Personalization in Real-Time Physical Activity Coaching Using Mobile Applications: A Scoping Review

Mobile monitoring for health and wellness is becoming more sophisticated and accurate, with an increased use of real-time personalization technologies that may improve the effectiveness of physical activity coaching systems. This study aimed to review real-time physical activity coaching applications that make use of personalization mechanisms. A scoping review, using the PRISMA-ScR checklist, was conducted on the literature published from July 2007 to July 2018. A data extraction tool was developed to analyze the systems on general characteristics, personalization, design foundations (behavior change and gamification) and evaluation methods. 28 papers describing 17 different mobile applications were included. The most used personalization concepts were Feedback (17/17), Goal Setting (15/17), User Targeting (9/17) and Inter-human Interaction (8/17), while the less commonly covered were Self-Learning (4/17), Context Awareness (3/17) and Adaptation (2/17). Few systems considered behavior change theories for design (6/17). A total of 42 instances of gamification-related elements were found across 15 systems, but only 6 explicitly mention its use. Most systems (15/17) were submitted to some type of evaluation. However, few assessed the effects of particular strategies or overall system effectiveness using randomized experimental designs (5/17). Although personalization is thought to improve user adherence in physical activity coaching applications, it is still far from reaching its full potential. We believe that future work should consider the theory and suggestions reported in prior work; leverage the needs of the target users for personalization; include behavior change foundations and explore gamification theory; and properly evaluate these systems.


I. INTRODUCTION
I T IS well understood that we need innovative approaches to address the alarmingly low levels of engagement in physical activity (PA) among the general population. The use of wearable and mobile monitoring technologies for this purpose has exploded from a standing start in the last 10-12 years. In this line, there has been important progress on the use of these systems to increase adherence to PA [1], both for the healthy population as well as for prevention and management of chronic diseases [2]- [4]. In particular, PA coaching applications are defined as systems that aim to motivate the user to change their activity behaviour by means of a coaching element [5]. A common motivational strategy used in PA coaching is Feedback, which is a way to stimulate such change by generating awareness of the user current behaviour.
Despite an increase in number, complexity and accuracy, these systems face the underlying challenge of user abandonment, which has been highlighted in recent publications [6]- [8]. Studies report that users may stop using such technologies once they have gathered enough information about their routine activities [8], [9]. Also, when it comes to commercially available solutions, these often target young and active people [10], who do not require special recommendations nor motivation. For users who need to be persuaded to become active, effects seem promising in the short-term but users do not feel additional inducement to use the devices [10], [11].
In order for these persuasive technologies [12], [13], to make an impact on user's behavior, researchers have highlighted the importance of including a strong theoretical basis considering different aspects of behavior change. In particular, a metaanalysis from Fanning et al. shows that mobile-based PA interventions tend to be more effective when relying on behavior change theories (BCTs) and models [14]. Some of the most used are the Social Cognitive Theory [15], the Transtheoretical Model [16] and the Self-Determination Theory [17]. Yet, relatively few health apps explicitly rely on these theories [18]- [21]. Also, as brought up by Fogg et al. in [12], for the behavior change program to have a sustainable impact, it is essential that automated systems engage people. In that context, gamification, defined as This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/ "the use of game design elements in nongame contexts" [22], has recently emerged in the design of persuasive health technologies [23], [24] with particular uses in health and fitness apps [25]. The field has been on a rapid rise [26], with evidence suggesting its potential in creating pleasant experiences for the users of technology.
Overall, the high levels of user abandonment have been commonly associated with a low perceived personal relevance and a lack of engagement [11], [27], which highlights the importance of exploring additional motivators to adopt sustainable healthy behaviors. A key factor that may determine persuasiveness to behavior change is related to creating personalized, or tailored, experiences to each individual [28].
As stated in [10], each individual is unique, and dynamic, in a sense that a strategy that works for one, might not work for another. It is believed that mobile-based interventions that are closely tailored to the individual's convictions and motivations are more likely to be observed and remembered [29]. Therefore personalization, or tailoring, helps in increasing the intended effects of communication, which can contribute to overcome the lack of adherence and effectiveness of these systems [30]. Furthermore, with mobile technologies allowing for more accurate, usable and engaging real-time support, there is an increase in alternative forms of personalization that can potentially make a difference in the effectiveness of PA coaching applications.
Hawkins et al. in [30] defined tailoring as "any of a number of methods for creating communications individualized for their receivers …". Since 2008 a number of papers were published related to tailoring technology-based health interventions [31]- [33]. The most recent work in this area, by op den Akker et al., deals specifically with real-time tailoring of PA coaching applications [5]. The authors report on a literature survey, with data collected until July 2013, and define a tailoring model relying on 7 different concepts: Feedback (FB) -presenting the measured amount of activity performed to the user (can vary in timing, content and representation); User Targeting (UT) -conveying that communication is designed specifically for the user; Goal Setting (GS) -creating and updating user-specific goals based on users' activity trends and patterns; Inter-human Interaction (IHI) -providing support by form of interaction with other humans; Adaptation (Ad) -directing information to individual's status on key behavioral factors; Context Awareness (CA)using users' external context to provide relevant information; and Self Learning (SL) -learning reactions of the users' to previous communications. However, the authors highlighted the lack of systems exploring the full potential of smartphones and available contextual information for the design of more complex personalization; the lack in application of Ad, CA and SL; the lack of clear specification of theoretical foundation for specific design decisions; and the lack of work demonstrating the effectiveness of tailoring in a more structured and controlled manner.
The exponential increase in evidence related to mobile PA coaching technologies motivated this scoping review. The aim is to systematically map the most recent developments on techniques used in these real-time systems that aim to motivate users in reaching their personal activity related goals. The specific objectives are to: i) expand the knowledge on personalization in real-time PA coaching applications, by presenting current advances in the field, ii) understand if previously reported gaps have been addressed and identify opportunities for future work, and iii) to provide a comprehensive analysis of these applications considering general system characteristics, behavior change theoretical foundation, use of gamification and system evaluation. Due to the scope of this review, it is clear that it does not cover all the work done on tailoring/personalization nor on physical activity coaching systems. Instead it explores a narrow topic that sits in-between these two fields.

