Importance of Sleep Data in Predicting Next-Day Stress ...€¦ · Importance of Sleep Data in...
1
Hold-out Test accuracies on self-reported stress, happiness, and health: • Without self-reported sleep features: • 79.62%, 78.24%, 83.55%, respectively. • When including self-reported sleep features: • 80.67%, 80.40%, 83.12%, respectively Using McNemar's test we find that including self-reported sleep features does not significantly improve the classifiers for the stress or healthy prediction, but does significantly improve the classifier for the happy prediction (p<0.15). Importance of Sleep Data in Predicting Next-Day Stress, Happiness, and Health in College Students Natasha Jaques 1 *, Sara Taylor 1 *, Ehi Nosakhare 1 , Akane Sano 1 , Elizabeth B. Klerman 2 and Rosalind Picard 1 Introduction Results Acknowledgements The ability to model and predict self-reported stress, happiness, and health could be beneficial for individual and public health, for education, and in the treatment and prevention of mental illness. Importance: Methods References [1] A. Keller et al. Does the perception that stress affects health matter? the association with health and mortality. Health Psychology, 31(5):677, 2012. [2] S. Cohen et al. Psychological stress and susceptibility to the common cold. New England J. of Medicine, 325(9):606–612, 1991. [3] H. Cheng and A. Furnham. Personality, self-esteem, and demographic predictions of happiness and depression. Personality and individual differences, 34(6):921–942, 2003. [4] Ruut Veenhoven. Healthy happiness: Effects of happiness on physical health and the conse- quences for preventive health care. Journal of happiness studies, 9(3):449–469, 2008. [5] A. Bogomolov et al. Daily stress recognition from mobile phone data, weather conditions and individual traits. In Int. Conf. on Multimedia, pages 477–486. ACM, 2014. esearch, 2007. [6] R. LiKamWa et al. Moodscope: building a mood sensor from smartphone usage patterns. In Int. Conf. on Mobile systems, applications, and services, pages 389–402. ACM, 2013. [7] N. Jaques et al. Multi-task, multi-kernel learning for estimating individual wellbeing. In NIPS 2015 Workshop on Multimodal ML, volume 898, 2015. [8] N. Jaques et al. Predicting students’ happiness from physiology, phone, mobility, and behavioral data. In ACII. IEEE, 2015. [9] Y. Xue et al. Multi-task learning for classification with dirichlet process priors. The Journal of Machine Learning Research, 2007. Data Self-reported health strongly relates to actual health and all- cause mortality [1] Stress increases susceptibility to infection and illness [2] Self-reported happiness is indicative of clinical depression [3], and has as strong an effect on longevity as cigarette smoking [4]. Automated mood recognition and prediction are hard: accuracies range from 55-80% (e.g. [5,6]), including the data we use in this work here [7,8]. Physiology: Accel., skin temp., Electrodermal Activity (EDA) Behavioral surveys Conclusions *Both authors contributed equally 144 college students participated in a 30-day study Total of 2,769 days of data were used 2 Division of Sleep and Circadian Disorders, Brigham & Women’s Hospital, Harvard Medical School {elizabeth_klerman}@hms.harvard.edu 1 Affective Computing Group, Media Lab, Massachusetts Institute of Technology {jaquesn, sataylor, ehinosa, akanes, picard} @media.mit.edu We show the accuracy for predicting next-day wellbeing is improved when including simple sleep features. This work was supported by the MIT Media Lab Consortium, NIH R01GM105018 and K24-HL105664, Samsung Electronics, and Canada’s NSERC program. The inclusion of self-reported sleep features improved the prediction of tomorrow’s state of self-reported of sad or happy above the accuracy achieved using features from smartphones and wearables. Changes in stress and health prediction were not statistically significant. Future studies of personalized prediction of mood ought to consider including self-reported sleep features in order to improve prediction. The top and bottom 40% of scores were assigned positive and negative labels, respectively. A hierarchical Bayes machine learning algorithm [9] was trained to predict each next-day wellbeing label on two data sets: (1) including self-reported sleep features (i.e., self-reported sleep latency, bedtime, and wake time), and (2) discarding sleep features. Both data sets include ~20 features computed from wearable sensors, phone, and online surveys. These features include changes in physical activity that are often associated with sleep/wake. Smartphone logs (call, sms, screen, location) Actigraphy Wellbeing reported on visual analog scale • Stressed out – Calm Relaxed • Sad – Happy • Sick – Healthy Figure 1: Distributions: Stressed-Calm, Sad-Happy, and Sick-Healthy
Transcript of Importance of Sleep Data in Predicting Next-Day Stress ...€¦ · Importance of Sleep Data in...
