This is an open-access article distributed under the terms of the Creative Commons Attribution License (https://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work, first published in the Journal of Medical Internet Research, is properly cited. The complete bibliographic information, a link to the original publication on http://www.jmir.org/, as well as this copyright and license information must be included.
Web-based mental health interventions have evolved from innovative prototypes to evidence-based and clinically applied solutions for mental diseases such as depression and anxiety. Open-access, self-guided types of these solutions hold the promise of reaching and treating a large population at a reasonable cost. However, a considerable factor that currently hinders the effectiveness of these self-guided Web-based interventions is the high level of nonadherence. The absence of a human caregiver apparently has a negative effect on user adherence. It is unknown to what extent this human support can be handed over to the technology of the intervention to mitigate this negative effect.
The first objective of this paper was to explore what is known in literature about what support a user needs to stay motivated and engaged in an electronic health (eHealth) intervention that requires repeated use. The second objective was to explore the current potential of embodied conversational agents (ECAs) to provide this support.
This study reviews and interprets the available literature on (1) support within eHealth interventions that require repeated use and (2) the potential of ECAs by means of a scoping review. The rationale for choosing a scoping review is that the subject is broad, diverse, and largely unexplored. Themes for (1) and (2) were proposed based on grounded theory and mapped on each other to find relationships.
The results of the first part of this study suggest the presence of user needs that largely remain implicit and unaddressed. These support needs can be categorized as task-related support and emotion-related support. The results of the second part of this study suggest that ECAs are capable of engaging and motivating users of information technology applications in the domains of learning and behavioral change. Longitudinal studies must be conducted to determine under what circumstances ECAs can create and maintain a productive user relationship. Mapping the user needs on the ECAs’ capabilities suggests that different kinds of ECAs may provide different solutions for improving the adherence levels.
Autonomous ECAs that do not respond to a user’s expressed emotion in real time but take on empathic roles may be sufficient to motivate users to some extent. It is unclear whether those types of ECAs are competent enough and create sufficient believability among users to address the user’s deeper needs for support and empathy. Responsive ECAs may offer a better solution. However, at present, most of these ECAs have difficulties to assess a user’s emotional state in real time during an open dialogue. By conducting future research with relationship theory–based ECAs, the added value of ECAs toward user needs can be better understood.
Meta-analyses have demonstrated that Web-based interventions for mental health have become reasonably successful treatments against common mental health problems such as depression and anxiety [
The higher rates of adherence in human-supported interventions can be explained in favor of therapists who do an effective job in motivating clients during their change process [
The following challenge is how these socioemotional processes could be handled within Web-based health interventions. As suggested by Bickmore [
Altogether, ECAs hold the promise that they can bring in social, emotional, and relational elements to the user interface. It is, however, less clear to what extent ECAs can (1) truly handle
This study was performed by means of structured data collection within the Web of Science and Scopus databases. The scoping review was chosen as research method. A scoping review aims to map the existing literature in a field of interest in terms of the volume, nature, and characteristics of the primary research [
This study is divided into two parts:
Is there a set of generic user support needs that are currently not sufficiently addressed within eHealth interventions requiring repeated use that may result in a lower user experience and therefore lower user adherence?
The Scopus database was searched with a combination of the concepts “support,” “Web-based intervention,” and “review.” For each of the concepts, multiple keywords were used (see
The search resulted in 93 studies. On the basis of our inclusion and exclusion criteria, we selected 18 studies. By checking the references of these selected studies, we found another 4 relevant papers. Finally, 22 papers were included. See
Inclusion criteria were as follows:
Papers had to address a Web-based intervention for a mental or physical disorder in which support was the subject of the study
Papers had to review multiple interventions/studies or present ideas based on literature or an earlier study
Exclusion criteria were as follows:
Papers that restricted themselves to a specific disease and/or intervention and did not generalize to eHealth within a broader context
Papers that described the creation of a Web-based intervention and did not take the empirical evaluation in scope
Papers on social media and support solutions that were studied separate from the Web-based intervention events
Papers that did not describe support in functional terms (eg, praise, reassurance) but only in technical delivery terms (eg, short message service [SMS], email)
Papers that analyzed Web-based interventions using high-level descriptive factors (eg, “interactive component,” “supervision,” “tailored”) without going into more detail
Flow diagram of the study selection of part 1 of the scoping review.