II. METHODS
A methodological scoping review [34], [35] was conducted to study real-time personalization in PA coaching mobile applications, and was built upon a prior literature survey and model published by op den Akker et al. in 2014 [5]. The protocol was drafted using the PRISMA extension for scoping reviews (PRISMA-ScR) checklist and explanation [36]. This extension provides reporting guidance for this specific type of knowledge synthesis. The checklist contains 20 essential reporting items and 2 optional items, which detail how to conduct and report scoping reviews.
This article does not aim to derive statistical evidence or conclusions from existing literature, as this is not applicable for scoping reviews.

A. Search Approach
The search strategy, used to identify potentially relevant studies, was based on [5]. Two searches were performed, one on July 12th, 2017 and another, to update the previous results, on July 10th, 2018. For the first search, 7 databases were selected as the source of information: PubMed; Association for Computing Machinery (ACM); ScienceDirect; IEEEexplore; PsycINFO; CINAHL; UCDlibrary-onesearch. A second search was conducted in PubMed, ACM and IEEEexplore, to include relevant papers published since the first search (from July 12th, 2017 to July 10th, 2018). These 3 databases were used as they covered all the selected studies from the first iteration. The search strategy was based on the previous review in this specific topic [5], and was conducted as follows: (personalized OR personalised OR personalization OR personalisation OR individualized OR individualised OR individualization OR individualisation OR tailored OR tailoring) AND ("physical activity" OR "daily activity" OR walking OR exercise OR exercising OR "activities of daily living") AND (coach OR coaching OR feedback OR motivate OR motivation OR stimulate OR stimulation OR promote OR promotion) AND (app OR application OR system OR device). When offered the option, keywords were searched in the entire text of the article.
To be included, papers needed to: be written in English; be published in conferences or journals over the last 10 years (from January 1st, 2007 to July 12th, 2017, when the first search was performed); deal with PA coaching systems, including either promotion of daily activities (e.g.,: walking, running), in-session coaching, or prevention of sedentary behavior; describe systems with some kind of personalization to the user; describe real-time coaching systems; describe systems that are smartphone-based and make use of embedded and/or external sensors to measure PA. Studies were excluded if they were: exergame-based; targeted at disease rehabilitation; targeted at specific exercises (e.g.,: rehabilitation exercises or machine exercises); with no direct connection between sensor and smartphone (e.g.,: systems in which a server is an intermediary of data synchronization between sensor and smartphone were excluded); with no real-time communication with the user. Also, papers were excluded if the full paper was not available. The real-time definition used in this study was the same considered in [5]. Following this definition, a real-time system is one that has a direct connection between sensor and feedback device, and that is able to communicate constantly with the user and provide immediate feedback on measured performance. For such real-time purposes, the feedback device could be either the sensor itself or a mobile phone/smartphone. The focus was on the latter modality as it offers more opportunities for richer processing and visual display. Some systems included in the previous review [5] were found through manual search, which means they could be missed in our database search. In such case, these systems were also included for analysis. Additionally, the research team searched for other papers with further system details by scanning the reference list from the included papers and through manual search in google scholar. The manual search terms included the author names and the system name.

