SLIDE 1
!
Clustering physical activity phenotypes using the ATLAS index on accelerometric data from an epidemiologic cohort study
Michael(Marschollek
!
! Sensors for physical activity (PA) measurement ! ! NIKE - - PDF document
! Sensors for physical activity (PA) measurement ! ! NIKE - - PDF document
Clustering physical activity phenotypes using the ATLAS index on accelerometric data from an epidemiologic cohort study Michael(Marschollek ! ! Sensors for physical activity (PA) measurement ! ! NIKE Bodymedia AIRUN+ SenseWear Aipermon
SLIDE 2
SLIDE 3
Michael(Marschollek
! ! !
Physical activity measurement in cohort studies
- NHANES study
- UK Biobank
- Berlin Ageing Study 2
- and many more…
- Common parameters: active energy expenditure (AEE), sedentary
time, number of steps per day, sleep, …
06/2/2011
Ac$Graph AM7164 Geneac$v Humo$on/ Xybermind
SLIDE 4
Michael(Marschollek
! ! !
Physical activity and health outcome
- PA is associated with risk of
– morbidity (cardiovascular/ musculoskeletal diseases; Sattelmair et al., Circulation, 2011 and Woodcock et al., Int J Epidemiol, 2011) and – mortality (e.g. Leitzmann et al., Arch Intern Med, 2007)
- Common PA recommendations include
– AHA/ ACSM > 5 x per week 30 minutes of moderate PA (3-6 METS) or 3 x per week 20 minutes of vigorous PA (> 6 METS) – WHO > 150/75 minutes of moderate/ vigorous PA, better 300/150 mins. – But: only ca. 3% ‚comply‘
- Problem: Dose-effect relationships and appropriate patterns remain
unclear, as are currently predominant PA patterns and their effects.
- Preliminary work: ATLAS index (Activity Types from Long-term
Accelerometric Sensor data; Marschollek, PLOS One, 2013)
– PA parameters: 1. intensity, 2. duration, 3. regularity
06/2/2011
SLIDE 5
Michael(Marschollek
! ! !
Objective
- to apply the ATLAS index on a large cohort accelerometer data set
to investigate whether physical activity archetypes can actually be extracted and if these match the ones proposed earlier by the author
06/2/2011
SLIDE 6
Michael(Marschollek
! ! !
Methods
- Data set
– NHANES (National Health and Nutrition Examination Survey) 2005- 2006, available at: http://www.cdc.gov/nchs/nhanes/nhanes2005- 2006/PAXRAW_D.htm – ActiGraph AM7164 uniaxial accelerometer – N=6386 persons
- Data processing
– Splitting the data set into single patient data sets (handling) – PA event detection – Computation of ATLAS index parameters: intensity, duration, regularity – X-Means clustering (Pelleg and Moore, 2000; WEKA‘s x-Means algorithm implementation, Ver 3.6)
06/2/2011
SLIDE 7
Michael(Marschollek
! ! !
PA parameter distribution
06/2/2011
SLIDE 8
Michael(Marschollek
! ! !
Clustering results
06/2/2011
SLIDE 9
Michael(Marschollek
! ! !
PA clusters
- C2 (N=4264):
– very low mean activity duration (3.1 per 1000 minutes) – very low intensity and regularity scores (1.09, 0.2) – comprises two thirds of the entire population (N=4264), could be named “insufficiently active”, as proposed by Lee at al (Am J Epidemiol, 2004) – inactive, sedentary lifestyle
- C4 (N=1892):
– fair mean activity duration (28 per 1000 minutes;
- ca. 40 minutes/day)
– higher intensity score (1.23) – low regularity score (0.07) – „irregularly active“
06/2/2011
! !
SLIDE 10
Michael(Marschollek
! ! !
PA clusters
- C1:
– low scores for mean activity duration (3.12 per 1000 minutes) and intensity (1.1), but high regularity (3.55) – „busy bee“, constant low-level activity
- C3 (N=72):
– mean duration: 200/1000 minutes or 287 minutes/day – intensity level is very high (1.78), and the regularity is low (0.01) – mixed group: „physical worker“ and „weekend warrior“ (?)
06/2/2011
! ! !
SLIDE 11
Michael(Marschollek
! ! !
Discussion
- method: Identification of distinct PA groups, just using ATLAS
parameters computed from accelerometer data, is feasible, and potentially useful for cohort studies (UK Biobank, German National Cohort, etc).
- Short-term activities, such as daily-life low-intensity ones (e.g. a short
walk, vacuuming, etc.), can be detected from wearable sensor data using algorithms, and can be associated with health-related outcome parameters.
- The ‚athlete group‘ as proposed earlier (Marschollek, PLOS One,
2013) could not be identified.
- Daily and seasonal PA variations so far are not accounted for
- ToDo: investigate associations between cluster groups and health-
related parameters.
06/2/2011
SLIDE 12
Michael(Marschollek
! ! !
Outlook
Inter-group comparisons for health outcome parameters (event duration threshold 5 mins, intensity threshold Q20; submitted to Int J Epidemiol):
- For C-Reactive Protein (CRP) and BMI (but not HDL), ANOVA shows
statistically significant differences between the groups 1-4
- Quartile comparisons e.g. for resting systolic RR (upper [≥ 126mmHg]
- vs. lower [<106mmHg]):
– Statistically significant differences for PA duration and regularity
➢Detecting and denoting low-intensity, short-term activity events is relevant
06/2/2011
SLIDE 13