The entire content, including the introduction, discussion, and references, of the 22 studies was checked from the users’ perspectives regarding usability and the needs they expressed. We applied grounded theory, applying the following phases:
What are main supportive features of ECAs that could potentially address user support needs?
The search aimed to create a generic idea of the capabilities of ECAs for supportive purposes. The Scopus and Web of Science databases were searched with a combination of the concepts “embodied conversational agents,” “Web-based intervention,” and “support.” For each of the concepts, multiple keywords were used (see
The systematic search resulted in a limited number (8) of studies. Moreover, these studies addressed a wide range of topics, from physical attributes [
We initiated the hand search on the following basis:
Finding synthesizing information on ECAs within a health or pedagogical (ie, e-learning) context with a focus on the delivery of support and motivating users. We started with the information found in [
Finding additional (founding) studies on the computers as social actors (CASA) effect as mentioned within [
Finding additional information on relationship building [
Finding additional information on theoretical models related to ECAs as touched upon in [
The entire search procedure resulted in including 53 studies (
Using grounded theory, the entire content, including the introduction, discussion, and references, of the selected studies was analyzed with the aim of finding specific information on user support as carried out by ECAs. As this information was scarce, we decided to formulate three concepts that we thought were most relevant for eHealth and covered substantial information of the ECA literature that was semantically related to the notion of user support.
We formulated the following three concepts:
Which
What kind of
How
Out of the three concepts we formulated, 8 themes coherently described a specific ECA topic.
Inclusion criterion was as follows:
Papers had to address embodied conversational agents (ECAs) interacting with users or studies on ECAs interacting with users
Exclusion criteria were as follows:
Papers that solely focused on virtual reality
Papers in which interaction between human users and ECAs was absent
Papers that described the design of an ECA but did not take the empirical validation in scope
Flow diagram of the study selection of part 2 of the scoping review.
The 22 analyzed papers suggest that a myriad of subtle interactions between users and computers play an important role in keeping a user motivated in continuing the Web-based intervention.
We formulated 8 themes according to our data extraction procedure. We further condensed these 8 themes into 2 main need-and-support concepts that in our view summarized the subject and that would help us during further analysis.
Users expressed the need for concrete feedback on their performance. Within the literature, this need is described as the principle of closure [
Users expressed the need for interest and support for the issues they are dealing with. This suggests that users of Web-based interventions could benefit from emotional support that acknowledges both the user’s endeavors during the change program and the originating issue the user is dealing with. This concept was based on the literature we found earlier [
The user needs and issues, mentioned in
The anonymity of Web-based interventions seems to play out both as a strength and weakness. Users feel encouraged to speak out but sometimes also feel isolated because of its anonymous nature. As formulated by McClay [
Both task-related and emotion-related system support could potentially counteract feelings of isolation.
Users seem to expect (and probably need) a deeper interest in their situation. Knowles et al [
This is a case for emotion-related system support.
User needs and issues and common user support mechanisms that can potentially fulfill these needs.
User need or issue | Support mechanism to fulfill the need | Source that describes the support mechanism |
1. Overcome users’ feelings of isolation | [ |
|
[ |
||
2. Deeper interest in the user’s situation | [ |
|
3. Interest in fundamental daily issues the user is struggling with | [ |
|
4. The ability for the user to refine the communication process | [ |
|
5. The user’s need for encouragement | [ |
|
6. Performance feedback mechanism for user responses | [ |
|
7. Users coping with experiences of negative affect during their change process | [ |
|
8. Creating a setting of accountability toward the user | [ |
aTask-related support: the confirmation that a user action has been successfully performed.
bEmotion-related support: acknowledgement of both the user’s endeavors during the change program and the originating issue the user is dealing with.
Users seem to wish for a form of deeper interest in their practical daily issues. This need is described by Knowles et al [
This is a case for emotion-related system support. In case the user requests practical advice for daily issues, task-related support on these issues can also contribute.
As mentioned by Donkin [
A noninteractive tool as a questionnaire is perfectly fit for structurally gathering experimental user data. However, it may be less appreciated as it “forces” the users to answer according to its rigid structure. Emotion-related system support provided alongside a more open interaction between user and system (ie, by means of bidirectional free text or free speech) could potentially (and so far theoretically) increase the user’s feeling of contributing to his or her own change process.
