yan kestens - methodologies for collection and analysis of gps data for health research
DESCRIPTION
Keynote address of the 'Spatial Analysis of GPS Data Workshop' held at Exeter, UK, the 16-17 May 2013.TRANSCRIPT
Methodologies for collection and analysis of GPS data
for health research
Yan KestensMontreal University, Social and Preventive Medicine
Montreal Hospital University Research Center (CRCHUM)
SPHERE Lab .org
Spatial analysis of GPS data, Exeter University, UK16th May 2013
Context
CHANGES
• Recent push in health research along the ‘space-time’ continuum
• A consequence/correlate of society’s ‘space-time convergence’ ?
• Space: From ‘place based’ to ‘people based’
• Time: From snapshots to continuous measures, from delay between collection and results to ‘real-time’
• Convergence of space and time, convergence of fields
• New methods, new developments, new continuums
Context
POTENTIAL
• Local trap / residential trap
• Space-time geography
• Potential path areas
• Activity spaces
• Network of usual places
• Multiple exposures
Context
TECHNOLOGICAL CHANGES
• Wearable sensors
• Ubiquity
• Connectedness
• 7 billion sensors
• Quantified Self - mHealth
Context
‘More data more often’
vs.
‘Less is more’
Spatial data collection for health
CAPTURE
PROCESSINGUSAGE
Web server
Acquisition server
Outputs /Applications
End users
GISAlgorithms
GSM towerSensors
Issues with GPS data capture
CAPTURE
Participation/adherence: privacy, participation burden
Device usage: co-occurrence, lose vs. on body, devicemanipulation
Device performance: Battery life, data storage space,precision/validity of data points (TTFF, drift, indoor/outdoor)
Temporal aspects: Epoch, survey duration, linkage with other sensors and GIS data
GPS data capture
CAPTURE
Most current devicesGPS trackersLow battery lifePb of integration with additional sensorsLimited capacity of data transmissionNot designed for health research
GPS in cellphonesBattery life major hinderSimultaneous usage with other applications not alwayspossible
GPS data capture
CAPTURE
Attempts to address these issues
Collaborations with engineers
Validation requirements
SenseDoc Multisensor Device
CAPTURE
Central Unit
GPS GPRSAccelerometer
Acquisition server
Central Unit
GPS GPRSAccelerometer
ANT ModuleMemory
SPHERE Lab .org
125 g
137 g
96 * 80 mm
115 * 59 mm
SenseDoc Multisensor Device
CAPTURE
Central Unit
GPS GPRS
Memory
Accelerometer
ANT Module
Acquisition server
Central Unit
GPS GPRS
Memory
Accelerometer
ANT Module
SenseDoc Multisensor Device
CAPTURE
Central Unit
GPS GPRS
Memory
Accelerometer
ANT Module
Acquisition server
Central Unit
GPS GPRS
Memory
Accelerometer
ANT Module
Glucose monitor
Galvanic skin response
Accele-rometer
HR monitorBlood
pressure
Other
SenseDoc Multisensor Device
CAPTURE
Acquisition server
GPS – SIRF IV
GPS performance validation
Spatial accuracy
Time to First Fix (TTFF)
Indoor – Outdoor
Fixed - Moving
CAPTURE
Average of dist_moy Column Labels
Row Labels Etrex HTC MS Qstarz Grand Total
Indoor
cold 13,6 9,0 7,7 16,7 12,3
Brick building, hallway 14,1 5,7 4,1 15,4 10,4
Brick building, window 14,4 12,4 7,6 15,5 12,5
Concrete building, window 12,3 11,6 19,4 14,4
hot 12,9 11,3 10,0 15,5 12,6
Brick building, hallway 11,2 7,2 6,3 15,8 10,5
Brick building, window 7,6 5,0 6,9 12,8 8,5
Concrete building, window 19,9 21,8 16,9 17,9 18,7
warm 14,0 13,5 20,3 15,8 16,1
Brick building, hallway 10,4 15,6 22,1 11,1 14,8
Brick building, window 7,8 10,4 21,7 13,2 13,7
