passive sensing of circadian rhythms for individualized...
TRANSCRIPT
Passive sensing of circadian rhythms for individualized models
of cognitive performance
Julie Kientz, Tanzeem Choudhury
Saeed Abdullah, Elizabeth Murnane, Mark Matthews, Matt Kay
Cognitive capabilities vary over time
2
Alertness: basic building block of cognitive performance
3
fatigue and sleepiness =
alcohol intoxication
4
Fatigue is involved in 30% of all road accidents in US
NTSB. Safety report NTSB/SR-99/01
5
36% increase in serious medical errors
6
Williamson, Ann, et al. "The link between fatigue and safety." Accident Analysis & Prevention 43.2 (2011): 498-515.
Negative impact on learning and problem solving
7
8
Alertness
Chronotype
Sleep CircadianMisalignment
Stimulants
9
Circadian Rhythm: biological processes following a roughly 24-hour period
circa: about, diem: a day
10
Almost every neurobehavioral process displays circadian rhythms
11
Chronotype: Individual differences in temporal preference resulting from circadian rhythms (early and late types)
12
Circadian Rhythms and Alertness
• Internal time dictates optimal peak alertness period
• Alertness drops during mid-day dip
• Sleep is a crucial factor
13
Studying alertness beyond controlled lab environment
14
Continuous assessment of alertness based on in-situ data in a real-world setup
15
• How do body clock, time of day, and stimulant intake impact alertness?
• Do phone usage patterns reflect fatigue and sleepiness?
• Can we automatically assess alertness using passively sensed phone data?
16
Methodology
17
Population
• University-aged individuals
• Massive risk of circadian misalignment
• Largest and most habituated technology users
18
Participants & Procedure• 20 participants
• 7 male, 13 female
• 18-29 years old
• Android users
• 40 days
19
• Data
• Daily sleep diary
• 4-times-per-day alertness assessment (EMA)
• Phone use logs
• Interviews
Sleep Data
MSFSC = MSF −0.5(SDF −(5∗SDW +2∗SDF)/7)
1 2 3 4 5 6 7 8 9
35
0
5
10
15
20
25
30
Chronotype
% o
f Sam
ple
Larks Owls
extremeEarlytype
moderateEarlytype
slightEarlytype
Normaltype
slightLatetype
moderateLatetype
extremeLatetype
Chronotype:
20
21
Smartphone Toolkit
22
PVT
From PVT to Alertness
• Median response time from a PVT session
• Establish individual baseline across all session
• Alertness is departure from baseline
23
Results
24
Alertness varies across time
25
Early and late types have different performance pattern
26
Daylight saving time (DST)
• Social clock-shifting
• Known to cause circadian disruptions
• 70 countries observe DST, impacting 1.6 billion people
27
Negative impact of DST
28
Stimulant intake
• Positive stimulants
- Caffeine intake, napping, doing exercise, nicotine intake
• Negative stimulants
- Alcohol consumption, having meals, relaxation
29
Stimulant intake
• 5.1% increase after positive stimulants
• 1.37% drop after negative stimulants
• Statistically significant (t = 2.2, p = 0.03)
30
Self-Assessment of Alertness
• Self-assessment
- Tiredness, energy and concentration level
• Response time differs significantly between high and low self-ratings
- fatigued individuals are usually aware of reduced capability
32
230 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22
2500
0
500
1000
1500
2000
App
licat
ion
Usa
ge E
vent
s
Entertainment Time & WeatherCommunicationProductivity BrowsingEmailSocial Media
Rhythms in App Use
33
Productivity vs. Entertainment
20
0
5
10
15
% o
f Usa
ge E
vent
s
EntertainmentProductivity
Mon Tues Wed Thu Fri Sat Sun
• Work days
• Free days
• Mid-week dip
34
Usage Differences by Chronotype
• Early Types
• 25% more productivity apps
• 18-28% fewer entertainment apps
• Late Types
• 22-68% more productivity apps
• 15-50% less entertainment apps
100
-100
-80-60-40-20
020
406080
Earl
y-La
te U
sage
Cha
nge
(%)
Morning(6AM-12PM)
Afternoon(12PM-6PM)
Evening(6PM-12AM)
Night(12AM-6AM)
EntertainmentProductivity
35
Internal Time
230 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22
1
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
InT
Performance
6
-8
-7
-6
-5
-4
-3
-2
-10
1
2
3
4
5
Usage
Entertainment CMC
Alertness
Productivity
InT = ExT - MSFSC
“To wake myself up, I’ll have to look at things on the phone like Facebook or Tumblr.”
“In morning classes, I have less attention and am very tired so I’ll browse the phone. Using tactics like social media, I focus on the screen to try to keep my eyes open.”
———
“Every time before I go to bed, I play a card game until I feel sleepy.”
“I use my phone when falling asleep. Especially if I’m having trouble falling asleep, I’ll play a game or talk to my boyfriend until I fall asleep.”
36
App Use and Sleep
• Less sleep: less productivity (r=0.43), more entertainment apps (r=-0.19)
• Adequate sleep: 61% more productivity apps
• Inadequate sleep: 33% more entertainment apps
• Nightly use events reflect sleep interruptions
37
Predicting alertness from phone data
• PVT is not suitable for longitudinal deployment
• Passive inferring of alertness can enable a new suite of HCI applications
• Can data from mobile phone predict alertness?
38
Predicting Alertness• Stochastic Gradient Descent (SGD) with Huber loss
function
• Standardize all features to have zero mean and unit variance
• L1-norm as regularization term
- α = 10-8
- learning rate: γt = γ0 · t−1/4 (with γ0 = 0.01)
39
Predicting Alertness
• 10 fold cross-validation
• RMSE of 11.39 across all participants
• Accurate enough for scalable deployment
Internal Time
Avg. time between phone usage sessions
Short Session frequency
Phone usage duration
Relative sleep need
Top-ranking features for predicting alertness
40
Implications & Applications
41
Task Scheduling
• When to do what?
- based on cognitive demand and assessed alertness
• Better team collaboration
- grouping members with similar circadian characteristics
42
Learning and Education
• Circadian disruptions adversely affect memory and learning abilities
• Learning and memorization aligned with individual alertness rhythms in school
43
Accident Prevention
• Assistive systems for drivers
• Continuous monitoring to prevent industrial accidents
44
Future work: • Circadian-aware technology • Informatics tools & intervention studies
Contributions: • In-situ alertness sensing • Manifestations of biological rhythms in mobile use • Automated alertness prediction
45