Towards more personalized digital health interventions: a clustering method of action and coping plans to promote physical activity

Promoting a healthy lifestyle is key in reducing the burden of non-communicable diseases such as type 2 diabetes, cancer, osteoarthritis, depression and cardiovascular diseases [1, 2]. Digital health interventions, an umbrella term for the usage of digital technology to support health [3], can be employed to promote a healthy lifestyle and has gained popularity because of its time- and cost-effectiveness [4‐6].

Digital health interventions have been found to be more effective when informed by a behaviour change theory in comparison with a-theoretical interventions [7]. Several theories have been developed (e.g. social cognitive theory, the health belief model, self-determination theory), but one of the most comprehensive models is the Health Action Process Approach (HAPA) model [8]. The HAPA-model is a two-phase model that guides individuals to change their behaviour, beginning with the development of an intention (motivational phase), followed by bridging the gap between intention and the actual behaviour (volitional phase) [8, 9]. Behaviour change techniques (BCTs) such as action planning (i.e. where participants select their own goals and decide what they want to do & how, where and when they want to do it) and coping planning (i.e. exploring solutions for possible barriers) are key components within the HAPA-model to bridge the intention-behaviour gap [10‐13].

Despite the effectiveness of action and coping plans in digital health interventions to change behaviour [14‐16], attrition rates remain high [14, 17, 18]. This reduces the impact of these interventions. An important reason for these high attrition rates might be that support offered by digital interventions to make such action and coping plans usually is abstract, generic and the same for each individual (a so-called “one size fits all intervention”) [14]. Nevertheless, people are different, and the context in which they behave varies between individuals [19]. There is a need to provide support in digital health interventions in a more personalized and contextualized way.

As yet, interventions provide already tailoring at the construct level or BCT level (e.g. participants who have an intention to change for example receive other BCTs than participants who do not have an intention to change) [20]. However, practical support on the content level (i.e. concrete operationalizations of BCTs such as action and coping planning), is not provided in a personalized and contextualized way. In previous studies [14, 16, 17], participants were considered as their own expert in terms of making plans: they specify themselves the content of their plans and take their own personal and context-factors into account. Nevertheless, this approach resulted in a low quality of plans, and participants experienced difficulties in formulating them [21, 22]. As a result, support at the content level is needed: this support should include suggestions of specific plans that are personalized to the individual (e.g. If someone is retired, he should not get the advice to walk to work) and contextualized to the individual (e.g. If someone is working from home, she should not get the advice to go for a lunch walk with a colleague).

A promising approach is to use intelligent algorithms and decision support systems [23]. The term “decision support system” (=DSs) is a broad concept covering all aspects of support during decision making, and provides automated recommendations where required and when available [23]. As such, a DSs could improve tailoring in digital health interventions by suggesting a relevant plan to do physical activity (PA) that is personalized and contextualized to the individual. Notwithstanding the potential of a DSs, a knowledge-base should first be developed in order to deliver such suggestions of plans [23, 24]. A knowledge-base is defined as a collection of facts, assertions, relationships, rules about a specific domain represented in a computer readable format [24]. The process of acquiring knowledge for the knowledge-base is defined as knowledge acquisition and may come from multiple sources (experts, books, research findings, etc) [24]. For our purpose, the knowledge-base should at least contain relationships between personal and contextual user information (i.e. information that relates to the individual itself such as demographic information, motivational stage, emotions; information that relates to the context of the individual such as physical and social environment, the weather, respectively) and PA plan characteristics (e.g. PA type, place of the activity, time of the activity, barriers to do the activity). Once a knowledge-base is developed, it may become possible to deliver context-aware personalized suggestions to a user, as the DSs can exploit the knowledge-base to find appropriate suggestions based on specific user information of that user (e.g. if a user is a female younger adult living with her partner in a rural environment and the knowledge-base contains relationships between these characteristics and certain outdoor physical activities, these activities can then be delivered as a suggestion to the user).

The main objective of this study is to empirically investigate whether user information relates toward specific action and coping plans (e.g. motivated users may rather plan vigorous physical activities compared to less motivated users; adults who work may rather experience barriers such as not having time for PA compared to retired adults). More specifically, this paper will rather address personal user information than contextual user information. As such, this paper provides a proof-of-concept on how knowledge can be acquired in order to develop such a knowledge-base. This includes a clustering method with a two-steps approach. The first step is to explore whether patterns in action and coping plans can be identified using clustering algorithms on available data. The second step is to examine whether these clusters of action and coping plans can be linked to specific user information.

The study explored the feasibility of applying a clustering method to develop a knowledge-base, which might be a first step towards more personalized suggestions on the content level in future digital health interventions. More specifically, this study investigated whether user information was related to specific action and coping plans. The results can be readily summarized. First, we were able to cluster action plans, coping plans and the combination of action and coping plans. Second, relating these clusters to user information was possible for action plans, but proved more difficult for the coping plans and specific combination of action and coping plans.