Concrete building, window 23,8 12,2 17,0 23,0 20,0
Outdoor
cold 7,8 16,6 11,0 17,6 13,0
Narrow streets 21,4 20,0 16,2 35,3 23,2
Open surroundings 2,8 12,0 1,4 1,1 4,3
Residential areas 2,4 4,1 0,9 2,5
Sky scrapers 4,7 17,8 22,2 33,0 19,4
hot 5,5 10,6 3,4 4,8 6,1
Narrow streets 12,9 18,6 4,9 9,5 11,5
Open surroundings 1,5 1,8 2,2 1,8 1,8
Residential areas 3,2 3,4 1,9 3,1 2,9
Sky scrapers 4,4 18,4 4,6 4,8 8,5
warm 8,6 9,1 6,5 10,0 8,5
Narrow streets 26,7 21,9 16,2 20,5 21,3
Open surroundings 3,0 5,4 3,3 4,1 3,9
Residential areas 4,1 4,6 2,8 5,0 4,1
Sky scrapers 5,0 8,9 7,4 15,2 9,1
Grand Total 10,3 11,4 9,5 13,0 11,0
CAPTURE
Average of ttff Column Labels
Row Labels Etrex HTC MS Qstarz Grand Total
Indoor
cold 136,3 255,0 33,2 86,3 102,3
Brick building, hallway 68,0 104,0 12,5 23,0 44,4
Brick building, window 252,0 406,0 9,5 193,0 187,9
Concrete building, window 89,0 77,5 43,0 69,8
hot 18,5 181,3 5,5 13,5 36,6
Brick building, hallway 6,5 82,0 6,0 2,5 16,0
Brick building, window 41,0 143,0 4,0 35,0 43,3
Concrete building, window 8,0 319,0 6,5 3,0 50,6
warm 101,7 293,3 46,5 204,7 149,5
Brick building, hallway 27,0 563,5 0,0 69,0 164,9
Brick building, window 107,0 26,0 84,5 191,0 113,0
Concrete building, window 171,0 20,0 55,0 354,0 168,6
Outdoor
cold 37,8 171,7 26,0 40,5 62,1
Narrow streets 44,0 247,0 36,0 40,0 91,8
Open surroundings 39,0 104,0 37,0 57,0 59,3
Residential areas 26,0 20,0 26,0 24,0
Sky scrapers 42,0 164,0 11,0 39,0 64,0
hot 16,5 36,1 21,9 10,1 21,5
Narrow streets 11,5 110,0 29,0 12,5 40,8
Open surroundings 10,5 15,0 4,5 1,0 7,8
Residential areas 8,5 10,0 7,5 3,0 7,3
Sky scrapers 35,5 9,5 46,5 38,0 31,6
warm 26,4 46,8 39,4 31,6 36,1
Narrow streets 40,0 45,0 45,0 40,0 42,5
Open surroundings 21,0 36,0 45,0 35,0 34,3
Residential areas 30,0 68,5 44,5 29,5 43,1
Sky scrapers 11,0 16,0 18,0 24,0 17,3
Grand Total 55,8 130,6 28,2 65,2 65,3
SenseDoc Multisensor Device
CAPTURE
Accelerometer
Marie-Lyse Bélanger, M.Sc. Student in kinesiologyAccelerometer validation using indirect calorimetryLab – 14 controlled exercises from sedentary to vigouros PAEleven adult subjectsCalculation of Vertical Magnitude Acceleration (VMAG)Testing of various bandpass filtersComparison with Actigraph GT3X performence
Best results obtained with Bandpass filter 0.1 Hz – 3.5 HzModelling of Energy Expenditure: Adj. R-square of .79Use of Vector Body Dynamic Acceleration (VEDBA)
SenseDoc Multisensor Device
CAPTURE
Battery life
Strong battery (3200 maH)
Axelle Chevallier, M.Sc. Student in electrical engineeringMohamad Sawan,
Battery optimisation algorithm- Movement- Location and movement
SenseDoc Multisensor Device
CAPTURE
Acquisition server
Battery life
SenseDoc Multisensor Device
CAPTURE
Data transmission
GPS Data sent over the air (cellphone network) every 30 minutes
Possible alerts depending on - Location- Activity- Time
Connection to other sensors (2.4 GHz ANT+) Heart rate monitor, footpod, RFID tags, etc.
SenseDoc Multisensor Device
CAPTURE
Challenges in developping new hardware
Hardware / Software / User Interface
Miniaturisation
From prototype to market
Challenges
Web server
Acquisition server
Outputs /Applications
End users
GISAlgorithms
GSM towerSensors
CAPTURE
PROCESSING
USAGE
Spatial data collection for health
CAPTURE
PROCESSINGUSAGE
Issues in data processing
PROCESSING
Continuous monitoring = Huge pile of data!!!
Issues in data processing
PROCESSING
Continuous monitoring = Huge pile of data!!!
