Estimating Heart Rate Variation during Walking with Smartphone
Mayu SUMIDA, 〇 Teruhiro MIZUMOTO, Keiichi YASUMOTO
Nara Institute of Science and Technology, Japan
ACM Ubicomp’13, September 8 -12, 2013Zurich, Switzerland
2Estimating Heart Rate Variation during Walking with Smartphone
• Goal: Walking support application for effective walking with appropriate physical load while keeping the walking advantage
• Challenge– Predicting heart rate (HR) with only available functions of a
smart phone to measure physical load
• Idea– Constructing HR prediction model by machine learning
adopting the oxygen uptake as one of input data
• Result– Less than 7 beat per minute mean error for various walking
routes/users
Overview
3Estimating Heart Rate Variation during Walking with Smartphone
Outline
1. Background
2. Related work
3. Heart Rate Prediction Method
4. Evaluation
5. Conclusion
4Estimating Heart Rate Variation during Walking with Smartphone
Background• Walking is not only simple and convenient
Effective for health promotion and maintenance
It is important to walk with appropriate physical load depending on individual physical condition
Walking with high physical load
Walking with low load
[1] Intensity versus duration of physical activity: implications for the metabolic syndrome. a prospective cohort study, BMJ Open (2012).
However
decrease the walking motivation give the risk of injury (to the elderly people, etc)
may result in no effect[1]
5Estimating Heart Rate Variation during Walking with Smartphone
Related Work 1/2
Walking Support System: MPTrain [1]
Regulate HR within an appropriate range during walking
Users have to attach a HR monitor directly on body Simplicity and convenience of walking are spoiled
[1] MPTrain: a music and physiology-based personal trainer, MobileHCI’06 (2006).
HR monitor
6Estimating Heart Rate Variation during Walking with Smartphone
Related Work 2/2
Predict HR from acceleration data by using Neural network
Showed that Neural network is effective to predict HR
Use previous predicted HR to next prediction Error is accumulated every prediction
HR prediction method proposed by Xiao et.al.[2]
[2] Heart Rate Prediction Model Based on Physical Activities Using Evolutionary Neural Network , ICGEC '10 (2010).
Can apply only to daily living situation as HR variation is rather small
7Estimating Heart Rate Variation during Walking with Smartphone
Contribution
• Problem Existing system requires attaching a HR monitor (costly, bother)
Existing HR prediction method cannot be used in walking
Predict heart rate by only available functions of a smart phone and provide effective walking through pace control
• Goal
Devise a heart rate prediction method by a smartphone
8Estimating Heart Rate Variation during Walking with Smartphone
Outline
1. Background
2. Related work
3. Heart Rate Prediction Method
4. Evaluation
5. Conclusion
9
Heart Rate Prediction Method
•How to predict HR?– Construct HR prediction model by machine learning
• What parameters can we use for training data?– Smartphone can measure many information
Estimating Heart Rate Variation during Walking with Smartphone
Light
Acceleration Temperature
HumidityLocation
DirectionStep count
Speed Distance
Gradient
10Estimating Heart Rate Variation during Walking with Smartphone
Consideration of Input Data• Heart rate is related to exercise intensity
– Gradient, walking speed and acceleration are available to predict heart rate
• We constructed model and evaluated HR– Model caused more 10 bpm mean error with actual HR
Walking SpeedGradientAccelerationAmplitude
We searched the parameter more related to exercise intensity and HR
11Estimating Heart Rate Variation during Walking with Smartphone
VO [ml/kg/min]
Time ( s)
Demand
Case of increment
Time( s )
Case of decrement
Demand
• Oxygen Uptake ( VO ) gradually converges to oxygen demand in 2 to 3 minutes[3]
Calculate oxygen demand and determine the trend then estimate VO by using oxygen demand, trend and time
VO [ml/kg/min]
This feature is similar to HR feature
Trend changes by whether oxygen demand increases/decreases
[3]Linear and nonlinear characteristics of oxygen uptake kinetics during heavy exercise. J. of Applied Physiology, 1991.
Oxygen uptake
Devised a novel technique to estimate VO
12Estimating Heart Rate Variation during Walking with Smartphone
How to calculate Oxygen demand• Can calculate oxygen demand by speed and gradient[4]
WalkingSpeed Oxygen
DemandGradient
[4] Lippincott Williams & Wilkins, Philadelphia, ACSM’s Guidelines for Exercise Testing and Prescription (7th edition.)
