610
P.-Y. S. Hsueh et al.
to guide a target user’s choice of actions by learning from real-world experiences
leading to desired health outcomes. The other regards how to make the learned
strategies more interpretable by representing the decision rules in a simpler form
that can be digested by users with ease as well as by providing the target user
with actionable goal-setting guidance, e.g., the predictive sub-goal performance on
outcome improvement. The multilayer challenges current smart learning methods,
which are often biased and uninterpretable [37, 38], to deliver interpretable actionable insights to individuals.
To address the first challenge, researchers could leverage a long line of studies
in adaptive clinical trial and adaptive treatment regime, which have established the
benefits of incorporating sequential multiple adaptive randomized trial (SMART)
methods to achieve best mean outcomes [39–42]. More recently, researchers have
considered supporting just-in-time delivery of adaptive interventions [43, 44]. To
address the second challenge, researchers have started developing interpretable
machine learning models for healthcare applications [45]. However, most of these
studies define interpretability as a measure of model characteristics (e.g., complexity) learned by the algorithms. The problem of such an operational definition is its
disconnection with actual user interpretation [46, 47].
To address both challenges simultaneously, most recently policy learning methods such as reinforcement learning are being studied to lend support to the
generation of easy-to-interpret adaptive strategies according to the observations of
target users [48, 49].
Another great domain wherein smart learning for mobile and IOT e-health can
make great impacts is to provide systematic support for operationalizing patient
education and coaching programs in daily-life settings. The prevalence of low health
literacy and self-efficacy in the population is one of the central problems in the
healthcare system. Due to the service-oriented nature of strict time scheduling,
physicians often cannot contribute significantly to patient education. A possible
solution is the use of automated educational systems, which adapt the delivery
of knowledge according to a learner’s needs and level of knowledge. Although
adaptive automated educational systems for other topics have been applied in
practice, the adaptive automated patient education system practically does not exist.
Several challenges have emerged therein. The key challenge lies on its long
development phase and non-standardized evaluation. Additionally, specific limitations in the domain of healthcare lie on the needed amount of personal information
about the user and the number of related time-consuming assessments on the
patient’s side. Furthermore, there exist challenges on the bandwidth of healthcare
experts needed for implementing such systems and the difficulties in explicating
their knowledge.
P.-Y. S. Hsueh et al.
to guide a target user’s choice of actions by learning from real-world experiences
leading to desired health outcomes. The other regards how to make the learned
strategies more interpretable by representing the decision rules in a simpler form
that can be digested by users with ease as well as by providing the target user
with actionable goal-setting guidance, e.g., the predictive sub-goal performance on
outcome improvement. The multilayer challenges current smart learning methods,
which are often biased and uninterpretable [37, 38], to deliver interpretable actionable insights to individuals.
To address the first challenge, researchers could leverage a long line of studies
in adaptive clinical trial and adaptive treatment regime, which have established the
benefits of incorporating sequential multiple adaptive randomized trial (SMART)
methods to achieve best mean outcomes [39–42]. More recently, researchers have
considered supporting just-in-time delivery of adaptive interventions [43, 44]. To
address the second challenge, researchers have started developing interpretable
machine learning models for healthcare applications [45]. However, most of these
studies define interpretability as a measure of model characteristics (e.g., complexity) learned by the algorithms. The problem of such an operational definition is its
disconnection with actual user interpretation [46, 47].
To address both challenges simultaneously, most recently policy learning methods such as reinforcement learning are being studied to lend support to the
generation of easy-to-interpret adaptive strategies according to the observations of
target users [48, 49].
Another great domain wherein smart learning for mobile and IOT e-health can
make great impacts is to provide systematic support for operationalizing patient
education and coaching programs in daily-life settings. The prevalence of low health
literacy and self-efficacy in the population is one of the central problems in the
healthcare system. Due to the service-oriented nature of strict time scheduling,
physicians often cannot contribute significantly to patient education. A possible
solution is the use of automated educational systems, which adapt the delivery
of knowledge according to a learner’s needs and level of knowledge. Although
adaptive automated educational systems for other topics have been applied in
practice, the adaptive automated patient education system practically does not exist.
Several challenges have emerged therein. The key challenge lies on its long
development phase and non-standardized evaluation. Additionally, specific limitations in the domain of healthcare lie on the needed amount of personal information
about the user and the number of related time-consuming assessments on the
patient’s side. Furthermore, there exist challenges on the bandwidth of healthcare
experts needed for implementing such systems and the difficulties in explicating
their knowledge.