Index Terms— Data mining, heart rate limitatio
training, wearable wireless sensors
ata i
n, inte
nd w
and fitness
ne system
onitoring on
ensors.
hysical
tes had
a cane, using accelerometer, pressure and gyroscope s
Dencker [32] uses accelerometers to evaluate daily p
activities in children aged 8 to 12. Previously, athle
to visit athletic centers or hospitals to monitor their health
and fitness using large fitness monitoring systems. However,
with recent advances in technology, people supplement or
even tailor this process using sensors and handheld systems
in their very own home or in the field. Moreover, these
systems have the power to give precise and real-time
feedback, using the data collected from sensors such as
accelerometers, pressure sensors, and gyroscopes. In this



ng system with wireless
terleaving high intensity
been the basis for athletic
his training method is a
ch helps strengthen and
em ([1] – [7]). Moreover,
tion, general fitness, and
ring interval training, the
is utilized, and both
ces are activated. Energy
he workout period.
interval training methods
fits-all optimal solution.
ive” rest which means
ty exercise during rest
12] , Newman et al. [13],
al. [15] all recommended
rest intervals in order to
remove blood lactate which causes muscles to ache and
furthermore, causes a person to feel
mentions, there is a relation betw
acid. Therefore, by monitoring
exercise, we can avoid exceeding the level at which the heart
rate causes person to feel exhauste
tired. As Borg et al. [16]
een heart rate and lactic
one’s heart rate during
d.
There are many studies related to data mining and the
n ([19]-[24]). However, we found
no study sing data mining for interval training monitoring
and scheduling.
In this paper, we present the fol
By using data mining decision tr
lowing key contributions.
ees, we find conditions
which maximize an individual’s amount of work. In our
limited testing of this approach on different individuals, we
obtained up to 22.73 % increase in the amount of exercise.
For an individual, we saw an improvement of up to 29.54 %
in the amount of calories burned within 2 weeks.
II. EXERCISE AND HEART RATE MODEL

During exercise, the quantity of blood pumped by the heart
increases to match the increased skeletal muscle demand. In
addition, heart rate also acts as an indicator of exercise
intensity. The more intense the activity, the faster your heart
Optimizing Interval Training Protocols
Using Data Mining Decision Trees
Myung-kyung Suh1, Mahsan Rof
Computer Science De
J. Kaiser2,3, Majid Sarrafzadeh1,3
reless Health Institute3
University of California, Los Angeles
n,ani

ouei1, Ani Nahapetian1,3, William
partment1, Electrical Engineering Department2, Wi
g high
od is a
en and
e is no
timized
which
ue. But
ibutes,
xercise
timized
paper, we focus on an interval traini
sensors and a smart phone.
Interval training consists of in
exercises with rest periods. It has
training routines for many years. T
well known exercise protocol whi
improve one’s cardiovascular syst
it helps with weight loss, rehabilita
the reduction of heart diseases. Du
body’s energy production system
aerobic and anaerobic energy sour
from these two sources is then efficiently distributed
throughout the body for the duration of t
Currently, there exists several
but there is no known one-size-
Usually experts recommend “act
people should continue low intensi
periods ([12] – [15]). Billat et al. [
Brooks et al.[14] , and DC Poole et
such low intensity activities during
smart
fitness
in an
mining
raining
ned up
nterval
I. INTRODUCTION

Recent advances in sensors, smart phones a
technology have made a new generation of health
monitoring systems available. The SmartCa
(Vahdatpour [31]) helps patient rehabilitation m
,majid}@cs.ucla.edu, kaiser@ee.ucla.edu
rval
ireless
medical and fitness domai
u

