13 Smart Learning Using Big and Small Data for Mobile and IOT e-Health
615
the average effect of a 30-minute bout of exercise on stress reported at the end
of that day is negative and highly significant. This effect has a high variability at
the individual level. In addition, there exists a significant negative average effect
of anticipated stress for a given day on the probability of exercising that day. This
effect also exhibits a high variability at the individual level.
Furthermore, a substantial proportion of individuals demonstrated significant
bidirectional effect of stress and exercise. This indicates that the anticipated stress
is associated with a lower chance for the individual to exercise, and exercising is
associated with a lower reported stress on the same day. However, the significant
effect in either direction is only demonstrated on a small number of individuals.
These findings highlight the limitation of the homogeneous assumption of
individuals and the necessity of investigating stress-behavioral pathways at the
individual level. This case study shows promise to research on personalized
medicine and maintaining health practices at the individual level.
13.2.3 Case Study: Actionable Learning Methods
Mobile and IOT e-health technologies that are able to continuously collect data over
an extended period of time could deliver interventions in an adaptive manner. In this
way, individuals can be provided with personalized coaching plans based on their
ongoing performance. Understanding the relationship between stress and exercise
behavior will facilitate transforming behavior patterns into actionable interventions.
In this case study, we leverage the stress-activity data to explore how to recommend
a sequence of actions to individuals in order to maximize the overall stress reduction.
One solution to that is to model the data through a sequential decision-making
process using reinforcement learning [64]. Particularly, Q-learning is a commonly
used reinforcement learning technique, which uses a sequence of state-actionreward triples to determine a sequence of decision rules to maximize the cumulative
reward. Such decision models have been developed and tailored in the healthcare
area, with regression models to approximate the expected reward given action and
state [39, 40, 42]. Another solution is to model the expected rewards through direct
maximization when a particular form of decision rules is of interest [65–67].
However, as mentioned in the introduction of this chapter, there still exist
multilayer challenges in making the adaptive strategies interpretable. The decision
rules of assigning interventions, i.e., policies, generated by black-box learning
algorithms can be difficult to explain, rendering them unappealing in practical
settings in the healthcare domain.
To make recommendations for e-health users, it is hence imperative to develop
actionable analytics methods that can provide highly relevant and person-specific
insights. As healthcare decisions often involve unclear choices, we need methods
that can account for human psychology to support the implementation of effective
engagement technologies beyond simply addressing clinical efficacy. For example,
[68, 69] have proposed a method to estimate an interpretable policy that guides the
615
the average effect of a 30-minute bout of exercise on stress reported at the end
of that day is negative and highly significant. This effect has a high variability at
the individual level. In addition, there exists a significant negative average effect
of anticipated stress for a given day on the probability of exercising that day. This
effect also exhibits a high variability at the individual level.
Furthermore, a substantial proportion of individuals demonstrated significant
bidirectional effect of stress and exercise. This indicates that the anticipated stress
is associated with a lower chance for the individual to exercise, and exercising is
associated with a lower reported stress on the same day. However, the significant
effect in either direction is only demonstrated on a small number of individuals.
These findings highlight the limitation of the homogeneous assumption of
individuals and the necessity of investigating stress-behavioral pathways at the
individual level. This case study shows promise to research on personalized
medicine and maintaining health practices at the individual level.
13.2.3 Case Study: Actionable Learning Methods
Mobile and IOT e-health technologies that are able to continuously collect data over
an extended period of time could deliver interventions in an adaptive manner. In this
way, individuals can be provided with personalized coaching plans based on their
ongoing performance. Understanding the relationship between stress and exercise
behavior will facilitate transforming behavior patterns into actionable interventions.
In this case study, we leverage the stress-activity data to explore how to recommend
a sequence of actions to individuals in order to maximize the overall stress reduction.
One solution to that is to model the data through a sequential decision-making
process using reinforcement learning [64]. Particularly, Q-learning is a commonly
used reinforcement learning technique, which uses a sequence of state-actionreward triples to determine a sequence of decision rules to maximize the cumulative
reward. Such decision models have been developed and tailored in the healthcare
area, with regression models to approximate the expected reward given action and
state [39, 40, 42]. Another solution is to model the expected rewards through direct
maximization when a particular form of decision rules is of interest [65–67].
However, as mentioned in the introduction of this chapter, there still exist
multilayer challenges in making the adaptive strategies interpretable. The decision
rules of assigning interventions, i.e., policies, generated by black-box learning
algorithms can be difficult to explain, rendering them unappealing in practical
settings in the healthcare domain.
To make recommendations for e-health users, it is hence imperative to develop
actionable analytics methods that can provide highly relevant and person-specific
insights. As healthcare decisions often involve unclear choices, we need methods
that can account for human psychology to support the implementation of effective
engagement technologies beyond simply addressing clinical efficacy. For example,
[68, 69] have proposed a method to estimate an interpretable policy that guides the