Behavior change is a complex process in which people receive support in order to improve aspects of their behavior, for instance regarding their health or lifestyle. Although there exist several theoretical approaches to model behavior change, including abstractions that can be applied to di erent use-cases, these are not easily translated into reusable components that can be integrated into implementable systems for persuasion. This work discusses the need for an ontology-based approach to modelling interactions in eHealth systems, with the goal of achieving behavior change. This contribution includes an analysis of current modelling needs in behavior change, specially regarding: stages of change, motivation & ability factors, plans & actions, argumentation, and domain modeling.
Improving quality of life and reducing health risks are increasingly important
concerns in our society. In di erent situations, including the presence of chronic
diseases, or the desire to adopt healthier daily habits (e.g. concerning diet,
physical activity, etc.), these improvements are only possible if an e ective behaviour
change is also produced. This change can span from small alterations to daily
routine to radical changes in lifestyle. It has been shown that personalized
interventions are crucial in order to maximize the e cacy of behaviour change.
Custom and tailored programs are nowadays feasible, thanks to advances in
personal data analytics and personalized digital health. Di erent models exist to
describe behaviour change strategies [
However, the e ort required in order to adapt these models to the appropriate technologies in a given use-case, remains prohibitive and leads to ine ective or partial implementations, with little or incomplete personalization. As a consequence, there is no clear methodology that allows to e ectively model the pro le of a patient, with the goal of using arti cial intelligence (AI) techniques to adapt and personalize treatments, recommendations, and other health-related interventions. Therefore, even if di erent digital solutions and AI techniques have been shown to provide signi cant improvement to personalized treatments, it remains challenging to reuse and apply these methods to other use cases, or to establish a well-de ned work ow for enabling tailored behavior change.
In this paper, we envision an ontology-based approach that establishes in a systematic way the di erent elements that can guide the implementation of personalized behavior change programs, using ontological models as a foundation layer. In concrete, we address: (i) the modeling of behavior change models themselves, i.e. the di erent states of a user, and the factors and barriers that have an e ect on their actions; (ii) the modeling of arguments that can be used to persuade or to in uence the user; (iii) the modeling of the interactions with the user, following agent-based paradigms for autonomous behavior, and negotiation.
Throughout the paper, we use smoking cessation as a running example for the analysis of semantic modelling needs for behavior change systems. 2
Behavior change is a challenging problem, especially regarding health-related
issues and lifestyle. There are di erent factors that need to be taken into
account in order to achieve e ective outcomes, including the attitude, emotional
issues, social pressure, self-perception, etc. [
As it can be seen in classic behavior change models, participants may fall
under di erent states, for which intervention may require complete di erent
strategies and approaches. Although di erent such models exist (e.g. Trans-theoretical
model of stages of change [
As an example, for our smoking cessation use-case, participants may enter a contemplation state when they consider registering to the program, after realizing the bene ts of quitting, while considering their own capacity to overcome addiction. Afterwards, once they enter the program, they may enter a preparation state in which they become aware of their consumption behavior thanks to the tracking of cigarettes through a chatbot. This step is crucial as they consciously learn about their own smoker pro les, which helps detecting which cigarettes are easily avoidable, what behaviors can be modi ed, and what effort is required to attain the desired cessation goals. Furthermore, during the cessation itself (trial/maintenance), the participants may be in di erent states depending on their chosen strategy: e.g., dealing with relapse and/or replacing smoking with alternative activities.
Using an ontology to represent these states within a behavior change model can help reusing concepts and relationships shared among them. For instance, the di erent intention states (e.g. precontemplation, contemplation, preparation) can be reused among di erent models, as well as their relationship with behavioral states. Moreover, information and awareness factors could feed from existing ontologies that already provide existing knowledge over a certain subject (e.g. ontologies for smoking cessation and prevention describing risks, facts, evidence, etc.). In sum, the modeling of behavior change models would require: R1.1: Stage modeling: Generic states found in most models provide a high-level view over a participant's state with respect to a desired behavior change program/initiative. From these generic stages, more speci c one can be derived depending on the use case.
R1.2: Modeling of Factors & Barriers: The di erent factors that my in uence the change of state, as well as the development of a desired behavior change are numerous, and can be classi ed in di erent ways according to existing models. Motivation, awareness, and ability factors are common examples.