Page 2

will beat. Thus, after starting exercise session, the in
heart rate variable can be observed. In contrast, i
after stopping exercise activity, the heart
decrease( CR Cole [30] ).
Fig 2.1 indicates heart rate variables during
modified Tabata interval training ( Table 3.2
observe that the heart rate increase after starting
decrease after taking a rest. If we can increase th
time until the heart rate reaches a certain high
exercise, it helps people exercise m
cr
mme
rate
22 m
). W
exerc
e am
level
ore without fatigue
if we can decrease the amount of time until the hea
reaches a certain level during rest, it will help incre
amount of exercise time within limited time.
ease in
diately
should
inutes
e can
ise and
ount of
during
. Also,
rt rate
ase the
12123

However, when observing the sa
the same individual (Jansen [25]) s
are dramatically different. Thus it
guess future heart rate values
equations. Instead, it may be bette
probabilistic approach since this
compare and infer certain charac
different heart rate curves of an indi
the information of the heart rate cu
value and the time constant of the cu
the frequency response of the c
individual conditions, we can maxi
an individual performs during
workout. As many studies ha
Christensen [8] there is a correlati
and heart rate. Since each indivi
unique conditions such as age, g
which will in turn affect the heart ra
information to help in the calculatio
heart rate and fatigue level. Also, th
person takes regularly, and the e
taken can drastically affect the
Tuininga [27], Conny M. A [28
[29]). Additionally, we need to tak
of time after having a meal since this could also affect the
heart rate, fatigue level and othe
exercise. In our experiments, we to
since these conditions can dramatic
heart rate level. To monitor heart r
Technologies Bluetooth pulse ox
experiment, which sends heart rat
second.
Accelerometers are currently
studied wearable sensors for activ
al
data obtained from 3 different axes,
and the calorie consumption can be
they are also widely embedded in s
Apple iPhone, Google G1, and No
smartphones can help individuals to

0
1
639
12771915255331913829446751055743638170197657829589339571
0209
1
time ( unit : 0.1 sec )
He
Heart rate variable
during Interval training
100
150
1084711485
ar
te variable
t ra
50
Fig 2.1. Heart rate during 22 minutes Tabata interval training

For example, if you can increase the time to reach he
113 in the first
from 135 seconds to 269 seconds, you can exercise 1
seconds more ( Fig 2.2
amount of rest time, exercise time will be increased.
f 22 minutes Tabata p
n reduce

reach heart rate 113 in 22
Tabata Interval Training Protocol
III. TWO PHASES TO INCREASE THE AMOUNT OF WORK

FOR AN INTERVAL TRAINING PROTOCOL.

Mizuo Mizuo et al. [10] denotes that rising heart rate
curves have an exponential hyperbolic shape, and the falling
heart rate curves are exponential when exercising.

HR of the Rising Curve = Ae-βtsinh(ωt) + C
HR of the Falling Curve = A(1-e-βt) + C

me heart rate curve from
everal times, the curves
is hard to predict and
with only the above
r to use a statistical and
approach can extract,
teristics from the many
vidual. By combining
rve, such as the heart rate
rve which characterizes
urve, along with other
mize the amount of work
an interval training
ve suggested including
on between fatigue level
dual has his or her own
ender, level of activity,
te values, we will use this
n of a user’s optimal
e use of pharmecuticals a
lapse time after they are
heart rate curve (Y S
], LUND-JOHANSEN P
e into account the amount
r conditions related to
ok into account all factors
ally affect an individual's
ate in this research, Alive
imeter was used in our
e variable data every 0.1
among the most widely
ity recognition. They are
so a very useful sensor for interval training. By analyzing
the accuracy of exercise
calculated. Additionally,
mart phones such as the
kia N95. Moreover, these
follow up the scheduled
speed during a specific time period, can also give feedback
using the cell phone’s sound, vibration, and other graphical
interfaces. Additionally, their calculation functionalities can
also be leveraged.
In this study, we used th
development. The iPhone is very light
inch multi-touch display, and su
e iPhone[33] for our
, about 133g, has a 3.5
pports both Wifi and
Bluetooth. Furthermore, the iPhone has an in-built
accelerometer and proximity sensor. By using an embedded
3-axis accelerometer in the iPhone, we can detect one’s
activity pattern.