B. Study Selection Procedure
The study selection procedure was divided into four phases. The identification phase consisted of retrieving articles from the databases and was performed by one researcher (FMG). After removing the duplicates, resulting papers went through the screening phase, first by title and then by abstract, by two researchers (FMG and ORR). In case of doubt, the paper was included and reviewed again in later stages for final inclusion. Then, a full-paper review was conducted by the two researchers (FMG and ORR) for eligibility, considering the defined inclusion/exclusion criteria. The included papers were then assessed for data extraction. Discrepancies on study selection and data extraction were solved by consensus. Cohen kappa coefficient was calculated, to measure inter rater agreement, in the title review (k = 0.734 on a random sample of 38 papers) and in the abstract review (k = 0.85 on a random sample of 100 papers) [37], revealing a substantial agreement between the two reviewers.

C. Data Extraction
A data-charting form was constructed by two researchers (FMG and ORR) based on the suggested in [36]. Following an iterative process, the authors identified the most relevant categories related to the objectives of this study and they were included or updated in the form. Discrepancies were solved by consensus.
The data was abstracted regarding 4 main categories. The first one included general system details on: target population, target activity, main features, inclusion of human (coach)-in-the-loop, platform used and market availability. The second aspect addressed was personalization, which was based on the model and framework provided in [5], and covered: coaching mode (if the communication is provided during daily activities, during exercise sessions or during sedentary behaviors), personalization concepts and mechanisms, technical implementation, communication properties addressed and inclusion of user profiling. The third category regarded the persuasive design considerations, and addressed the use of theoretical foundation and the inclusion of gamification elements [23]. The fourth, and last, category was on evaluation methods and included information on: study design (based on the framework in [38]), study population, study/intervention description, outcome measures and persuasive strategies comparison.
Theoretical foundation was only extracted if theories were explicitly stated in the paper. The presence of gamification elements was extrapolated for all systems based on [23]. The systems were, however, distinguished based on explicit or implicit use of gamification. Explicit use was considered in papers that included gamification-related terms in the system description (e.g.,: "games", "game-based", "game-like", "gamification", …), while implicit use was attributed to those systems that include gamification-related elements without explicitly reporting to do so.
The details of the data-charting form for each of the included systems were independently extracted by two researchers (FMG, ORR). After data extraction, classification discrepancies were resolved by mutual agreement.

D. Data Analysis
Two researchers (FMG and ORR) went through the taxonomy table for all the included studies, in an attempt to find relevant insights to the research objectives posed for this review. Common patterns, contradictory results, and gaps were also analyzed for all studies.

III. RESULTS
We retrieved 1274 results from the first database search from which 341 were duplicates. To the remaining results, 200 new papers (without duplicates found) were added from the second database search, summing a total of 1133 results for screening. 555 met the title review criteria and, from these, 154 met the inclusion criteria in the abstract review. From the full-text review, 18 out of the 138 were then selected for inclusion, to which we added 8 more papers with further description of the included systems and 2 more with a system included in the review of op den Akker et al. In total, we included 28 papers [39]- [66] that covered 17 different PA coaching applications that explore personalization strategies in their design. A flow diagram representing the full process is shown in Fig. 1.
16 out of the 28 included papers (some describing the same system) were new compared to the previous review, contributing with 10 new apps and 1 updated version of an already reported system. 5 of the systems found by their team were not retrieved in our database search, as these were probably identified through the manual search they have performed in Google Scholar and their personal libraries. From those 5 systems, we only included 1 (with 2 associate papers), which respected our criteria for being a mobile phone app. Therefore, a total of 17 systems were included in our study, and the analysis is presented in the following sections of this paper.

A. General System Characteristics
An overview of the systems' details is provided in Table I and includes a descriptive summary of the systems' objectives, main features, platform and market availability.
App screenshots of some of the included systems are presented in Fig. 2.