As noted by Donkin et al [
Encouraging users during the intervention can likely be achieved by emotion-related system support.
Somewhat comparable with the statement of Donkin et al [
Providing a direct task-related response such as “I have received your answers, thank you for your time and effort. Please allow me to comment on your answers” would immediately acknowledge the user’s effort invested. By later analyzing the user responses and by providing feedback via email, a second, more profound task-related support mechanism could be implemented.
As formulated by Todd [
As described by Bradbury [
For such a setting, certain preconditions are necessary, such as participants who understand and agree with the benefits of their expected future behavior. Other preconditions are concrete goal setting and performance monitoring. Task-related machine support can play a positive role by reminding the users of their goal set and by indicating which of these goals have (not yet) been met. Note that accountability might be harder to trigger among users who have been assigned to health interventions by their doctors and who did not primarily opt to participate by themselves.
A large body of studies on ECAs refer to the CASA effect [
Computers that display flattery texts toward their users are preferred by their users compared with computers that do not display such texts.
Computers that textually praise other computers are better liked than computers that praise themselves, and computers that
Users who are partnered with a computer on basis of a color (eg, the blue team) will have a more positive opinion about the computer and cooperate more with it than users who have to partner with a computer of the opposite, differently colored team.
As an explanation of the CASA effect, it has been proposed that humans have a strong innate tendency to make social connections with other humans and other living creatures such as pets. This human tendency becomes real when objects such as personal computers (PCs) demonstrate activities that could be socially interpreted by their users [
Themes on supportive embodied conversational agents.
Theme | Explanation | Sources |
1. Computers as social actors | Humans treat media in the same way as they treat other humans. | Systematic search: [ |
2. Open dialogue between user and computer | Embodied conversational agents (ECAs) have the ability to have an open verbal dialogue with users. | Systematic search: [ |
3. Visible conversational partner | Interaction with a “talking face” leads to more trust and believability. | Systematic search: [ |
4. Human-ECA relationship | Interactions with an agent can lead to a relationship, which is important to keep users engaged over time. | Systematic search: [ |
5. Measures of the human-ECA relationship | Human-ECA relationship quality can be measured. | Systematic search: [ |
6. Responsive verbal and nonverbal communication | Computers should have the ability to notice and respond to verbally and nonverbally expressed emotions from their user to create a more natural interaction. | Systematic search: [ |
7. Impact of ECAs on user motivation | There is evidence that ECAs can motivate users, which is highly dependent on ECA implementation, context, task, etc. | Systematic search: [ |
8. Methodological issues within ECA research | Most experiments into ECAs face similar methodological issues, which have to be taken into account when interpreting the research. | Hand search: [ |
aeHealth: electronic health.
The theme that follows is the ability of computers and ECAs to have an open verbal (textual or speech) dialogue with users. Within regular, day-to-day human-computer interaction events, a user who interacts with his or her information technology system will typically activate predefined menu options such as the “save as” option. Subsequently, the computer will respond to the request by presenting a pop-up window, which will enable the user to type in the file name of the document. In such a closed dialogue scenario, the interactions between the user and the software traditionally have a task-specific character (ie, serve to reach the specific goal of saving a document), have a short duration, and are typically initiated by the user (and not by the computer). In contrast, ECAs enable more open-ended and more relationship-oriented interactions. Interactions between ECAs and users can span multiple question-and-answer pairs and can therefore be interpreted as a dialogue.
The ELIZA (software created by Joseph Weizenbaum at the MIT Computer Science and Artificial Intelligence Laboratory Cambridge, MA, USA) study [
Later studies create richer dialogue contexts to explore the capabilities of computers interacting with humans. One of the examples is a study that has shown that a robot taking the role of a museum guide who uses, for example, empathy and humor in his conversation style led to a more positive attitude toward the robot than the same robot without this enhanced conversation style [
Main theories and effects of visible embodied conversational agents.
