6.3 Model development
6.4 User evaluation of the recommendations
6.5 Conclusions
7
Guided home-based resistance exercise 1137.1 Introduction
7.2 Methods
7.3 Results
7.4 Discussion
7.5 Conclusions
Lifestyle interventions to
improve blood pressure
6. Rasch-based lifestyle recommendations
6.1 Introduction and Context
Hypertension, or high blood pressure, is a key risk factor for cardiovascular diseases [243]. Ef-fective hypertension management is therefore a major concern for public health. Besides blood pressure medication, non- pharmacological lifestyle interventions have been proven successful in hypertension management [57]. Appropriate lifestyle modifications may not only lower or control blood pressure in hypertensive patients but also effectively delay or prevent hypertension in non-hypertensives [145]. The ESH and ESC endorse a wide variety of lifestyle interventions for the reduction of BP: salt restriction, moderation of alcohol consumption, a diet rich in vegetables, fruits and low-fat dairy products, weight reduction, regular exercise and smoking cessation [145].
6.1.1 Recommender systems
Despite the obvious health benefits, adherence to lifestyle recommendations in the hypertensive population is generally low [231]. Changing behavior requires effort from the user and support from the environment. The American Heart Association stated the potential of mobile health applications to provide the information and support that is necessary to counsel and motivate individuals attempting to improve their lifestyle [33]. The first of these digital coaching services are recommender systems which rely on the pervasiveness of mobile media to support the patient at times where the care provider is unable to do so. These systems are outperforming traditional means of lifestyle recommendation for heart failure patients [41]. The content of these systems has focused on advice with a high expected health benefit by tailoring the recommendations based on the health needs of the user (e.g. weight loss recommendations for obese patients).
6.1.2 Feasibility
Besides the health benefit other factors also contribute to adherence. The classical Health-Belief Model [311, 313] coins (amongst other factors) the barriers for the patient to follow the recommen-dation as an additional predictor of adherence. This notion reoccurs as a prominent determinant
of behavior in many other psychosocial theories where it is called the behavior difficulty, cost or feasibility[13]. In this paper we will call this factor feasibility, though the other terms will also be used interchangeably.
Recent work [175] has incorporated a notion of feasibility. For each monitored behavior, feasibility was measured as the frequency with which the user performs it. The authors relate the resulting feasibility parameters to self-efficacy (a person’s belief in their own ability [11]).
Subsequently, the feasibility and the benefit of the behavior were weighed against each other to find recommendations that are both feasible and beneficial. The resulting recommender system outperformed random control by most measures, including adherence to the recommendation [176].
This approach makes excellent use of the user’s data, yet also has its limitations. The method measures seperate self-efficacy parameters for each behavior, while usually self-efficacy is seen as a trait of the person that can influence a large amount of behaviors simultaneously, leading their feasibility to be correlated within an individual.
Modeling the feasibility in health behaviors as a combination of the behavior’s inherent diffi-culty level and the individual’s personal level of self-efficacy allows knowledge about a behavior’s difficulty to be generalized to different individuals. This has two benefits: first of all, knowing the relative difficulty of behaviors amongst each other can be used as a priori information to start an intervention with behaviors that are more feasible; the other benefit is that after observing the engagement of the user in a small subset of the behaviors, knowledge about how difficult those behaviors are can predict the self-efficacy level of the user, which then can in turn be used to predict how likely it would be for the user to engage in the remainder of the behaviors that were not observed (given the corresponding difficulty levels). This interaction between a pre-determined level of behavior difficulty with a personal level of self-efficacy can be used to remedy some common problems in recommender systems, such as dealing with the cold-start problem (i.e. system needs too much data from the user before it can generate personalized content) or being able to reason about behavior that is hard to monitor or observe with the technology at hand.
6.1.3 The Rasch model as a feasibility model
We have selected the Rasch model [295] as a simple, one-dimensional user model for the modeling of the feasibility of behaviors in the form of a generic ability level with the benefits outlined above.
The Rasch model is based on item-response theory (IRT), which deals with modeling a latent trait (i.e. a person’s ability) given the difficulty of a set of tasks that relate to that latent trait (i.e. behavior difficulty). The classical use of the Rasch model is in psychometrics (e.g. mathematical capability testing) but it has recently been applied successfully on lifestyle behaviors such as dietary habits [90], salt intake restriction [150], mindfulness habits [193] and even across multiple dimensions of health behavior [36] as a measure of health performance. The success of the Rasch model in describing engagement and difficulty of behaviors as well as the performance of individuals in these behaviors across a multitude of health-related categories motivates the use of the Rasch model as a user model of feasibility, capturing both the notion of a behavior-specific difficulty and a person-specific ability level.