B. Personalization
The personalization concepts implemented in each app are discriminated in Table II, where the systems were categorized according to the coaching mode: over daily life activities (12 apps); during exercise sessions (3 apps); or to reduce sedentary behavior (2 apps). The most frequent used concepts were Feedback (used in all 17 apps), Goal Setting (15 apps), User Targeting (9 apps) and Inter-human Interaction (8 apps). The less frequent were Self Learning (4 apps), Context Awareness (3 apps) and Adaptation (2 apps). Move2Play and INTELiRun covered the highest number of personalization concepts, both with 6 instances, followed by Sweetch and AAFS with 5 instances.
The following subsections describe the different personalization concepts and mechanisms used by the systems. 1) Feedback: Feedback (FB) is the most obvious form of personalization and is used by all of the included apps.
Considering the intention of the FB, we can separate the included systems in two main categories, those that aim at promoting PA (daily activities or exercise) and those that aim at reducing sedentary behavior (SitCoach and B-Mobile). On11 and Analytic, Social, Affect feature both intentions. They monitor sedentary behavior and try to reduce it, but also monitor active periods and provide coaching to achieve certain activity goals.
In terms of timing, FB initiative can be with the user, when information is provided only if the user looks at a glanceable display (e.g.,: UbiFit Garden, BeWell), or with the system, when it provides cues for the user to walk/run faster or slower based on user's speed (e.g.,: Haptic Personal Trainer, u4fit). Another aspect of timing relates to the type of coaching, which can be during daily activities or during exercise sessions.
Considering the content of FB, only 2 papers seemed to address this property, reporting specifically on the phrasing of FB. The authors from the paper on the AAFS system, tested three types of messages: encouraging, neutral and discouraging. Also, in the SitCoach study, the authors tested motivational messages phrased using 4 different persuasive strategies based on the social influence theory.
Regarding FB representation, the simplest form is through text messages and/or text notifications. Most systems also make use of some kind of visual feedback, either through graphs or more complex visual displays such as avatars (e.g.,: Move2Play) or virtual ecosystems that change based on the user state (e.g.,: UbiFit Garden and BeWell). Audio feedback was explored in 3 systems, TripleBeat, Haptic Personal Trainer and SitCoach. These last two also explored tactile cues as a mechanism for providing feedback.
2) Goal Setting: Goal setting (GS) is used by most of the included applications (15 out of 17) and is a strategy commonly associated with FB.
Many of the applications, in particular the ones that provide coaching during life activities, include simple daily and/or weekly numerical goals such as number of steps, distance or activity duration. On the other hand, applications that provide coaching during a workout session can provide goals in the form of session routines that the user can follow step-by-step and that can be based on the type of activity, pace, duration, or time spent in a specific heart-rate range (e.g.,: TripleBeat and u4fit).
Some systems allow the user to choose from a set of high-level goals and then present more specific objectives or suggestions according to that goal choice. TripleBeat proposes a workout schedule based on a general goal, for example, to lose fewer calories but burn more fat, or to improve cardiovascular/ respiratory health. On11 allows the users to select from three types of goals: Keep Healthy, Lose Weight, or Burn Calories, and then suggests appropriate activities to help them achieve those goals. Regarding the creation of overall activity plans/calendars, only a few systems seem to implement this feature (e.g.,: CAMMInA and u4fit). In u4fit, for example, the schedule is composed of several sessions per week for several weeks. In general, most systems let the user set or review their goals. Only u4fit and AAFS allow the training goals or plans to be set or adapted by a professional (human-in-the-loop).
Several systems consider the user's characteristics (e.g.,: age, weight, preferences, physical activity level) in the definition of goals, combining GS and UT. iBurnCalorie provides a daily personalized caloric estimate that the user needs to expend, based on personal information provided during the registration process and daily requests. StepbyStep takes into account the user's baseline level of walking, then automatically sets a daily walking goal reflecting a 10% increase over the baseline level. The personal trainers from the u4fit system use information from the user's physical activity profile and his progress to adapt the plan provided. In the AAFS system, the goals are set based on a baseline of activity for each day of the week and are kept up to date based on the daily progress and the activity pattern throughout the day. Move2Play recommends the appropriate amount of activity based on characteristics present in user and domain models. Also, the Sweetch app continuously adapts goals based on the user's real-world behavior and weight.
BeWell+ is a particular case that uses community adaptive GS, adapting the goals based on the performance of the users compared to other similar users that perform slightly better.
In terms of representation of goals, most systems are making use of visual displays like progress bars, goal lines in graphs (see systems figures), or cues in virtual ecosystems, rather than relying solely on textual format.
3) User Targeting: A variety of forms of User Targeting (UT) are used by 9 of the included applications.
The most transparent way of implementing UT is implemented by Move2Play, SitCoach, Sweetch and iBurnCalorie apps, which include the users' names or nicknames in the main screen and/or in the textual feedback.
Other less transparent approaches involve adapting the information based on the user characteristics (e.g.,: age, weight, preferences, physical activity level). As described previously, this is mostly used in pair with GS strategies: the iBurnCalorie makes use of personal information provided during the registration process (age, gender, height, and weight during the registration process) and daily estimates of the users' mean caloric food intake per day; StepbyStep takes into account the user's baseline level of walking; u4fit system uses information from the user's physical activity profile and his progress; the AAFS system, uses the users' baseline of activity, daily progress and activity pattern throughout the day; Move2Play system adapts to the characteristics present in user and domain models (e.g., age, gender, physical activity fitness); the Sweetch app makes use of the user's real-world habits and weight data; and INTELiRun incorporates age, height, weight, heart rate and injury history. On11 is a system that combines UT with FB by providing estimations of burnt calories to the users based on personal information from the user profile interface (gender, age, height, and weight).
A less reported and detailed strategy of UT is to consider the users' preferences to provide personalized suggestions or recommendations. For example, Move2Play recommends the appropriate type activity, based on the user activity preferences, in a way that will lead to the fulfilment of the daily plan and that the user will enjoy. Also, On11 provides walking suggestions and recommends detours based on personal preferences.