Walking speed and gradient during walking vary unexpectedly
⇒ It is difficult to calculate VO by Oxygen demand
Periodically calculate oxygen demand with fixed time interval then estimate VO
Accelerometer GPS Gyro
We use dead reckoning
13Estimating Heart Rate Variation during Walking with Smartphone
K1
VO Estimation Method (1/2)
K0
t2t0
VO [ml/kg/min]
time[s]
Current
Previous
Current
( 1 ) Calculate demand
( 3 ) Estimate oxygen uptake variation
( 2 ) Determine trend of variation
by comparing Kc with Kpre
( 1 ) Calculate demand
( 3 ) Estimate oxygen uptake variation
( 2 ) Determine trend of variation
by comparing Kc with Kpre
K1
K0
t1t0
VO [ml/kg/min]
time[s]
Current
Previous
Current Previous
Kpre < Kc Kc < Kpre
Apply incremental model to estimate VO
Apply decremental model to estimate VO
V1
V1Current VO
Current VO
14Estimating Heart Rate Variation during Walking with Smartphone
VO Estimation Method (2/2)
Continue to apply previous trend model to estimate VO
K2 =
K0
t2t0
VO [ml/kg/min]
time[s]
Current
Current
Kc = Kpre
V2
K1Previous
V1
Current VO
t1
Previous
( 1 ) Calculate demand
( 3 ) Estimate oxygen uptake variation
( 2 ) Determine trend of variation
by comparing Kc with Kpre
( 3 ) Estimate oxygen uptake variation
15Estimating Heart Rate Variation during Walking with Smartphone
t3
K3=
• We can obtain oxygen uptake variation by repeating this process
time [s]
V0
0 t1
K1
K2
VO [ml/kg/min]
t2 t4
K4K5=
t5
No change
Up
Up Down
No changeV4
Example of VO Estimation
V5
16Estimating Heart Rate Variation during Walking with Smartphone
Overview of Input Data
• As result of preliminary experiment– Constructing model by gradient, amplitude
(vertical and horizontal direction) and oxygen uptake was the best
Gradient
Speed
Oxygen uptake (VO)
Amplitude
Location
Acceleration
Input Data
Measured valueDead-reckoning
Calculate Oxygen Demand
17Estimating Heart Rate Variation during Walking with Smartphone
Constructing HR Prediction Model
• Construct model by three-layered neural network• There is no liner relation ship between HR and input data
18Estimating Heart Rate Variation during Walking with Smartphone
Outline
1. Background
2. Related work
3. Heart Rate Prediction Method
4. Evaluation
5. Conclusion
19Estimating Heart Rate Variation during Walking with Smartphone
Purpose and Setting for EvaluationPurpose
Setting
• Evaluate the heart rate prediction accuracy of our method
• 18 subjects ( twenties / 15 male, 3 female )• Each subject walked 5 different routes to collect data
We extracted accurate altitude from the map published by government
20Estimating Heart Rate Variation during Walking with Smartphone
Devices for Collecting Data
Hardware Sensor Sampling time
SUUNTO t6dHeart ratemonitor 2s
Xperia active 3-axesaccelerometer 20ms
GPS 3s
Y-axis
X-axis
• We asked each subject to equip • A smart phone to measure acceleration and location• A heart rate monitor to measure heart rate as training data
21Estimating Heart Rate Variation during Walking with Smartphone
Model• Collected 90 training data (18 subjects×5 routes)
acceleration amplitude, gradient, VO and measured HR
• Construct model of each subject of each route– The prediction when a new user walks on a new route
Subject S
Remaining 17 subjects
Route R
Remaining 4 routesUse 68 data as training data
If we evaluate the model of subject S of Route R
×
Use the data of subject S of Route R as test data
Estimating Heart Rate Variation during Walking with Smartphone
Accuracy Definition22
time [s]
HR
[bpm
]
measured predicted
• We calculate mean absolute error as accuracy• Absolute error: the difference between measured HR by HR
monitor and predicted HR by our model every 24 seconds
Absolute error
The minimum time that we can use all parameters in same time
23Estimating Heart Rate Variation during Walking with Smartphone
• Borg Scale[5] classifies physical load into 15 levels (6 ~ 20) called RPE
[5] Psychophysical scaling with applications in physical work and the perception of exertion, Scandinavian Journal of Work Environment Health (1990).[6] Perceived exertion: a note on ”history” and methods. ACSM J. of Med Sci Sports Exerc.(1973).
• RPE (Ratings of perceived exertion) corresponds to one tenth of HR [6]
Borg ScaleIf error is less than 10 bpm, difference of physical load is low
Physical Load scaling method
24Estimating Heart Rate Variation during Walking with Smartphone
Accuracy of Each Subject
1 2 3 4 5 6 7 8 9 1011121314151617180
2
4
6
8
10
Mea
n Ab
solu
te E
rror
[b
pm]
Subject
• All of subject were less than 10 bpm MAE• We achieved average 6.78 bpm MAE
HR is known that it varies 7bpm even during rest situation
6.78 bpm
25Estimating Heart Rate Variation during Walking with Smartphone
Accuracy of Each Route
A B C D E0
2
4
6
8
10
7.576.46
5.55
9.05
5.25
MAE
[bpm
]
Route
• All of route were less than 10 bpm MAE• Routes with down slope or flat were more less error• Almost subjects with low accuracy were low accuracy
in route A and D having steep slope, especially 6.78 bpm
Our method can predict HR with low error even if a new user walks on a new route
26Estimating Heart Rate Variation during Walking with Smartphone
Example of Heart Rate Prediction
Almost accurately follow HR variation
Route A by subject 103.48bpm mean absolute error
Altitude of Route A
Altit
ude[
m]
Predicted
27Estimating Heart Rate Variation during Walking with Smartphone
Example of Heart Rate Prediction
Route A by subject 216.69 bpm MAE
Altitude of Route A
Altit
ude[
m]
Predicted low HR in up slope
PredictedPredicted appropriate HR in down slope
28Estimating Heart Rate Variation during Walking with Smartphone
Effectiveness of oxygen uptake
• We also evaluated effectiveness of introducing oxygen uptake (VO) by other data set– The MAE without VO and with VO were 16.71
bpm and 6.41 bpm
Using oxygen uptake as a parameter is effective for HR prediction
MAE
16.71
6.41
29Estimating Heart Rate Variation during Walking with Smartphone
Conclusion
• Heart rate prediction method for walking support system by a smartphone• Adopt Oxygen uptake similar to the feature of heart rate
variation to train model by neural network
• Our method could estimate the HR with accuracy of about 6.78 bpm on average when 18 subjects walked on 5 routes
We considers user’s condition, weather (temp and humid) , etc.
Future work
From evaluation result