A. Phase 1: Finding conditions which affect optimal time
constants.

By modifying an already existing interval training method,
Tabata Protocol (Table.3.2), we collected 10 sets of data
which include several attributes which may affect an
individual’s heart rate and exercise potential. During
art rate
rotocol
34
the
exercise period o
). Similarly, if you ca
using

minutes Fig 2.2 Increase in the amount of time to
55 2569
Increase in the amount of exercise time
20
0
80
60
140
120
2141 23
ec )
ate 100
Heart r
40
Modified interval trai
data mining d
1215 429 643 857 1071 1285 1499 1713 1927
Time ( unit : 0.1 s
Initial interval training protocol
ning protocol
ecision tree

Page 3

performing Tabata protocol, a participant should
0.5 second or 0.75 second during certain pe
25cm-tall stool. For rest periods, we as
participants took full rest during the rest period.
minute exercise session datase
shown in Table.3.1. One Tabata protocol is also
4 exercise periods and 3 rest periods.
To maximize the amount of work, the time co
exercise period should be maximized until th
reaches a certain high level. In contrast, the time
a rest period which has the exponential curve c
should be minimized in order to reduce the tim
certain low level heart rate. As you can see in Fig
should change the original exponential sine hype
to the modified one which has a bigger time con
Before starting the exercise session, other att
as age, height, and weight can be recorded in th
During the 22 minute session, the user's heart ra
monitored and recorded, since it is coming into
e. Additionally, we should record the inreal tim
step
riods
sume
E
ts includes 13219 attrib
divid
nstan
e hea
cons
harac
e to r
. 2
rboli
stant
ribute
e dat
te sho
our
accu
exercise after finishing our workout session. In our
accureans the differenc
ount of se
entioned ee
taineding3-axis accelerometer on the ibe ob by us the

ta,
een the sch
se performed.
actual exercis
tual
.1. s in one data set
g Protocol
every
on a
d the
ach 22
utes as
ed into
t of an
rt rate
tant of
teristic
each a
.2, we
c curve
.
s such
a base.
uld be
system
racy of
da
eduled
With these datasets, the knowl
mining techniques provides the ext
of information that we need. Us
mining techniques, a J48 decision t
a powerful data mining tool. The J4
follows the following simple alg
classify a new item, creating a de
attribute values of the available tr
Thus, whenever it encounters a tr
attribute that discriminates the
clearly. Among the possible values
any value for which there is no am
the data instances falling within it
value for the target variable, then
and assign to it the target value th
the other cases, it looks for anothe
highest information gain. Hence it
until it either gets a clear decision
attributes gives a particular targe
attributes. In the event that it runs o
cannot get an unambiguous result from the available
the in
am
m
acy m
exerci
, the sp d or movement in the
e betw
exerci and the ac
AttribuV ue
{ma
female}
A es
Gender
Days afternumeric
le,
period
Age num
oes
done today
e Tim
Inaccuracy
numeric
eric
Number f exercis

Height (cm) numnumeric eric
Weight(kg) num
of exercise
(%)
numeric
eric
As
information, it assigns this branch
majority of th
if more data sets are accumulated i
will be adopted and can ach
classification.

B. Phase 2: Modified interval
constructed J48 decision tree
According to the constructed J48 decision tree in Phase 1,
conditions which extend exercise
periods related to the time const
example, if the user has slept fo
previous night or has eaten with
system can tell the participant to ex
words, the system can tell user to ex
of time in order to maximize the ex
minute duration.
Exercise information such as
a person starts exercising and stop
obtained from Phase 1. In phase
until the maximal heart rate reache
in the exercise period. They shou
e can
Phone.
Table 3