4) Inter-Human Interaction:
Inter-human Interaction (IHI) is used in 8 of the included systems and can also be covered through a variety of different strategies.
Social comparison is the most common form of IHI. The iBurnCalorie app makes use of a trending graph at the bottom of the home screen that provides an overview of the user's status (activity trend or driving trend) compared to the social trend. Also, Move2Play and StepbyStep include a feature that ranks users in a leaderboard according to their achieved results. Move2Play tries to ensure a fair competition by considering the relative effort from the users, based on their fitness and physical condition. A more complex strategy is implemented by TripleBeat, where the user can compete with other virtual or real users, or the actual user on past runs. This competition is defined by how well users achieve their predefined goals, not by who is faster or exercises longer, aiming to promote healthy goal achievement. In the paper, the authors detail the strategies used for this purpose and to create a fair competition, which are based on a novel performance score function and similar opponent selection. BeWell+ app implements social comparison but implicitly, in combination with GS and FB, by comparing the performance of individual users with other peers, but in this case with users that are similar in behavior and that perform slightly better (through similarity matching).
Some systems incite teamwork among the users. Such can be done either through group-based competition, as done in Move2play and the social version of the Analytic, Social, Affect system, or by simply allowing the users to create shared goals with others, which is also provided as an option in Move2Play app.
Another form of IHI relies on enabling sharing of results and providing support among a network of users. The social version of the Analytic, Social, Affect system includes an electronic message board where participants can post comments or suggestions to the other participants. In Move2Play users can connect with friends, see their progress and compare their  results, and this is made easier with the integration with popular social networks. The u4fit application allows users to create, save, and share their workout session results on Facebook.
INTELiRun also allows users to share their running data on multiple social media platforms, to send and receive challenges from others and to find similar fitness running mates (via a matching feature in the app).
A less frequent strategy of IHI is to include interactions with a human coach, which is one of u4fit's main features. In u4fit a coach (chosen by the user) creates a tailored workout plan, analyses the training data gathered from the app and has the option of modifying the plan and motivating the user by means of the internal messaging system. 5) Adaptation: Only 2 system covers the concept of Adaptation (Ad) as a personalization strategy.
The AAFS combines Ad with FB, by providing tailored messages based on the user's score on self-efficacy and stage of change questionnaires and also on the user's own baseline level of physical activity (UT). With that information users are identified as one of 8 personas and recommended one of six feedback strategies. The INTELiRun app also prompts the user with a personality questionnaire, and provides feedback specific to each personality type. 6) Context Awareness: Context Awareness (CA) is used in 3 out of the 17 included systems. However, the papers do not provide much details on its use and implementation, except for the one describing On11.
In Move2Play, they reported on a PA recommendation algorithm that takes into account a Domain Model, which holds stable facts/knowledge about how we generally exercise and what factors affect the amount of activity. The model integrates factors such as day of the week, the month or season and the weather. On11 encourages users to walk by adding detours into their usual travel routes, such as home-workplace route or routes to their frequently visited destinations, and suggesting mini-walks around their workplace such as walking to the coffee lounge. The implementation of the recommendation system is detailed in the paper and takes into consideration the user's current location, location history, date and time.
The Sweetch app is context aware in the sense that it notifies the user to do activity only when the user's calendar indicates available time and recommends specific activities based on the user's surrounding locations. 7) Self Learning: Self Learning (SL) is covered in 4 of the included systems.
The BeWell+ system uses SL in the way it implements the similarity matching algorithm for GS and FB, by repeating the matching process as new behavior data from the user is available. In that way, the system will incrementally set more challenging goals to the user, by selecting as frame of reference higher performing people that are still relatively similar to the user. Move2Play includes a user model (containing physical fitness, activity patterns and activity preferences) that is built incrementally with the use of the system. This model is fed into the rule-base system that provides recommendations on the amount and type of activity, hence combining SL with FB, UT and GS. As mentioned previously, the system seems to be partly in conceptual phase and no technical details are provided. The AAFS intends to be a continuously adapting system that takes into consideration the user's routine, for GS, and their progress through time regarding the psychological constructs, for Ad. The system uses a smart reference module that automatically plots in a graph self-adjusting goal lines for each individual based on their routine of activity. The goal lines provide information on the percentage of total activity that the user should achieve at different times of the day and are built based on the past user activity, on that day of the week and specific time of the day, but with a slight increment. Furthermore, users are prompted with self-efficacy and stage of change questionnaires that, combined with the classification of their activity pattern, allow the system to automatically select the best feedback strategies for each user. The Sweetch app uses machine learning to create insights about the individual's life habits (e.g.,: schedule, activity patterns, driving and walking routes, surroundings) and then, using advanced algorithms adapts the timing and content of FB, CA recommendations and GS. The system learns what types of messages result in better compliance for the specific user in terms of for example, day of the week, time, location and types of messages. However, no details are provided regarding its implementation.