Our study revealed that some user characteristics related toward specific action plans. 1) Users with a higher BMI were more likely to choose outdoor leisure activities (walking, biking, running). 2) Women, users that did not perform PA regularly yet, or users who had a job, were more likely to choose for household activities. 3) Younger users were more likely to choose for active transport and different sports activities (fitness, swimming, tennis). Of these younger adults, users with a higher BMI were more likely to choose for active transport whereas users with a lower BMI would choose for different sports or work-related activities. Overall, these findings suggest that with the approach used in this study, it is feasible to find relations between action plans and specific personal user information. Consequently, the knowledge acquired from these findings might be used to define relationships in a knowledge-base and to ultimately personalize suggestions for action plans.

Although we could identify relatively consistent pairs of barriers and solutions formulated in the clusters of coping plans, we concluded that no logical link was found between user information and coping plans or specific combinations of action and coping plan. This was concluded since these results could not be compared to previous studies, nor theory. The reason that no logical link was found might be due to the fact that 8 clusters were identified for both coping plans and for specific combinations of action and coping plans. This is a relatively large number of clusters to relate user information to and may explain why, although statistical differences were found, these differences were not straightforward to interpret. Moreover, clusters of the specific combinations of action plans and coping plans were not considered to be valuable because it was difficult to meaningfully distinguish one cluster from another. Advanced clustering techniques with a larger and more heterogeneous sample may identify more valuable clusters in future research [30]. Also, the reason for the large numbers of clusters might reflect the fact that it is not possible to cluster coping plans or specific combinations of action and coping plans. In that case, only suggestions of action plans could be formulated rather than suggestions of coping plans or combinations of both. Another, maybe more important explanation, might be that the current paper only analyzed personal user information (i.e. demographic information and motivational stage). Consequently, relationships of other user information with plan characteristics remain unexplored. As it is known that PA is not a stable but dynamic (i.e. time-dependent) behaviour that varies throughout the day and from day to day [31, 32], it is more likely that barriers to certain physical activities, that are in a sense more hypothetical than action plans, relate more to contextual (e.g. the weather) and dynamic user information (e.g. emotions) than personal and rather stable user information (e.g. demographic information). Thus, future research should investigate whether more contextual and dynamic user information relates more strongly towards plan characteristics.

Considering the above findings, the clustering method used in the current study might be a feasible approach to acquire knowledge for a knowledge-base. However more user information will be needed to deliver personalized suggestions on the content level. First, we will illustrate how the results of the current study might be used to develop a knowledge-base, and then we will discuss what other user information might be needed to deliver more context-aware personalized suggestions. To develop a knowledge-base, acquired knowledge should be organized into a structure. Ontologies are one of the most popular approaches to structure these knowledge-bases as they are well specified [33], and can be combined with intelligent algorithms, which makes it possible to deliver personalized suggestions (e.g. for example, a user who does not perform PA regularly yet may get a suggestion to do a household activity). Furthermore, Larsen [33] shows that ontologies are already increasingly used by behavioural scientists. Indeed, several ontologies in the PA and behaviour change domain already exist [33]. For example, the Physical Activity Concept Ontology (PACO) structures different physical activities [34], the HAPA ontology structures all constructs of the HAPA-model [35], the Behaviour Change Interventions Ontology (BCIO) is a broader ontology that structures knowledge about interventions, their contexts, effects and evaluations [36, 37]. Even though ontologies such as the BCIO and HAPA ontology provide structures and their relationships on an abstract construct level (e.g. ‘intention’ influences ‘planning’, ‘planning’ positively influences the ‘Intenders’ to ‘Actors’ transitions relationship [35]), they still lack detailed concretization of these constructs and their interrelations at the content level [33]. Yet, ontologies with detailed concretizations at the content level are needed to deliver context-aware personalized suggestions. The current paper provides an approach to acquire knowledge for such an ontology. Here follows an example of how an ontology with the findings from the current paper may be used to deliver personalized suggestions. Suppose Rosy (older woman, higher BMI, high education, pre-intender for PA) logs into an m-health intervention. Which suggestions could the system deliver to get her more active the following week? Since the ontology contains relationships between user information and plan characteristics, the system can exploit the ontology with the help of intelligent algorithms and deduce which specific plan suggestions would match Rosy’s user profile. Based on the findings of the current paper, the suggestions ‘do a household activity’ or ‘do an outdoor activity such as walking, biking or running’ could be delivered to Rosy. In addition, Larsen [33] highlights the importance of combining ontologies with other ontologies in the field and asks the scientific community to update ontologies as new evidence emerges. As such, findings from the current paper could take other ontologies (such as the BCIO or the HAPA ontology) to a higher level by adding detailed concretizations at the content level. Consequently, the approach used in this study may contribute to the refinement of ontologies related to behaviour change interventions.

