Virtual coaches have the potential to address the low adherence common to eHealth applications for behavior change by, for example, providing motivational support. However, given the multitude of factors affecting users’ attitudes toward virtual coaches, more insights are needed on how such virtual coaches can be designed to affect these attitudes in a specific use context positively. Especially valuable are insights that are based on users interacting with such a virtual coach for longer. We thus conducted a study in which more than 500 smokers interacted with the text-based virtual coach Sam in five sessions. In each session, Sam assigned smokers a new preparatory activity for quitting smoking and provided motivational support for doing the activity. Based on a mixed-methods analysis of users’ willingness to continue working and their relationship with Sam, we obtained eight themes for users’ attitudes toward Sam. These themes relate to whether Sam is seen as human or artificial, specific characteristics of Sam (e.g., caring character), the interaction with Sam, and the relationship with Sam. We used these themes to formulate literature-based recommendations to guide designers of virtual coaches for behavior change. For example, letting the virtual coach get to know users and disclose more information about itself may improve its relationship with users.@en

Background: Smoking and physical inactivity are two key preventable risk factors of cardiovascular disease. Yet, as with most health behaviors, they are difficult to change. In the interdisciplinary Perfect Fit project, scientists from different fields join forces to develop an evidence-based virtual coach that supports smokers in quitting smoking and increasing their physical activity. Intervention content, design and implementation as well as lessons learnt are presented in the hopes of guiding other research groups working on similar projects. Methods: Numerous approaches were used and combined to support the development of the Perfect Fit virtual coach. Approaches include literature reviews, empirical studies, collaboration with end-users, content and technical development sprints, interdisciplinary collaboration and iterative proof-of-concept implementation. Findings: The Perfect Fit intervention integrates evidence-based behavioral change techniques as well as new techniques focused on identity change, big data science, sensor technology and personalized real-time coaching. Intervention content of the virtual coaching matches communication preferences and individual needs of end users. Lessons learnt include ways to optimally implement and tailor interactions from the virtual coach (e.g., ‘explain why user is asked for input’, ‘tailor timing and frequency of intervention components’). With regards to the development process, lessons learnt include strategies for effective interdisciplinary collaboration and technical development (e.g., ‘Find a good balance between wishes of end-users and legal possibilities’). Discussion: The Perfect Fit development process was interactive, iterative and challenging at times. We hope that our experiences and lessons learnt can inspire and benefit others.@en

Behavior change applications often assign their users activities such as tracking the number of smoked cigarettes or planning a running route. To help a user complete these activities, an application can persuade them in many ways. For example, it may help the user create a plan or mention the experience of peers. Intuitively, the application should thereby pick the message that is most likely to be motivating. In the simplest case, this could be the message that has been most effective in the past. However, one could consider several other elements in an algorithm to choose a message. Possible elements include the user’s current state (e.g., self-efficacy), the user’s future state after reading a message, and the user’s similarity to the users on which data has been gathered. To test the added value of subsequently incorporating these elements into an algorithm that selects persuasive messages, we conducted an experiment in which more than 500 people in four conditions interacted with a text-based virtual coach. The experiment consisted of five sessions, in each of which participants were suggested a preparatory activity for quitting smoking or increasing physical activity together with a persuasive message. Our findings suggest that adding more elements to the algorithm is effective, especially in later sessions and for people who thought the activities were useful. Moreover, while we found some support for transferring knowledge between the two activity types, there was rather low agreement between the optimal policies computed separately for the two activity types. This suggests limited policy generalizability between activities for quitting smoking and those for increasing physical activity. We see our results as supporting the idea of constructing more complex persuasion algorithms. Our dataset on 2,366 persuasive messages sent to 671 people is published together with this article for researchers to build on our algorithm.@en