C. Behavior Change Theories
Only 6 out of 17 analyzed systems explicitly included behavior change theoretical foundations (see Table II). All were classified into the category of "real-time coaching over daily life activities". The most reported BCT was the Goal-Setting Theory (GST) [67], used in 4 of the systems. Other included theories were the Transtheoretical Model (TTM) [16], [68], Presentation of Self in Everyday Life [69], Cognitive Dissonance Theory (CDT) [70], Social Cognitive Theory (SCT) [71], Self-Efficacy from SCT [72], Self-Regulatory Principles of Behavior Change [73], Social Influence Theory (SIT) [74], Operant Conditioning Principles [75], [76], and Self-Determination Theory (SDT) [77]. 3 systems used a combination of BCTs, all including the GST. The papers did not clearly report design decisions based on these theories.

E. System Evaluation
The reviewed papers reported on the evaluation of 15 out of the 17 included systems (see Table III) [39]- [41], [44], [46]- [48], [51], [53]- [62], [64], [66], [78], [79]. 6 of these systems were evaluated in more than one separate study, with 4 being submitted to both nonexperimental and intervention studies. In total, 10 systems were submitted to nonexperimental studies (including user experience, system functionality, validation, usability and user acceptance). 9 systems were evaluated in terms of system effectiveness, 5 in randomized experiments, and 4 in quasi-experiments. None of the quasi-experimental studies included a control group. The quasi-experiment and randomized experiment studies, involved between 10 and 55 participants, and 27 and 199 participants, respectively. Regarding duration, the quasi-experiment studies lasted between 1 session and 3 months, and the randomized experiment studies between 10 days and 6 months. From all studies, only 4 compared the effects of different persuasive and/or personalization strategies.

IV. DISCUSSION
In this scoping review we identified 28 studies, published between 2007 and 2018, describing 17 real-time PA coaching applications that used some form of personalization. The review contributed with the analysis of 10 new mobile applications and 1 updated version of an already reported system, compared to previous results reported by op den Akker et al. in 2014 [5]. Furthermore, we took a comprehensive approach, following the PRISMA-ScR, gathering detailed information on general system characteristics, personalization, behavior change foundation, gamification and system evaluation.
The global picture of having 17 personalized real-time PA coaching systems being published in the last 10-11 years reveals a considerable interest of the scientific community in this field of  research. However, we expected more contributions taking advantage of the current complexity and accuracy of smartphones and monitoring technologies in implementing more advanced real-time personalization. Prior work [5] has provided a clear conceptual framework for tailoring/personalization, and provided numerous ways to explore different tailoring mechanisms for the design of new and more complex systems. Yet, none of the reviewed papers referred to such work or any other personalization or tailoring theories. Hence, many gaps remain to be addressed in this field.

