We define active assistance technology as any technology involving automated processing of health or behavior change information that is ongoing as the user interacts with the technology. In other words, the technology continues to process the health-related information during interactive use and may adapt its responses. This contrasts with passive information technology such as storage devices, videos, and website design. It also contrasts with interactive systems that do not process information related to health or behavior change. For example, an interactive system can process user choices on presentation format (eg, video or text) and adapt to these choices during a session. However, this is not active assistance because the responses are not related to the semantic content of the health messages, only to their formatting.

In this way, the concept of active assistance draws attention to the distinction between semantic and nonsemantic information processing during an interactive session. This is important, because semantic processing entails a degree of delegation of health decision making to an automated system, which can free up human specialists. It also has more serious consequences if incorrect.

Furthermore, active assistance takes place in an environment in which citizens and experts participate actively in the behavior change intervention (combining push and pull). In other words, the role of the technology is not merely to deliver a fully expert-led intervention where users follow instructions with minimal understanding. Instead, the technology helps users to reflect and learn about the obstacles to successful behavior change. A desirable feature is that users should feel they have ownership or control of the intervention [8].

A concept related to active assistance is persuasive technology, which is “designed to change attitudes or behaviors of users through persuasion and social influence” [9]. Although many aspects of persuasion are relevant to active assistance, persuasive technology is not necessarily active in the way that we are defining it; passive techniques such as visualization and website design may suffice. However, some aspects of persuasion would benefit from active technology. In particular, the potential of the technology for unbiased information processing can help with self-monitoring and correction of inaccurate health beliefs, both of which are required in the persuasion model [10,11]. Similarly, many behavior change theories such as the transtheoretical model [12], social cognitive theory[13-15], and theory of planned behavior [16] imply the need for health education, accurate beliefs, and self-monitoring to effect behavior change and sustain it over the long term. Independently of persuasion and self-monitoring, automated information processing can also overcome some of the problems of self-report in measuring behavior change such as adherence [17].

In the context of these requirements, the following are key examples of active assistance technology that can support behavior change. The technologies may be used together or independently.

Some previous literature reviews have addressed topics related to active assistance.

Webb et al [2] conducted a systematic review and meta-analysis of the effectiveness of Internet-based interventions. The review found that interventions grounded in behavior change theories are associated with increased effectiveness, with the theory of planned behavior having the largest positive effect. The review used a taxonomy of behavior change techniques and found that using a combination of these techniques tended to be more effective than using only one. Effectiveness was also increased by communication methods such as instant messaging. Although the review did not consider the role of active semantic processing, the findings that a combination of techniques is more effective suggest an increased necessity for such automated processing due to the added complexity of interventions that combine more than one technique.

Fry and Neff [18] reviewed the effectiveness of interventions using automated prompts and reminders, finding that more than half of those (11 out of 19 reviewed) showed positive results. The findings suggest that tailoring of prompts and making them more interactive can make them more effective. However, the review did not specify the role of the technology in determining the timing and content of prompts: whether these are manually designed in advance, or whether they can be adapted using automated methods.

Lustria et al [4] defined second-generation interventions as those that are tailored to individuals. The definition did not specify whether the tailoring is an ongoing automated adaptation process that happens during continued user interaction. However, the review has some interesting findings. Many tailored interventions have components for developing self-regulatory skills and self-monitoring. For example, the user can often keep online diaries and send them to an expert for analysis, who then provides feedback. Automation of some of the semantic analysis steps in this process would reduce the need for human expert involvement, thereby reducing costs. Similarly, Revere and Dunbar [19] conducted an earlier review (1996–1999), finding that tailored interventions tended to be effective but not enough evidence that tailoring was superior to generic interventions. The review mentioned the possibility of dynamic tailoring but did not focus on the distinction between dynamic and offline computer tailoring.

Bickmore and Giorgino [20] reviewed the technologies used in dialog systems and presented some examples that have been evaluated as effective. Similarly, Corkrey and Parkinson [21] reviewed interactive voice response systems from 1989 to 2000, finding a growing research area but few applications at that time.