Background Despite their increasing prevalence and potential, eHealth applications for behavior change suffer from a lack of adherence and from dropout. Advances in virtual coach technology provide new opportunities to improve this. However, these applications still do not always offer what people need. We, therefore, need a better understanding of people’s needs and how to address these, based on both actual experiences of users and their reflections on envisioned scenarios. Methods We conducted a longitudinal study in which 671 smokers interacted with a virtual coach in five sessions. The virtual coach assigned them a new preparatory activity for quitting smoking or increasing physical activity in each session. Participants provided feedback on the activity in the next session. After the five sessions, participants were asked to describe barriers and motivators for doing their activities. In addition, they provided their views on videos of scenarios such as receiving motivational messages. To understand users’ needs, we took a mixed-methods approach. This approach triangulated findings from qualitative data, quantitative data, and the literature. Results We identified 14 main themes that describe people’s views of their current and future behaviors concerning an eHealth application. These themes relate to the behaviors themselves, the users, other parties involved in a behavior, and the environment. The most prevalent theme was the perceived usefulness of behaviors, especially whether they were informative, helpful, motivating, or encouraging. The timing and intensity of behaviors also mattered. With regards to the users, their perceived importance of and motivation to change, autonomy, and personal characteristics were major themes. Another important role was played by other parties that may be involved in a behavior, such as general practitioners or virtual coaches. Here, the themes of companionableness, accountability, and nature of the other party (i.e., human vs AI) were relevant. The last set of main themes was related to the environment in which a behavior is performed. Prevalent themes were the availability of sufficient time, the presence of prompts and triggers, support from one’s social environment, and the diversity of other environmental factors. We provide recommendations for addressing each theme. Conclusions The integrated method of experience-based and envisioning-based needs acquisition with a triangulate analysis provided a comprehensive needs classification (empirically and theoretically grounded). We expect that our themes and recommendations for addressing them will be helpful for designing applications for health behavior change that meet people’s needs. Designers should especially focus on the perceived usefulness of application components. To aid future work, we publish our dataset with user characteristics and 5,074 free-text responses from 671 people.@en

Smoking tobacco and physical inactivity are key preventable behavioural risk factors of cardiovascular disease (CVD). Computerised coaching systems can help individuals to modify risky behaviours, thereby preventing CVD. However, most reported eHealth or computerized coaching systems are hard to reuse in slightly different settings. To provide an open-source, reusable computer coaching system, we developed Perfect Fit. The reusability is manifested by building around the open-source text- and voice-based contextual assistant framework Rasa. Rasa provides a simple, standard interface to many popular messaging and voice channels, and custom connectors are easily implemented. A set of algorithms have been developed and connected to Rasa to drive and personalize the conversation flow and the coaching process. Such algorithms make use of data stored in a devoted database. Furthermore, Perfect Fit adheres to best practices and standards in software engineering. The modular design of Perfect Fit will allow researchers to connect the virtual coach to any messaging or voice channel with only modest modification. Perfect Fit is available under open-source license in GitHub and is currently in prototype-phase. Concluding, Perfect Fit will deliver a virtual coach that can easily be adapted and reused in different settings. The coach helps individuals to achieve and maintain abstinence from smoking and sufficient physical activity (PA).@en

A human therapist continuously adapts how they persuade a client to adhere to a behavior change intervention based on theoretical expertise, past experience with the client as well as other clients, and the client's current situation. We aim at incorporating these elements into the persuasive communication of a conversational agent that acts as a virtual coach for smoking cessation and physical activity increase. The focus thereby is on investigating how three coaching elements, goal-setting, data-monitoring and the assignment of activities, can be designed to enhance treatment adherence. A first experiment is currently finished to 1) get user input for the interaction design based on interaction scenarios, 2) gather data for and test a reinforcement learning-approach to persuading people to do small preparatory activities for smoking cessation and increasing physical activity, and 3) gain insights into the acceptance and perceived motivational impact of the virtual coach used to persuade people.@en

Duurzame gedragsverandering is moeilijk, zelfs wanneer het huidige gedrag een groot risico vormt voor de gezondheid. Gedragsverandering wordt gemakkelijker als het nieuwe gedrag past bij hoe een individu zichzelf ziet (identiteit). Virtuele coaching is veelbelovend om gedrags- en identiteitsverandering te ondersteunen, omdat het altijd beschikbaar is in de eigen omgeving en optimaal kan worden gepersonaliseerd.@en

Deep Reinforcement Learning (RL) is a promising technique towards constructing intelligent agents, but it is not always easy to understand the learning process and the factors that impact it. To shed some light on this, we analyze the Latent State Representations (LSRs) that deep RL agents learn, and compare them to what such agents should ideally learn. We propose a crisp definition of ’ideal LSR’ based on a bisimulation metric, which measures how behaviorally similar states are. The ideal LSR is that in which the distance between two states is proportional to this bisimulation metric. Intuitively, forming such an ideal representation is highly favorable due to its compactness and generalization properties. Here we investigate if this type of representation is also desirable in practice. Our experiments suggest that learning representations that are close to this ideal LSR may improve upon generalization to new irrelevant feature values and modified dynamics. Yet, we show empirically that the extent to which such representations are learned depends on both the network capacity and the state encoding, and that with the current techniques the exact ideal LSR is never formed. @en

Recent years have seen a surge of algorithms and architectures for deep Re- inforcement Learning (RL), many of which have shown remarkable success for various problems. Yet, little work has attempted to relate the performance of these algorithms and architectures to what the resulting deep RL agents actu- ally learn, and whether this corresponds to what they should ideally learn. Such a comparison may allow for both an improved understanding of why certain algorithms or network architectures perform better than others and the devel- opment of methods that specically address discrepancies between what is and what should be learned.@en