A. Personalization
Most systems made use of the more simple forms of personalization such as Feedback (17), Goal Setting (15), User Targeting (9) and Inter-human Interaction (8), while only a few systems covered the concepts of Adaptation (2), Context Awareness (3) or Self Learning (4).
In what concerns to Feedback (FB), further work is still needed to find the best way to communicate the information to the user, considering the different communication properties (intention, content, timing and representation). Such properties were implicit in most included studies, but few specifically explored this topic. Two new studies addressed the content of feedback messages, one compared encouraging, discouraging and neutral phrasing and the other compared different persuasive strategies based on a BCT. However, many aspects related to content and intention of the communication are still to be addressed, and can contribute to a better understanding of 'what' to communicate to the user. One possibility would be to study system variations with FB intended at promoting physical activity and/or sedentary breaks. Moreover, the variety in representation forms of FB, together with the need to address the timing (initiative, moment and frequency) of communication, provide numerous opportunities for tackling the lack of user engagement. All these aspects are related with the intensity of coaching provided, which also remains to be evaluated in detail.
Goal Setting (GS), normally used in pair with FB, was the second most used concept. The most common form of GS was through simple numerical goals or, in some cases, in the form of training sessions and routines. Compared to the previous review, there was a considerable number of new applications considering the user characteristics, however most only relied on information at the time of registration and only two considered the user's progress through time. Also, two new systems allowed a professional to adapt activity goals (human-in-the-loop), with one study addressing the effects of supervised coaching. Finally, there were only a couple of systems making use of PA plans or schedules and none adapting to the user's own schedule. The AAFS presented it conceptually in future work, suggesting the use of information on the users' schedule (sleep, travel, work and leisure time), tracked manually by the user or automatically using location data, to determine changes in activity goals.
User Targeting (UT) and Inter-Human Interaction (IHI) were used in more than half of the included systems, with UT being more predominant in the new applications (8/10) compared to IHI (3/10). The instances of UT were, in some cases, in the form of including the user's name or nickname, but mostly in combination with GS, considering information on the user's characteristics (e.g.,: age, weight, height, PA level). Still few systems explored the progress in PA level. Besides Move2Play, which was included in the previous survey, only 2 other systems, u4fit and AAFS, considered the users' PA data for personalization. Also, few systems considered the user's preferences to provide personalized suggestions or recommendations.
IHI appeared in the form of social comparison, teamwork, network support and support from a real coach. The most common form was social comparison through individual or group-based competition, followed by network support, which was included in some systems through sharing of results or support from other users. Only 2 systems used IHI by including teamwork features (e.g.,: goal sharing) and only 1 new system, u4fit, included support from a real personal trainer.
Regarding Adaptation (Ad), there is still room for research exploring communication matched to the users' psychological traits (e.g.,: personality, stage of change, self-efficacy, player type). In our review, only the AAFS and INTELiRun covered the concept of Ad. The AAFS system provided FB adapted to user's scores in questionnaires on stage of change and self-efficacy, and INTELiRun adapted FB to the users' personality traits, which were also inferred from questionnaires.
The more sophisticated approaches, Context Awareness (CA) and Self Learning (SL), are still far from taking the most of existing technologies and reaching its full potential in PA coaching systems. Only 3 out of 17 included systems covered CA (Move2Play, Sweetch and On11). However, only the paper describing the On11 provides examples and implementation details on the use of such strategy. Based on frequently visited destinations (e.g.,: home-workplace), On11 proposes different travel routes and suggests mini-walks around the workplace.
Similarly, there was a lack of systems exploring SL (4/17), with only 2 new systems (AAFS and Sweetch) covering the concept compared to the previous survey. SL was used in combination with GS to incrementally set more challenging goals in BeWell+ and Move2Play. The latter included a user model that was continuously updated. The AAFS updated its mechanisms for GS and Ad by learning the user's routine of activity and psychological changes, respectively. The Sweetch app, seemed to take the most advantage of SL, by learning the user's life habits and compliance with certain types of messages, and combining it with FB, CA and GS. Overall, SL is a strategy that can be used as improvement to any system, as it involves learning with the users' interactions with the app. Also, it is closely related to UT as it relies on the creation of user models that adapt through time. Therefore, there is still space for more work using intelligent SL algorithms to dynamically optimize other personalization strategies.
Besides addressing the existing gaps, new contributions should provide a more technical description of the systems. Some of the included applications lacked in details on the systems' architecture, on how the personalization strategies were implemented, and what algorithms were used, which would be relevant for facilitating future research.