For the purpose of building personalized health models, Fernandez-Luque et al [22] reviewed the research on extracting content from social networking sites (eg, blogs, forums, or search patterns). These techniques have relevance in behavior change interventions but are still at an early stage and present many ethical and technical challenges.

There have been no comprehensive reviews of active assistance technologies in health-related behavior change.

We used the following strategy: <technology-related keywords> AND <health psychology-related keywords> AND health AND “behavior change”.

Technology- and psychology-related keywords were combinations of the following: [automated OR technology OR Internet OR “mobile phone” OR intelligent OR computer-based OR interactive OR agent-based OR adaptive OR “context-aware” OR “machine learning” OR “pattern recognition” OR robotic OR “virtual reality” OR semantic OR “knowledge-based” OR “decision support” OR ontology OR dialog OR “natural language”] AND [assistance OR intervention OR personalization OR persuasion OR adherence OR compliance OR motivation OR affective OR emotion OR reminder OR prompt].

We conducted some preliminary searches on a wide range of databases including CINAHL, EMBASE, Inspec, ISI Web of Science, PsycINFO, and ScienceDirect. However, we found that Google Scholar had a wide enough coverage to allow it be used instead of these databases. This is consistent with recent empirical studies such as those of Howland et al [24] and Walters [25]. Therefore, we decided to use only Google Scholar and PubMed. The date range was January 2005 to January 2012. Only articles written in English were included.

We used the definition of active assistance above. The studies should have included at least one of the three active assistance technologies listed and have been intended for interactive use by clients or patients attempting to change their behavior. Studies may have described a new technology or design to be used in behavior change interventions, or they may have reported an evaluation of an actual intervention using the technology. Qualitative studies are necessary to evaluate usability and acceptability of active technology. Similarly, prototypes and works in progress help to provide an overview of the current research concepts and their maturity.

We excluded the following kinds of study: (1) interventions merely delivered using the Internet, CD-ROM, or other medium, where the technology only facilitates transmission of information from expert to user, (2) feedback, in which there is no automated processing of personal health-related information—for example, receiving emails from a human counselor would be excluded, (3) tailored-offline interventions, in which the computer processing is used by health professionals to tailor the intervention before or after the user interacts with the technology—this is the case where the semantic content processing is not part of the interactivity with the client (although a health professional may have interactive access), (4) simple data collection or preprogrammed reminding without any pattern recognition or inference (eg, pedometers); interventions where the semantic content of reminder messages were configured by the client were also excluded for this reason—for example, the planning tool of Soureti et al [26] is a passive technology and therefore excluded, (5) fixed games or simulations, in which there is no automated player with semantic processing capability, and (6) general guidelines or research roadmaps in which different options are discussed.

We divided studies into the following categories: (1) quantitative and qualitative evaluations of interventions using active technology, (2) pilot studies of new technology, and (3) prototypes or designs that are being developed or tested.

For each study, we asked the following questions. What kind of active technology was used? Was it effective? What was the role of the technology in the intervention? Was it theoretically grounded?

From Table 1, it is clear that dialog systems were used widely in interventions using active technology (19 studies). Of these, embodied conversational agents (ECA) were components of 8 studies. ECAs are visual (embodied) characters that can conduct conversation with a user. They are agents in the sense that they can sense and respond to their environment [68]. The agent’s environment might include a virtual game environment, the text inputs of a user, physical behavior data, or all three of these. Similarly, its responses might include conversational output, actions within the virtual environment, or effects on the real environment in the case of a robotic system [52,57,66]. Within the ECA context, motivation and empathy were central themes in 6 studies.

Ecological momentary assessment [69]) is a process of capturing the momentary experiences of participants—for example, using online diaries for the purposes of research. Not only are the participants’ environmental and behavioral circumstances recorded but so are their mental states. These can include, for example, their current goals, beliefs, mood, and emotions. Ecological momentary interventions [70]) are based on ecological momentary assessment. Levin and Levin [50] conducted a feasibility study on pain monitoring and recommended management using interactive voice response. Sorbi et al [61], in their pilot study, obtained a positive result for migraine attack prevention based on experience sampling using random mobile calls. Randomization overcomes memory bias. Another study that has relevance for the automation of ecological momentary assessment is the semantic analysis of blogs describing combat experience [48].