13219 attribute
high}
Disease nam
numeric
e string
Number of exercise
period
Number of rest period Hours of slenumeric ep numeric
Hours after meric … … a meal nu
Medicine
Hours after t
medicine a
string
meric
13200 Heart rate
variables
(every 0.1 s)
numeric
aking
nu
22 minutes Tabata Interval Trainin
Warm up
stepping on the stair per 0.75 second for 5 minutes
6 set
exercises
Rest
6 set
exercises
Rest
6 set
exercises
Rest
Cool down
6 x (stepping on the stair per 0.5 second for 20 seconds +
10 seconds rest)
1 minute rest
6 x (stepping on the stair per 0.5 second for 20 seconds +
10 seconds rest)
1 minute rest
6 x (stepping on the stair per 0.5 second for 20 seconds +
10 seconds rest)
1 minute rest
stepping on the stair per 0.75 second for 5 minutes
Table 3.2.Uncontrolled Exercise Protocol for 10 data in Phase 1.

edge discovery and data
raction and classification
ing data along with data
ree is generated by Weka,
8 Decision tree classifier
orithm. First, in order to
cision tree based on the
aining data is considered.
aining set it identifies the
various instances most
of this feature, if there is
biguity, that is, for which
s category have the same
it terminates that branch
at we have obtained. For
r attribute that gives the
continues in this manner
of what combination of
t value, or it runs out of
ut of attributes, or if it
a target value that the
e items under this branch possess. In addition,
n the database, the system
ieve a more accurate
training based on the
periods and shorten rest
ants were obtained. For
r more than 6 hours the
in the last 2 hours, the
ercise more. Or in other
ercise for a longer period
ercise within the 22
the heart rate value of when
s exercising can also be
2, people should exercise
s the optimal rising curve
ld also start exercising
when the minimal heart rate reaches the optimal falling
curve during the rest period. Participants should satisfy
conditions in decision tree information obtained from phase
1. By combining conditions related to time constants and
increase in the amount of exercise, we can obtain better
result which makes users exercise more without fatigue.
As data is accumulated in the data base, the decision tree
obtained from phase 2 also can be updated. As mentioned,
the decision tree will become more accurate as more data is
accumulated in the data base. This means the decision tree
will be modified and adapted, as the number of data
increases.

tes alttributValue
Activity level
{low,
medium, Time constant numeric

Page 4

Fig.3.1. Interval Training Command Program. This
accelerometer functionality and graphical, sound, vibration
on iPhone.

syste
functi
for seven
pated in five
ified Tabata
ery 0.75 second
2 minutes 30 seconds, 6 x (stepping on the stair every 0.5
second for 20 seconds + 10 seconds rest), 1
fterwards t complete one
satisfying the conditions of the decision
phase. Wob d
2.7ine amnt of calo
dsec s. proves our m
rcise more within a given heart rate threshold rang

inute rest). m
t
bt
d ph se of he exercise
ee o ained
me
bud win 5 m
ds ethohel
of
p peo
Table 4.1 Information about individuals who participated in the modified
Tabata interval training exercise

Studies have shown that an individual's heart rate will
return to a resting rate after 3 minutes. And as such, in our
experiments, individuals were asked to take a 3 minute break
between the 5 different exercise sessions.


m uses
onalities
is from 1.36 to 22.73 percent.

B. 22 minutes Interval Training

Individual #7 in table 4.1. parti
minute modified Tabata interval t
finished 10 sets of 22 minutes Tab
different conditions on different da
one’s conditions in Table 3.1 is stored and used to generate
the J48 tree after completing 10 interval trainings. The
generated J48 tree after finishing 10 Tabata interval trainings
is shown below. This gives the conditions which maximize
the time constant of exercise periods and minimize the time
constants of rest periods. Conditions mentioned in Fig.4.2
can be updated and modified as data is accumulated.