Issues in data processing
PROCESSING
Continuous monitoring = Huge pile of data!!!
Issues in data processing
PROCESSING
Transforming raw GPS data into meaningful and usefulinformation
- ‘Putting things into context’- Activity locations- Trips between locations
Activity location detection
PROCESSING
Development of kernel-density based algorithm to transform raw data into history of activities and trips
ArcGIS ArcToolBox (see www.spherelab.org toolssection)
- Input: raw GPS data- Output:
- Location of activity places- Activity places timetable- Trip timetable with origins and destinations
Thierry et al. (2013) IJHG
Issues in data processing
PROCESSING
Activity location algorithm validation method
- Artificial track generation with controlledparameters (noise, stop time)
- Testing of algorithm performance in relation to track characteristics and algorithm parameters
Thierry et al. (2013) IJHG
Issues in data processing
PROCESSING
Activity location detection kernel-based algorithm
Thierry et al. (2013) IJHG
Aft Akd
(a) M
ean
num
bero
f sto
ps fo
und
(b) M
ean
dist
ance
to tr
uest
op(c
) Ave
rage
time
diff.
rela
tive
totr
uest
op d
urat
ion
*10m radius missing
*10m radius missing
Noise categories Noise categories
Issues in data processing
PROCESSING
Activity location detection kernel-based algorithm
Thierry et al. (2013) IJHG
Aft Akd
(a) M
ean
num
bero
f sto
ps fo
und
(b) M
ean
dist
ance
to tr
uest
op(c
) Ave
rage
time
diff.
rela
tive
totr
uest
op d
urat
ion
*10m radius missing
*10m radius missing
Noise categories Noise categories
Aft Akd
(a) M
ean
num
bero
f sto
ps fo
und
(b) M
ean
dist
ance
to tr
uest
op(c
) Ave
rage
time
diff.
rela
tive
totr
uest
op d
urat
ion
*10m radius missing
*10m radius missing
Noise categories Noise categories
Issues in data processing
PROCESSING
Activity location detection kernel-based algorithm
Thierry et al. (2013) IJHG
Aft Akd
(a) M
ean
num
bero
f sto
ps fo
und
(b) M
ean
dist
ance
to tr
uest
op(c
) Ave
rage
time
diff.
rela
tive
totr
uest
op d
urat
ion
*10m radius missing
*10m radius missing
Noise categories Noise categories
Aft Akd
(a) M
ean
num
bero
f sto
ps fo
und
(b) M
ean
dist
ance
to tr
uest
op(c
) Ave
rage
time
diff.
rela
tive
totr
uest
op d
urat
ion
*10m radius missing
*10m radius missing
Noise categories Noise categories
Issues in data processing
PROCESSING
Activity location detection kernel-based algorithm
Thierry et al. (2013) IJHG
Aft Akd
(a) M
ean
num
bero
f sto
ps fo
und
(b) M
ean
dist
ance
to tr
uest
op(c
) Ave
rage
time
diff.
rela
tive
totr
uest
op d
urat
ion
*10m radius missing
*10m radius missing
Noise categories Noise categories
Issues in data processing
PROCESSING
Linkage of GPS locations with additional information
- Temporal linkage: - Additional wearable sensors (Accelerometer,
Heart rate monitor, continuous glucose monitor)
- Spatial linkage: GIS data – exposure at any givenlocation/time – descriptive vs. causal understanding
- Spatio-temporal linkage – spatio-temporal GIS
Issues in data processing
PROCESSING
Mapping – visualisation
Tool for communication / counseling, etc.
Issue of privacy – artificial blurring
Usage
Using GPS to locate behaviour and assess exposure
Improving the understanding of mechanisms linkingenvironments to health behaviours and profiles
Using GPS to prompt recall and gain additional insight
Using GPS to support qualitative studies (go-along, geo-ethnography, geo-tagged photos, environmentalperception, etc.)