As previously stated, more user information in relation to plan characteristics should be acquired in order to develop a knowledge-base for context-aware personalized suggestions. Future studies might use the same approach of the current study to acquire more data from possible end users to shape the knowledge-base. However, the current approach should be enriched with other user information that might be placed on two continuums: First, relatively stable user information (i.e. information that does not change over a certain period of time) versus more dynamic user information (i.e. information that varies over a certain period of time). Second, personal user information (i.e. information related to the individual itself) versus contextual user information (i.e. information related to the context of the individual). The current paper addressed demographic info and motivational stage as stable and personal user information. Other personal and relatively stable user information worth investigating might be (perceived) motor skill competence or physical health. For example, recent research of Drenowatz [38] showed that a higher motor skill competence could be linked to more club sports participation. However these findings were only identified in children [38]. Another example shows that patients with chronic back pain perform more physical activities in the morning than in the evening compared to controls [39]. Second, exploring whether relatively stable and contextual user information relates to plan characteristics might provide important information as well. It might be interesting to explore whether users from various home or work environments make different plans or encounter different barriers to do PA. For example, research already showed that neighborhoods supporting a safe, enjoyable and social experience are associated with more leisure time walking among adults [40]. Third, as PA is not a static but dynamic behaviour [31, 32], it would be useful to examine whether certain dynamic personal user information (e.g. emotions, fatigue, pain) and dynamic contextual user information (e.g. weather, agenda of the day) relates to certain plan characteristics. This would enable the knowledge-base to deliver personalized suggestions based not only on relatively ‘stable information’ but also on ‘dynamic information’. For example, when the user has a busy day at work (dynamic contextual info), he might need other suggestions than when a user has more time; or when a user is stressed (dynamic personal info), the user might need other suggestions than when a user is more relaxed that day.

Some considerations should be taken into account for future research. To acquire more dynamic information, future studies should collect data on a smaller timeframe (day to day or even within days) as compared to the current study (only once at the start of the study). Relatively stable information might still be collected at the start of the study and/or at another moment depending on the study length (e.g. motivational stage may change after 3 weeks in an intervention). To acquire more dynamic information, ecological momentary assessment (EMA) might be used because this method makes it possible to collect real-time data based on repeated measures and observations that take place in participants’ daily environment [41]. For instance, during a 7-day EMA study the emotional state of a participant might be asked for example every 3 to 4 h, together with the question to make a PA plan for these following hours. This example would make it possible to relate certain emotions toward specific plans.

If a knowledge-base were to be developed based on the approach of the current study and the above-mentioned suggestions, future interventions to promote PA can exploit the knowledge-base in order to deliver context-aware personalized suggestions. Notwithstanding, these future interventions might take further steps toward context-aware personalized suggestions. First, despite the fact that more information in the knowledge-base may result in more context-aware personalized suggestions, one should be careful with asking too many questions at the start of a personalized intervention (in order to determine the new user’s profile). Using smart technologies such as wearables and apps (e.g. to measure stress, PA level, location, agenda, weather) could limit the number of questions. Second, the current study highlights the importance of a knowledge-base to deliver context-aware personalized suggestions in e-and m-health interventions. Nonetheless, it is unlikely that this approach will fully capture the complexity of behaviour change to provide context-aware personalized suggestions. Other approaches could complement the approach used in the current study. One of these approaches is ‘reinforcement learning’ [42]. In its most basic level, the system learns by measuring a success criterion for a given suggestion: if the success criterion is met, the probability of suggesting this suggestion a second time increases [42]. For example, the success criterion can be based on the user’s rating for a certain suggestion, or the user’s behaviour after that suggestion (e.g. if a user gets a suggestion of a plan to go for a walk and the user eventually goes for a walk). More advanced derivations of reinforcement learning should be explored in future interventions, for example success criterions of similar users [42, 43]. Third, another approach that might complement the current approach is the systems ID approach. The approach used in this study is still a ‘nomothetic’ approach (i.e. making ‘aggregated’ conclusions of relationships of user info and plan characteristics), whereas ‘ideographic approaches’ might deliver more context-aware personalized suggestions (i.e. making individualized conclusions of relationships between user info and plan characteristics by examining within-person variation over time). The systems ID approach is an ‘ideographic approach’ and learns from run-in periods to provide personalized suggestions (e.g. for a certain user it might be better to suggest a walking activity on a weekend day, whereas for another user it might be better to suggest a walking activity on a weekday) [31, 42]. The disadvantage of such a run-in period is that no context-aware personalized suggestions can be delivered at the beginning of such an intervention (which is also the case when using reinforcement learning on its own).

Until now, attrition rates in digital health interventions to promote PA are high which might be due to the lack of context-aware personalized suggestions in these interventions. The approach used in the current study might be a feasible approach to acquire knowledge for a knowledge-base, however more data from possible end users should be acquired in future research. This might result in a first type of context-aware personalized suggestions on the content level, as the system can provide initial suggestions based on the knowledge-base when a PA intervention has just started and the system did not have time to learn about the user. Over time, suggestions could be refined based on other approaches like reinforcement learning. Moreover, this approach extended prior efforts to personalize digital health interventions by providing context-aware personalized suggestions on the content level rather than on the construct level.