Figure 4.1. Result of the 2 phase.
the standard.) This shows when the
nd
( The amount of work in the 1st phase is
person follows conditons determined by
J48 tree obtained from the 1st phase, the increase of the amount of exercise
for an individual
cipated in the original 22
raining (table 3.2.). She
ata exercise sessions with
ys. Information related to
IV. EXPERIMENTAL RESULTS

A. 5 minutes 30 seconds Interval Training
individuals

Seven different individuals in table 3.1 partici
exercises for the 1st phase which uses the mod
Protocol (Warm up - stepping on the stair ev
for
Aheyd 2na
tr
the 1st
2
an 30
exe
e
ou
This
servean i prov ment
riesrne3 % thith
ond
from
up to
inutes
ple to
e.

Fig 4.2. J48 decision tree which shows conditions which make the time
constant of 4 rising curves which have a characteristic of exponential sine
hyperbolic curves maximize. J48 decision tree which makes the time
constant of 3 falling curves which have a characteristic of exponential
curves minimize are also same. Sleep: the amount of hours to take a sleep
before exercising. Inaccuracy: the inaccuracy between the pedometer data
and scheduled data. Meal: the amount of hours after having a meal.

3 3 4 4 5 5

Person
6 6


Person
7 7

Person
Person

Person
1 1
Person
2 2
PersonPersonPersonPersonPersonPersonPersonPerson
Gender Gender
Age Age
Height Height
(cm) (cm)
Weight Weight
(kg) (kg)
Activity Activity
male
29
female
30
male
27
male
26
female
25
female
25
male
35
170 170 170.8 167.7 158 164 177
62 58 62 79.3 50 50 71
medium
none
low
none
low
none
high
none
medium
none
low
none
low
none
Disease Disease

Page 5

ithin 22 Fig 4.3 The percentage of increase in the amount of
minutes.

The original schedule of modified Tab
training is shown in Fig 4.4. When the partici
conditions in Fig 4.5 and exercises until the c
rate reaches the level in the rising curve wh
maximum
he individual resumed exercising when the
eart rate recovered to the lowest level of the heart rate curve
with the minimum time constant. The changed schedul
exercise is shown on Fig.4.6. We observed that the rest
period of the 2nd phase exercise is sparser and short
the original schedule.

work w
ata interval
pant follows
urrent heart
ich has the
l's previous time constant among the individua
rising curves. T
h
e of
er than
Fig 4.7.
1090.4 sec more than 115.

The results of 3 sets of 22 minutes exercise data satisfying
the J48 decision trees in Fig.4.2 and Fig.4.6 shows more
improvement in the amount of exercise. We see an
improvement of up to 29.54 percent compared with the
results the original protocol.


Fig 4.4. The orginal schedul of Tabata Protocol. Intensity 120
per 0.5 sec. Intensity 75: 1step per 0.75 sec.

: 1 step

Fig 4.5. The changed schedule of Tabata Protocol when following the
constructed J48 decision tree. Intensity 120 : 1 step per 0.5 sec.
Intensity 75: 1step per 0.75 sec.

After finishing the second 10 sets of Tabata interval
training which fulfills conditions in Fig 4.2, new J48 tree
was generated with 20 accumulated interval training data.
The conditions which maximize the amount of work are
shown in Fig.4.6. Also other decision trees which satisfy
conditions in Fig 4.7 are obtained by backtracking. 10 sets of
e second phase shows the
increase of the amount of work up to 24.09 percent (Fig 4.3)

Tabata interval training done in th

Fig 4.6. J48 tree which shows conditions which make the amount of work
maximize. Within_X_Y : X to Y percent increase of the amount of exercise.
TimeX: Heart rate at X time unit. (Each time unit: 0.1 sec)


J48 tree which shows conditions which make heart rate at

Fig 4.8. The percentage of increase in the amount of work within 22
minutes satisfying two J 48 trees in Fig 4.2 and Fig.4.6
C. Improvement in cardiovascular build-up
For an individual, person 7 on table 4.1, the amount of
time required to reach the same heart rate (137) increases as
we repeat the interval training protocol. This means person 7
has adapted to the Tabata interval training protocol, and as a
result, increases their endurance to complete the exercise and
benefits the effect of cardiovascular build-up.