Slides for my PhD defense. Title: "Personalization of energy expenditure and cardiorespiratory fitness estimation using
wearable sensors in supervised and unsupervised free-living conditions" - Full text: http://www.marcoaltini.com/uploads/1/3/2/3/13234002/20150919_thesis.pdf
2. PHYSICAL ACTIVITY & HEALTH
• Lack of physical activity is a major
problem today
– Epidemics quickly expanding
(hypertension, diabetes, etc.)
2
3. PHYSICAL ACTIVITY & HEALTH
• Lack of physical activity is a major
problem today
– Epidemics quickly expanding
(hypertension, diabetes, etc.)
• Wearable technology & continuous
monitoring:
– Better understand relations between
physical activity and health
– Drive behavioral change 3
8. Combined data streams
• Higher accuracy
• Detect activities
• Strong link between heart rate,
oxygen uptake and energy
expenditure
WEARABLE TECHNOLOGY FOR
ENERGY EXPENDITURE ESTIMATION
16. • How can we dynamically personalize
heart rate-based models to improve
EE estimation at the individual level?
• Can we move beyond behavioral
aspects of physical activity (e.g. EE,
steps) and estimate cardiorespiratory
fitness as a proxy to health status?
16
RESEARCH QUESTIONS
17. • How can we dynamically personalize
heart rate-based models to improve
EE estimation at the individual level?
• Can we move beyond behavioral
aspects of physical activity (e.g. EE,
steps) and estimate cardiorespiratory
fitness as a proxy to health status?
17
RESEARCH QUESTIONS
18. Current solutions:
• Population based models: everyone is
the same
• Laboratory calibrations are performed
to determine normalization
parameters (e.g. running heart rate)
and personalize models
- i.e. context-specific HR
18
INDIVIDUAL DIFFERENCES IN PHYSIOLOGY
19. • Use wearable sensors and machine
learning methods to determine
context
• Use physiological data during specific
contexts to predict normalization
parameters and personalize EE models
without laboratory calibrations
19
OUR APPROACH
30. Heart Rate Heart Rate Normalized
30
PHYSIOLOGY IS PERSON-SPECIFIC
31. PHYSIOLOGY IS PERSON-SPECIFIC
Heart Rate Heart Rate Normalized
31
dynamic
walking
running
biking
28% 33%29%3%
0.60
kcal/min
0.58
kcal/min
1.13
kcal/min
0.81
kcal/min
1.25
kcal/min
0.89
kcal/min
1.38
kcal/min
0.92
kcal/min
• Reduces error up to 33%
• Does not require individual
calibration or laboratory recordings
32. RESEARCH QUESTIONS
• How can we dynamically personalize
heart rate-based models to improve
EE estimation at the individual level?
• Can we move beyond behavioral
aspects of physical activity (e.g. EE,
steps) and estimate cardiorespiratory
fitness as a proxy to health status?
32
33. CARDIORESPIRATORY FITNESS
ESTIMATION
• Cardiorespiratory fitness is a widely used
marker of overall health
– Higher CRF showing lower risk of all cause
mortality
Current solutions:
• Maximal and submaximal tests: can be
risky for individuals in suboptimal health
conditions, expensive, require medical
supervision, laboratory equipment, spot
measurement only
33
36. HR
CRF model
Activity type,
walking speed,
daily routine
Contextualized
HR
36
CRF ESTIMATION USING
CONTEXT-SPECIFIC HR
37. HR
CRF model
Activity type,
walking speed,
daily routine
Contextualized
HR
CRF
37
CRF ESTIMATION USING
CONTEXT-SPECIFIC HR
38. HR
CRF model
Activity type,
walking speed,
daily routine
Contextualized
HR
CRF
38
• 10.3% error reduction when using
low level context
• 22.6% error reduction when
combining low and high level
context
CRF ESTIMATION USING
CONTEXT-SPECIFIC HR
39. RESEARCH QUESTIONS
• How can we dynamically personalize
heart rate-based models to improve
EE estimation at the individual level?
• Can we move beyond behavioral
aspects of physical activity (e.g. EE,
steps) and estimate cardiorespiratory
fitness as a proxy to health status?
39
40. • How can we dynamically personalize
heart rate-based models to improve
EE estimation at the individual level?
• Can we move beyond behavioral
aspects of physical activity (e.g. EE,
steps) and estimate cardiorespiratory
fitness as a proxy to health status?
40
RESEARCH QUESTIONS
41. • How can we dynamically personalize
heart rate-based models to improve
EE estimation at the individual level?
• Can we move beyond behavioral
aspects of physical activity (e.g. EE,
steps) and estimate cardiorespiratory
fitness as a proxy to health status?
41
RESEARCH QUESTIONS
42. CRF
HR
42
CRF model
EE ESTIMATION PERSONALIZED BY
CRF: HIERARCHICAL MODELS
43. CRF
EE model
43
CRF
HR
CRF model
EE
HR
ACC
EE ESTIMATION PERSONALIZED BY
CRF: HIERARCHICAL MODELS
44. CRF
EE model
44
CRF
HR
CRF model
EE
HR
ACC
EE ESTIMATION PERSONALIZED BY
CRF: HIERARCHICAL MODELS
45. CRF
EE model
45
CRF
HR
CRF model
EE
HR
ACC
EE ESTIMATION PERSONALIZED BY
CRF: HIERARCHICAL MODELS
46. CRF
EE model
46
CRF
HR
CRF model
EE
HR
ACC
EE ESTIMATION PERSONALIZED BY
CRF: HIERARCHICAL MODELS
47. EE ESTIMATION PERSONALIZED BY
CRF: HIERARCHICAL MODELS
CRF
EE model
47
CRF
HR
CRF model
EE
HR
ACC
• No need for explicit
HR normalization
• RMSE reduced by
18% on average
48. CONCLUSIONS
• We personalized EE estimation models
without the need for individual
calibration in laboratory settings
– reduced RMSE up to 33% (HR
normalization and hierarchical modeling)
• We proposed new methods for context
recognition and CRF estimation in free-
living without requiring laboratory tests
– reduced CRF estimation error up to 22.6%
48