Embodied conversational agent (ECA) theory | Explanation | Source |
Theory of social inhibition/facilitation | When in the presence of others, people perform learned tasks better and novel tasks worse. Empirical results have demonstrated that this principle also applies for the presence of ECAs. | [ |
Social agency theory | By adding a visible ECA as a screen tutor, the social interaction schema is primed, which will cause the learner to try to understand and deeply process the computer-delivered instructions. | [ |
Social modeling/social learning theory | Humans derive their knowledge, attitudes, behavior, and goals by observing and imitating the surrounding social agents. | [ |
Situational dependency | Pedagogical agents are helpful when there is a need to increase companionship and decrease complexity. | [ |
Social exchange theory | People prefer equitable relationships in which the contribution of rewards and costs are roughly equal. This equity principle also applies to human-computer relationships. | [ |
Persona effect | The presence of a lifelike character in an interactive learning environment—even one that is not expressive—can have a strong positive effect on a student’s perception of his or her learning experience. | [ |
Image principle | The image of an ECA is not a key factor for learning; instead, the level of animation of the ECA is the key factor for learning. | [ |
The next theme is the visibility of the conversational computer depicted as a (either static or animated) human face. According to Lisetti [
Besides empirical research, there are multiple theories that support this notion. The theories that were mentioned in the included sources are listed and explained in
Despite these positive experimental results and theoretical support for a visible, human-like PC, the visibility subject is somewhat controversial. Strong claims against the human face are provided by Norman [
The fourth theme is the concept that regular human-computer interaction events result in a relationship. Routine interactions between a user and his or her computer should be regarded as contributions to this human-computer relationship, as is argued by Bickmore et al [
The question arises whether an ECA with a relationship-focused design could behave and be perceived as a competent social actor. This quality of the ECA as a conversational partner is impacted by the following:
The literature found mentions two regular measures with regard to the human-ECA relationship.
Working alliance is a construct that originates from the psychotherapy literature and has been described as “the trust and belief that the helper and patient have in each other as team-member in achieving a desired outcome” [
A second important human-computer relationship measure is rapport. Rapport has been described as “the establishment of a positive relationship among interaction partners by rapidly detecting and responding to each other’s nonverbal behavior” [
Within human-to-human communication, the exchange of nonverbal information plays a key role. Social psychologists assert that more than 65% of the information exchanged during a person-to-person conversation is conveyed through the nonverbal band [
The user first expresses his or her need and accompanying emotion through verbal and physical interaction with the machine, for example, through detectable gestures, usage of the keyboard, or spoken language.
Then, the system generates an affective reply, through words, speech, and animation with the intention to respond to the user’s need.
This response affects users in such a way that they become more involved in their further interaction with the computer.
Others, such as Doirado [
Concerning the importance of the affective loop and BDI, there are 2 stances:
Meta-studies and reviews [
Schroeder et al reviewed 43 studies and concluded that pedagogical agents have a small but significant effect on learning as ultimate outcome. Within their study, Schroeder et al [
Altogether, the evidence for ECAs capable of motivating users is inconclusive. ECAs, whether they are nonresponsive or responsive, provide a positive user experience as a result of their entertainment capabilities. Responsive ECAs when specifically designed to detect user frustration and to empathically respond to it have also empirically demonstrated positive effects on user attitudes. However, these positive effects have not yet been found in ecologically valid contexts. Instead they were found within constrained contexts such as games with clear win-and-lose rules and as a result of deliberately induced user frustration.
The inconclusiveness regarding ECA evidence as mentioned within the previous theme is claimed to be caused by methodological issues [
Different modalities used for output: (synthesized or natural) speech or text
Different levels of responsive emotional behavior: from textual responses projected alongside a static ECA to fine-grained ECA facial expressions intended to mirror the user’s facial expressions
Different roles: tutor, peer, interviewer, coach
Different implementations/different computer code applied as artificial intelligence to steer the ECA
Many of these issues can be resolved by using a common, open research platform for ECAs, such as the Virtual Human platform (as provided by University of Southern California (USC) and the Institute for Creative Technologies (ICT), Los Angeles, USA; see also [
Concerning the duration of the change programs, several studies (eg, [
Altogether Dehn and van Mulken [
Part 1 of this scoping review addressed the following research question:
Is there a set of generic user support needs that are currently not sufficiently addressed within eHealth interventions requiring repeated use that may result in a lower user experience and therefore lower user adherence?
We found various user needs and issues related to support, which we divided into the following two main categories:
It appeared that both task-related support and emotion-related support are regularly expressed user needs. Both needs therefore merit further attention in terms of research that aims to improve user adherence.