Statistically, the Rasch model assumes a unidimensional latent variable (i.e. a user’s ability of, attitude towards or self-efficacy in living a healthy lifestyle) given observations of the patient’s engagement in manifest behaviors (of which the difficulty is assumed to have a transitive ordering).
This information is captured on a single scale (the Rasch scale) which corresponds to a continuous latent trait (i.e. self-efficacy or ability). Behaviors are then modeled as probabilistic functions on
6.1 Introduction and Context 101
Figure 6.1: Item-characteristic curves (ICC) from the Rasch model that represent the probability of engagement for three different recommendations (related to weight loss, salt restriction and exercise) as a function of the patient’s ability.
this latent trait. Such a function is referred to as the item-characteristics curve (ICC), a function that describes the probability of engagement of a user in the behavior given their ability score on the Rasch scale. An ICC for a behavior n with difficulty βnreturns P(Xni= 1) (i.e. the probability of a person i engaging in n) given that person’s ability δi:
P(Xni= 1) = eβn−δi
1 + eβn−δi (6.1)
This means that as the user’s ability level δi increases, the probability of them engaging in some behavior also increases logistically. Vice versa, as the difficulty of a behavior increases, the probability of some user engaging in it decreases. The β of a behavior is proportional to the number of people in a population who perform it, while the δ of an individual is proportional to the number of the behaviors he or she performs. Figure 6.1 illustrates the concepts of behavior difficulty, patient ability and the item-characteristic curves visually with examples of 3 behaviors related to hypertension management. The x-axis represents the latent trait (i.e. the ability of the user to live a healthy lifestyle). Each ICC represents the probability of engagement in a certain behavior given the ability. They show that some behaviors are easier than others (e.g. behavior 1 always higher probability than behavior 2), reflecting the idea of a transitive ordering of the behaviors in terms of their difficulty. In the remainder of this paper the the difficulty of a behavior will refer to the mean of its ICC (i.e. 50% probability point). The actual likelihood of engagement also depends on the ability. For example, person X with an ability of 1.9 has a 50% likelihood to ask for low sodium (behavior 2), is more likely to compare calories (behavior 1, 75%) but less likely to use an activity tracker (behavior 3, 25%). Person Y, that has a higher ability (2.9) is more likely to do all these behaviors.
6.1.4 Feasibility-based recommendation strategies
In this work we use the Rasch model (fitted to a target user populations’ behavioral patterns) to obtain two measures. The first is the difficulty of a behavior (which defines its ICC). This metric is independent of the user but does provide prior knowledge about the difficulty ordering of lifestyle interventions. The second parameter is the ability of the user (which can be estimated from the engagement patterns of the user in a subset of the behaviors). This metric can be used to estimate the
probability of engagement (see equation 6.1) in each of the behaviors for that specific user. Based on these measures two distinct advice generation strategies were motivated from a perspective of behavioral change theory. The engagement maximization strategy uses only the behavior difficulty measure while the motivation maximization strategy also uses the user’s ability. The following sections describe these strategies.
Engagement maximization strategy (engagement maximisation strategy (EMS))
For any patient with an arbitrary ability level, the Rasch model predicts a higher probability of engagement in behaviors that are relatively easy when compared to more difficult behaviors.
According to the Campbell paradigm [114], such a high probability of engagement implies that the benefit of performing these behaviors is outweighed by their costs. Thus, maximizing the probability of engagement is equivalent to minimizing the cost-to-benefit rate of the behavior. The health-belief model also poses a positive contribution of the behaviors’ cost-to-benefit rate towards the individuals’ engagement in the behavior. The engagement maximization strategy (EMS) will therefore always recommend the most feasible behaviors from the Rasch scale (from the subset of behaviors that the patient is not already engaged in). This notion is supported by research on decision-support systems for the energy savings domain where it was shown that people ranked more feasible measures as more preferable [206].
Motivation maximization strategy (motivation maximisation strategy (MMS))
While advising the most feasible behaviors can already be a strong enhancement for recommender systems, it could also be argued that for persons with a higher ability level such advice would not be appropriate for a number of reasons. First of all, the mismatch between ability and the limited challenge of easy recommendations could lead to sub-optimal motivation [299]. In addition, persons who perform most easy behaviors might have other reasons for not engaging in the small set of remaining easy behaviors (e.g., physical disability to do certain exercises or allergies that prevent consumption of certain foods). These persons will not desire advice about these behaviors. Finally, theories on cognitive development suggest that the experience of mastering a difficult task could be rewarding by itself and contribute to self-efficacy [12]. These arguments suggest that the patient’s ability level (as derived from the Rasch model) should be taken into account for personalization.
Therefore the MMS will select behaviors for which difficulty is close to the patient’s ability on the Rasch scale.