Using GPS data to assist clinical practice (mHealth)
USAGE
Usage: Prompted recall
GPS / accelerometer data provides limited information on:
- What people are actually doing- Decision processes- How they feel
GPS-prompted recall can help gather additionalinformation
USAGE
Usage: Prompted recall
Example 1: Bike share study (PI: Gauvin)
Pilot study (n=25) on combined use of cellphones and accelerometers for gathering of:
- GPS data- Nature of activities- Transportation modes- Accelerometry (PA)- Momentary Impact Assessment (feelings)
USAGE
Usage: Prompted recall
Bike share studyN=25, study period=7 days
Accelerometer:- PA assessment
Cellphone:- GPS data – sent to server every hour- Feelings (real-time questionnaires)
Daily online prompted recall data collection using the MWM (Mobility Web Mapping) application:- History of mobility- Nature of activities- Transportation modes
USAGE
Usage: Prompted recall
Example 2: RECORD GPS Study (Chaix & Kestens)
GPS + Accelerometer
MWM prompted recall survey (Mobility Web Mapping) after reception and processing of GPS data:
- Validation of activity places and trips- Nature of activities- Transportation modes
USAGE
Usage: Prompted recall
MWM prompted data a useful tool to improve activitylocation algorithm / data collection
- Match/mismatch between algorithm detection and reported timetable (locations/times)
- Preliminary analyses: N=80
USAGE
88.5%
11.5%
GPS raw data
GPS data
Missing data
Usage: Prompted recall
Analyses comparing activity locations, trips and corresponding timetables obtained through:
- Spherelab GPS algorithm- MWM GPS-prompted recall
N=80Median of 88.5% of survey period with usable (raw and interpolated) GPS data (11.5% of period with missing data)
USAGE
88.5%
11.5%
Proportion of survey time with GPS data
GPS data
Missing data
Usage: Prompted recallUSAGE
MWM Algorithm
A A <50m
A A >50m
A T
T A
T T
AA >50m Misplaced activity
AA-AT-TT Early departure
AA-AT-AA False trip
AA-TA-AA Missing trip
AA-TA-TT Late departure
TT-AT-TT Late arrival
TT-TA-TT False positive
TT-AT-TT False negative
TT-TA-AA Early arrival
Usage: Prompted recallUSAGE
96.0%
Correctly classified
Misplaced activity location
Early departure
False trip
Missing trip
Late departure
Late arrival
False positive
False negative
Early arrival
Other
4%
Proportion of valid GPS time with match / mismatch withprompted recall data
USAGE
+ +
Trimble Juno SC GPS + Arcpad
Actigraph GT3X
Polar HR monitor
7-day data collection
Usage: Support for clinical interventions
Dyn@mo Intervention – CIRCUIT Clinic – Ste-Justine Pediatric Hospital (PI: MH Henderson)
Spatio-behaviouralindicators -ArcToolBox
Interactive map-based web application
Application supports lifestyle counseling
USAGEDyn@mo Intervention – CIRCUIT Clinic – Ste-Justine Pediatric Hospital (PI: M. Henderson)
Usage: Support for clinical interventions
USAGEDyn@mo Intervention – CIRCUIT Clinic – Ste-Justine Pediatric Hospital (PI: M. Henderson)
Usage: Support for clinical interventions
USAGEDyn@mo Intervention – CIRCUIT Clinic – Ste-Justine Pediatric Hospital (PI: M. Henderson)
Usage: Support for clinical interventions
How does GPS compare to regulardestinations?
Comparing spatial distribution of- 7-day GPS data- Regular destinations collected through an online interactive
mapping questionnaire (VERITAS)
How does GPS compare to regulardestinations?
89 participants of the RECORD GPS Study
VERITAS activity locations • Total of 1,314 locations• Median of 14 loc./ind.
How does GPS compare to regulardestinations?
89 participants of the RECORD GPS Study
7-day continuous GPS monitoring
0
10
20
30
40
50
60
70
80
90
100
Pe
rce
nta
ge o
f su
rvey
du
rati
on
w
ith
GP
S fi
xes
Proportion of GPS survey duration with valid GPS data
5 Days & 07:07:153 Days & 10:25:25
6 Days & 04:45:20
5 Days & 07:07:15
1000
250
500
100
0.00
0.10
0.20
0.30
0.40
0.50
0.60
0.70
0.80
0.90
1.00
Within 100 m Within 250 m Within 500 m Within 1000m
Proportion of total survey time spent within range of VERITAS locations
87%
85%
78%
66%
1000
250
500
1000
100
0
20
40
60
80
100
120
140
Shortest distance between a GPS detected location (unspecified category)
and a VERITAS location (specified category)
(median value; n=1,314)
VERITAS CATEGORIES
Dis
tan
ce in
met
ers
CONCLUSIONS
GPS opens great possibilities
Capture – Processing – Usage
Multidisciplinarity – Health – Transportation - Geography –Engineering
Applications very diverse
For epidemiology – validity is key
Thank you!
SPHERE Lab .org
Benoit Thierry from SPHERELAB Claire Merrien from RECORD
Participants of all the studies!