Part 2 of this scoping review addressed the following research question:
What are main supportive features of ECAs that could potentially address user support needs?
Information was scarce and a direct answer to this question could not be found. However, we were able to find relevant information on the ECA features of
Furthermore, we made two distinctions:
As described within
In contrast to nonresponsive ECAs, responsive ECAs are capable of performing more complex motivational tasks as described within the needs 3, 4, and 7. First, responsive ECAs are capable of having a dialogue with the user during which concrete daily issues the user is facing can be effectively discussed (need 3). Further research should focus on effective countermeasures for users losing interest interacting with responsive ECAs during longer-term interactions (eg, 4-10 weeks with daily contact) [
Dialogues between the user and ECA on deep, personal issues (need 2) are currently technically too complex to realize. Smooth interactions are a necessary condition for ECAs to become and remain a trustworthy counterpart. None of the ECAs found are capable of truly meeting this condition of smoothness. As a result of future progress within the artificial intelligence field, this may change for the better. For the moment, these dialogues should be best carried out by a human support provider.
User needs with supportive elements, associated embodied conversational agent (ECA) features, and the needed level of responsiveness of the ECA.
User need or issue | Supportive element | Associated ECA features | Needed responsiveness |
1. Overcome users’ feelings of isolation | Computers as social actors; visible conversation partner; human-computer relationship |
A nonresponsive embodied conversational agent (ECA) is sufficient | |
2. Deeper interest in the user’s situation | Computers as social actors; open dialogue; visible conversation partner; human-computer relationship; responsive verbal and nonverbal communication | No ECA is currently likely to be able to address this user need | |
3. Interest in fundamental daily issues the user is struggling with | Computers as social actors; open dialogue; visible conversation partner; human-computer relationship; responsive verbal and nonverbal communication | A responsive ECA is necessary; further research is advised | |
4. The ability for the user to refine the communication process | Open dialogue | A responsive ECA is necessary; further research is advised |
|
5. The user’s need for encouragement | Motivational effects | A nonresponsive ECA is sufficient | |
6. Performance feedback mechanism for user responses | Computers as social actors; visible conversation partner; human-computer relationship | A nonresponsive ECA is sufficient | |
7. Users coping with experiences of negative affect during their change process | Responsive verbal and nonverbal communication; motivational effects | A responsive ECA is necessary; further research is advised |
|
8. Creating a setting of accountability toward the user | Computers as social actors; human-computer relationship | A nonresponsive ECA is sufficient |
The ECA literature of part 2, for example [
This review has several limitations. Due to the nature of this study as a scoping review, no quantitative analyses were done, and selection of the studies was done by interpretation of the researchers. No exclusion criteria were applied with regard to the quality of the studies to ensure broad coverage of the studied topics.
As we looked for generic user support needs in part 1, we did not take the type of mental and/or physical disorder into account. In addition, we left out factors such as user personality. The rationale was to separate the subject of user experience from the user’s characteristics, but it is not certain that this separation always holds. Although we included user experience in our search string, we left out more fine-grained search terms, such as for user-centered design, to keep the search focused on the core issues. This focus on generic user needs has resulted in a broad overview of the needs and the possibilities of ECAs to address these needs, but when designing Web-based intervention for a specific target group, more research is needed to understand their specific needs for support.
Within part 2, we were aiming for on-screen solutions that could be added to the eHealth environments in
We conclude that users of self-guided eHealth interventions can likely profit from the support of
Responsive ECAs are also relevant from other perspectives. Psychological experiments extensively make use of questionnaires to gather user data. As touched upon in part 1, need 4, questionnaires are structured yet limited communication tools by design. The ECA’s sensors that deliver real-time signals on the user’s BDI during experiments can provide an additional source of user information for analysis. As an alternative to the sensor and artificial intelligence technology working in real time, logs on intervention usage (eg, number of log-ins, time in between log-ins) could predict lower user motivation.
To successfully
Finally, we would like to propose a research framework. Following the advice of Dehn and van Mulken [
Proposal for a theory-based framework for supportive electronic health (eHealth) embodied conversational agents (ECAs).
Keywords search part 1 and 2.
belief, desire, and interest
embodied conversational agent
electronic health
intelligent tutoring system
None declared.