612
P.-Y. S. Hsueh et al.
13.2 Predictive and Reinforcement Learning for Life
Coaching
In the first theme of case studies, we introduce the potential of collecting heterogeneous personal health data via mobile/IOT e-health technologies and discuss how
to handle the issues regarding the volume, variety, veracity and velocity of such
data sources. In particular, we illustrate a number of smart learning approaches that
help with not only association mining from population-level “big data” but also the
identification of N-of-1 behavioral signature from personal-level “small data.” We
also emphasize on the importance of a smart learning framework and discuss how
policy learning approaches, which are inspired by reinforcement learning tasks such
as Q-learning, can be used to improve interpretability and actionable personalization
insights. The case study will be presented in two examples of application: exercise
behavior prediction and personalized stress management.
13.2.1 Background: Stress-Activity Data
Psychosocial stress contributes to heart diseases by affecting health-related behavior
patterns [8]. Exercise helps reduce psychosocial stress, thus in turn alleviating heart
disease risks [57]. There potentially exists a bidirectional relationship between stress
and exercise behavior. Therefore, understanding how exercise behavior influences
stress can provide insights of developing stress management tools; reversely,
understanding how stress affects exercise behavior can help build healthy lifestyle
interventions. To explore and test the bidirectional relationship, a single cohort,
12-month randomized controlled experiment has been conducted (refer to [58] for a
complete description of the protocol and data curation process). The dataset used in
this case study would be referred to as stress-activity data hereafter.
To achieve personalized stress management, how to collect real-time stress
assessment and exercise statistics continuously is a crucial but challenging problem.
There is a growing interest in using mobile and IOT technologies, such as
smartphones and wearable devices, for collecting health-relevant data and delivering
health guidelines [59]. The stress-activity data are collected from 79 participants
who were followed for up to a 1-year period with physical activity objectively
monitored by actigraphy and stress recorded via ecological momentary assessment
on a mobile app. Physical activity can be continuously monitored via mobile devices
(e.g., Fitbit) over a long time horizon. The Fitbit device tracks users’ daily physical
activity, including step counts, walked distance and burned calories. Fitbit wirelessly
transmits individual activity data in real time, thus preventing loss of data for
individuals who may otherwise not return the device. To collect stress data, an
electronic diary in mobile devices was used to capture momentary and summary
aspects of users’ stress experience. Self-evaluated stress level was recorded at
the end of each day via the electronic diary with specifically designed questions.
P.-Y. S. Hsueh et al.
13.2 Predictive and Reinforcement Learning for Life
Coaching
In the first theme of case studies, we introduce the potential of collecting heterogeneous personal health data via mobile/IOT e-health technologies and discuss how
to handle the issues regarding the volume, variety, veracity and velocity of such
data sources. In particular, we illustrate a number of smart learning approaches that
help with not only association mining from population-level “big data” but also the
identification of N-of-1 behavioral signature from personal-level “small data.” We
also emphasize on the importance of a smart learning framework and discuss how
policy learning approaches, which are inspired by reinforcement learning tasks such
as Q-learning, can be used to improve interpretability and actionable personalization
insights. The case study will be presented in two examples of application: exercise
behavior prediction and personalized stress management.
13.2.1 Background: Stress-Activity Data
Psychosocial stress contributes to heart diseases by affecting health-related behavior
patterns [8]. Exercise helps reduce psychosocial stress, thus in turn alleviating heart
disease risks [57]. There potentially exists a bidirectional relationship between stress
and exercise behavior. Therefore, understanding how exercise behavior influences
stress can provide insights of developing stress management tools; reversely,
understanding how stress affects exercise behavior can help build healthy lifestyle
interventions. To explore and test the bidirectional relationship, a single cohort,
12-month randomized controlled experiment has been conducted (refer to [58] for a
complete description of the protocol and data curation process). The dataset used in
this case study would be referred to as stress-activity data hereafter.
To achieve personalized stress management, how to collect real-time stress
assessment and exercise statistics continuously is a crucial but challenging problem.
There is a growing interest in using mobile and IOT technologies, such as
smartphones and wearable devices, for collecting health-relevant data and delivering
health guidelines [59]. The stress-activity data are collected from 79 participants
who were followed for up to a 1-year period with physical activity objectively
monitored by actigraphy and stress recorded via ecological momentary assessment
on a mobile app. Physical activity can be continuously monitored via mobile devices
(e.g., Fitbit) over a long time horizon. The Fitbit device tracks users’ daily physical
activity, including step counts, walked distance and burned calories. Fitbit wirelessly
transmits individual activity data in real time, thus preventing loss of data for
individuals who may otherwise not return the device. To collect stress data, an
electronic diary in mobile devices was used to capture momentary and summary
aspects of users’ stress experience. Self-evaluated stress level was recorded at
the end of each day via the electronic diary with specifically designed questions.