Hold-out Test accuracies on self-reported stress, happiness, and health:• Without self-reported sleep features:
• 79.62%, 78.24%, 83.55%, respectively. • When including self-reported sleep features:
• 80.67%, 80.40%, 83.12%, respectively
Using McNemar's test we find that including self-reported sleep features does not significantly improve the classifiers for the stress or healthy prediction, but does significantly improve the classifier for the happy prediction (p<0.15).
Importance of Sleep Data in Predicting Next-Day Stress, Happiness, and Health in College Students
Natasha Jaques1*, Sara Taylor1*, Ehi Nosakhare1, Akane Sano1, Elizabeth B. Klerman2 and Rosalind Picard1
Introduction
Results
Acknowledgements
The ability to model and predict self-reported stress, happiness, and health could be beneficial for individual and public health, for education, and in the treatment and prevention of mental illness.
Importance:
Methods
References[1] A. Keller et al. Does the perception that stress affects health matter?
the association with health and mortality. Health Psychology, 31(5):677, 2012.
[2] S. Cohen et al. Psychological stress and susceptibility to the common cold. New England J. of Medicine, 325(9):606–612, 1991.
[3] H. Cheng and A. Furnham. Personality, self-esteem, and demographic predictions of happiness and depression. Personality and individual differences, 34(6):921–942, 2003.
[4] Ruut Veenhoven. Healthy happiness: Effects of happiness on physical health and the conse- quences for preventive health care. Journal of happiness studies, 9(3):449–469, 2008.
[5] A. Bogomolov et al. Daily stress recognition from mobile phone data, weather conditions and individual traits. In Int. Conf. on Multimedia, pages 477–486. ACM, 2014. esearch, 2007.
[6] R. LiKamWa et al. Moodscope: building a mood sensor from smartphone usage patterns. In Int. Conf. on Mobile systems, applications, and services, pages 389–402. ACM, 2013.
[7] N. Jaques et al. Multi-task, multi-kernel learning for estimating individual wellbeing. In NIPS 2015 Workshop on Multimodal ML, volume 898, 2015.
[8] N. Jaques et al. Predicting students’ happiness from physiology, phone, mobility, and behavioral data. In ACII. IEEE, 2015.
[9] Y. Xue et al. Multi-task learning for classification with dirichlet process priors. The Journal of Machine Learning Research, 2007.
Data
Self-reported health strongly relates to actual health and all-cause mortality [1]
Stress increases susceptibility to infection and illness [2]
Self-reported happiness is indicative of clinical depression [3], and has as strong an effect on longevity as cigarette smoking [4].
Automated mood recognition and prediction are hard: accuracies range from 55-80% (e.g. [5,6]), including the data we use in this work here [7,8].
Physiology: Accel., skin temp., Electrodermal Activity (EDA)
Behavioral surveys
Conclusions*Both authors contributed equally
144 college students participated in a 30-day studyTotal of 2,769 days of data were used
2 Division of Sleep and Circadian Disorders, Brigham & Women’s Hospital,
Harvard Medical School{elizabeth_klerman}@hms.harvard.edu
1 Affective Computing Group, Media Lab, Massachusetts Institute of Technology
{jaquesn, sataylor, ehinosa, akanes, picard}@media.mit.edu
We show the accuracy for predicting next-day wellbeing is improved when including simple sleep features.
This work was supported by the MIT Media Lab Consortium, NIH R01GM105018 and K24-HL105664, Samsung Electronics, and Canada’s NSERC program.
The inclusion of self-reported sleep features improved the prediction of tomorrow’s state of self-reported of sad or happy above the accuracy achieved using features from smartphones and wearables. Changes in stress and health prediction were not statistically significant.
Future studies of personalized prediction of mood ought to consider including self-reported sleep features in order to improve prediction.
The top and bottom 40% of scores were assigned positive and negative labels, respectively.
A hierarchical Bayes machine learning algorithm [9] was trained to predict each next-day wellbeing label on two data sets:
(1) including self-reported sleep features (i.e., self-reported sleep latency, bedtime, and wake time), and
(2) discarding sleep features.
Both data sets include ~20 features computed from wearable sensors, phone, and online surveys. These features include changes in physical activity that are often associated with sleep/wake.
Smartphone logs(call, sms, screen, location)
Actigraphy
Wellbeing reported on visual analog scale• Stressed out – Calm Relaxed• Sad – Happy• Sick – Healthy
Figure 1: Distributions: Stressed-Calm, Sad-Happy, and Sick-Healthy