Most studies (18) were prototypes or design concepts. A total of 17 were feasibility or usability studies. Only 6 were RCTs measuring effectiveness for behavior change [29,30,36,39,45,67]. All of those also included a qualitative self-report of user experience. Bickmore and Picard [29] reported mixed results for an agent-based dialog system (FitTrack), with short-term improvement in physical activity during the intervention but reduced activity after follow-up. Interacting with an agent led initially to more physical activity than the control (nonagent) for sedentary participants (45 out of 101). Users enjoyed interacting with the relational agent more than with the nonrelational agent, but there was no significant difference in physical activity between the two agent conditions (both increased during the intervention). Additionally, users found the agent’s repetitiveness to be annoying. The authors suggested that long-term interaction with the agent might prevent the drop in physical activity. Bickmore et al [30] reported positive acceptance and increased physical activity for FitTrack with older adults (n = 21) over 2 months, but there was no significant decrease in loneliness.

The study of Bickmore et al [36] involved mechanisms for reducing the repetitiveness of a dialog system and enabling it to tell stories. The effect on long-term user engagement was positive, but the effect on actual behavior was negative. The authors proposed several explanations, such as increased enjoyment during interaction inhibiting activity. Their most recent research [67] reported a positive effect (n = 70) for sustained step count over 12 weeks, but there were mixed effects for satisfaction and motivation.

The study of Consolvo et al [39] involved a personalized display of the results of an activity recognition system and reported a positive effect on actual behavior and user experience (detailed below).

Jin [45] reported a positive effect of a virtual agent on stress management self-efficacy and enjoyment (n = 60).

Behavior change models were used in 14 studies. Motivational interviewing [71] was the most widely used (8 studies). Of these, 5 were dialog systems, in which motivational interviewing was used as a general philosophy for dialog design [35,41,47,51,52]. Other behavior change theories that guided the designs were the transtheoretical model [31,35,40,47,65], self-efficacy [41,45], theory of planned behavior [28], and the health belief model [45].

Articles not indexed in Google Scholar or PubMed would have been missed—most scholarly publications, however, are captured by Google Scholar. The review required the mention of “health” and “behavior change” in the articles. We did not include gray literature such as white papers and unpublished reports. We selected the date range (2005–2012) to focus on recent developments, but this may also have excluded innovative earlier work.

The review required specific mention of a key technology. There may be some interventions that use active technologies, but the studies did not mention this. Similarly, some studies mentioning only general intelligent technology were excluded from the full-text review because they could not be categorized. This may be a limitation because included studies need to involve significant interdisciplinary communication between technology specialists and health specialists. On the other hand, it may be a strength, as such communication is important for understanding a particular technology in context.

Since we limited the search to behavior change, it is also possible that many of the technologies are being applied in other areas of health informatics. For example, we found some prototypes early in the date range (2006–2007) but found no subsequent study relating to behavior change. In these cases, citation searching sometimes revealed further development of the techniques and algorithms, but no application in the health domain.

The potential of active technologies for dynamic and unbiased information processing is not being fully exploited in current health behavior change research. Most research has focused on specialist areas, such as dialog and ECA systems, and has been largely restricted to the study of persuasive dialog in respect of relational behavior and motivation of behavior change.

In addition to the potential benefits of active technologies, there is a need for a thorough understanding of the potential risks. Expected benefits such as that of dynamic tailoring of the content and presentation of information can be measured using established evaluation methods (eg, [87,88]). Risks such as misinformation due to the unexpected behavior of an algorithm may be more difficult to uncover. It is important to study realistically complex scenarios of the uses of active assistance systems. Such studies need to reveal how the system components interact to produce information, and how these components might in turn interact with wider systems such as the Semantic Web, clinical records, and personal health records. Wider still, we note that many health behaviors are socially mediated; therefore, active assistance research needs to bridge cognitive and sociotechnical aspects in order to deliver maximum public health benefit.

To exploit the full potential of active assistance technology, health behavior change researchers need a deep understanding of how the different components of information systems might change the intervention—its safety, effectiveness, efficiency, and acceptability. This requires more collaboration between disciplines such as health psychology, computer science, cognitive science, health informatics, medical sociology, and public health and health promotion.