R1.3: Composition: The combination of di erent states and factors is desirable, considering that di erent models can establish alternative interactions patterns among them.
R1.4: Triggers: Moving form one stage to the other may happen in di erent manners, and often requires to be de ned in terms of certain triggers, events or other type of signaling elements, which should be included in an ontological model. 3
Having de ned a general behavior change model, the next step is to describe the plans and actions related to a certain stage of change. These plans refer to speci c activities performed during a program or a treatment, which tend to be highly domain speci c.
Ability factors in a behavior change model may include action plans, which can di er depending on the goals. As an example, in smoking cessation these action plans may include tracking cigarette consumption, preparing oneself for the cessation period (e.g., planning distracting actions), or reducing unnecessary cigarettes. Nevertheless, we can identify the following key elements: R2.1: Goals: Participants of a health program may de ne di erent goals which may also depend on the stage they are in. For instance, on smoking cessation, the general goal is to quit smoking entirely. However, there may be intermediary phases, for instance during preparation or trail, for which the goal might be to reduce the cigarette consumption, or to at least identify those cigarettes that can be replaced by other activities. Goals may also be linked to constraints (e.g., duration) so that they can be evaluated. R2.2: Planning: Following the goals, a set of activities can be de ned according to a plan. For example, this may translate to monitoring context of consumed cigarettes (reporting mood, need level, circumstances, etc.), adoption of alternative activities replacing smoking (e.g., sports), or changing certain daily habits.
R2.3: Feedback: During a behavior change program, it is crucial to periodically assess the situation, in order to check if current measures are e ective, or if amendments must be made to a plan. For instance, in case of relapse, it may be needed to understand the circumstances of failure. Or conversely, in case of positive results, how they can be maintained.
R2.4: Personalization: Plans and actions need to be adapted to the speci c conditions and context of each participant, to maximize e cacy. People do not respond in the same way to a treatment or a program, for example in smoking cessation a participant may struggle more with social smoking, while others may have troubles dealing with stress. A participant behavioral pro le should be modeled, in our case, through ontologies in order to capture these speci cities and patterns. 4
As we mentioned before, domain-speci c semantic models are important, in order to re ect accurately the di erent aspects relative to a health program. These ontologies may cover a number of di erent aspects, some of which we described throughout this paper. In this section we focus on those aspects that are generally not extensible to other domains.
Domain speci c vocabulary requirements can be summarized as: R5.1 Pro le data: Data models must include domain speci c information related to the user pro le. This may include data intrinsic to the user (e.g. information of the participant history, behavior patterns, self assessment before, during and after the program, etc.) R5.2 Health issue data: Ontologies may describe a pathology, a health issue/problem, including possible complications, relationship with co-morbidities, diagnosis, etc. R5.3 Treatments/Medication: In some cases certain procedures, treatments and medications may be part of a behavior change program. In those cases, existing standards related to these information elements must be incorporated.
R5.4 Messages and Communication data: Information about a health problem, motivation and encouragement messages, among others, are fundamental in order to engage with participants, via di erent communication means (e.g. chatbot, facebook, email, etc.) In certain cases these materials are well known and can be reused or adapted to a degree. 5
As we have seen, behavior change applications can bene t form existing theories, represented as behavior and intention states, in uenced by di erent types of factors and barriers. Using ontology-based models as explained above, can provide a solid foundation for developing behavior change applications, considering not only the personal circumstances of a participant, but also the speci cities of the health issue that is being addressed.
In summary, the success of these behavior change applications will require a combination of these models, and their reuse by agent systems that include them as part of their knowledge/beliefs/goals. The challenges and future work include: (i) The design of vocabularies and ontologies for description, discovery and exchange within behavior change agents; (ii) The development of speci c domain models that can help enriching agent-based systems in areas such as physical rehabilitation, medication adherence, physical activity, sleep training, etc.; (iii) Agent coordination and negotiation to incorporate computational persuasion into the agent execution logic; (iv) The speci cation of cooperation protocols for participating agents, nding common problems and targeting community-based interventions; (v) The implementation of the proposed model, and evaluation on a real environment with a considerable number of participants; (vi) Ensuring privacy protection, using di erent approaches spanning from obfuscation to anonymity guarantees.