B. Behavior Change Theories and Gamification
Although there is a strong body of literature on BCTs, only 6 of the 17 included systems were based on these, and the papers failed to clearly present design decisions taken from such theories. These findings are aligned with the discussed in [5]. However, some of the systems followed a design based on motivation theory. For example, MPTrain/TripleBeat refers to the Persuasive Technology theory from Fogg et al., 2003. Also, Move2Play has been designed considering motivation as the core system part, to tackle user abandonment, and is built upon informative, social and gamification strategies.
Also, despite the recent hype in gamified fitness apps [25], there was a lack of systems considering the existing game or gamification theories, with only 6 out of the 17 included applications explicitly reporting its use. However, a total of 42 instances of elements that could be related to gamification were found in 15 systems, mostly related to progress and rewards.
Gamification and BCTs share similar constructs [80], [81], but they also relate and can be mapped to personalization strategies. For example, BCTs can be directly associated with Ad (e.g.,: adapting to the stage of change) or IHI (e.g.,: considering the Social Influence Theory). On the other hand, gamification elements can be considered in the representation of FB (e.g.,: progress bars, avatars) or in combination with UT (e.g.,: avatars), with GS (e.g.,: levels and rewards), or with IHI (e.g.,: social interaction). Therefore, we believe both BCTs and gamification should be considered in the design of personalized PA coaching applications, to help creating highly individualized, engaging and effective experiences both in the short and long-term. Also, it is important to consider the theory behind gamification to understand how to best leverage it for motivation and to avoid any potential detrimental effects of misuse. For example, the simple integration of external rewards (e.g.,: points or badges) without considering a design driven towards increasing intrinsic motivations (e.g.,: sense of progress), might engage users in the short-term but fail to do so in the long-term.
A related conceptual question that remains to be explored regards the understanding of whether the creation of an app should be driven by gamification and other persuasive strategies, or by BCTs.

C. Evaluation
Near half of the evaluation studies assessed the effectiveness of the proposed system (10/22), but, half of them did not include a control group and near half (6/10) had sample sizes of less or equal than 30. Also, the same proportion of studies (6/10) evaluated the outcomes in the short-term, with intervention duration of less or equal than 1 month, remaining unclear the long-term effects these systems have on users. An exception is u4Fit whose effectiveness was analyzed using a retrospective observational study using data collected in more than one year. Therefore, there is still a need to assess in a structured and controlled manner the long-term effects of using these real-time personalized systems, with an analysis on user adherence and attrition rates.
Besides, there is still lacking evidence on the individual effects of particular personalization strategies, as normally these systems are tested as a whole. Some studies make use of different versions of the same system, that are compared in small scale studies. However, adopting the optimal methodological approaches is time and resource consuming and therefore challenging to put into practice. Furthermore, as commented in [5], whether or not personalization in real-time PA coaching applications increases engagement and motivates behavior change has yet to be rigorously examined.

D. Other
The majority of the systems were targeted at coaching over daily activities (10/17), with fewer targeting exercise-based (3/17) or sedentary-based coaching (2/17). However, 2 of the systems that provided coaching on daily activities also promoted sedentary breaks or provided simple feedback on sedentary behavior. These results are in accordance with the highlighted in [82], that an emerging area of PA intervention research is focused not only on increasing PA, but also on decreasing sedentary behavior.
Few of the included systems were targeted at particular populations, such as the elders or chronic patients. These individuals have specific needs, which can inform design decisions on personalization mechanisms, motivation elements and behavior change constructs. Such can be facilitated through user-centered design approaches, as done for the It's Life app [50]. We believe that system personalization can be particularly relevant in the context of disease prevention and management, having the potential to increase the acceptability of these applications by creating relevant and targeted user experiences. This has also been highlighted in literature, for example, in a study on the opinions of cancer survivors for mobile PA applications [83].
A particular factor that raised concern was that only 3 of the included systems were available in the app stores, which reveals the existing barrier in knowledge transfer and implementation of research work to society. This could be related to a discontinuation of the development of such systems, which in some papers were only presented in the form of early concepts or prototypes.
This work is extending the knowledge on this topic primarily through inclusion of new research that has been published in the 5 years that have passed since the last substantive review in the field. In these 5 years we have witnessed a significant number of new systems and associated research studies based on personalized coaching apps. Informed by the previous work, we have explored if previously reported gaps were addressed in the new contributions, we identified new gaps and we provide suggestions for future work. We took a more systematic approach to this scoping review, which allowed us to perform a detailed analysis on each particular aspect reported by op den Akker et al., but also to explore the topic of gamification and to identify the type of interventions being used to test these solutions. We present a comprehensive analysis of 17 PA coaching systems, which maps the general characteristics, personalization strategies, the theoretical foundations, and the evaluation methods used by these technologies. Also, the information is presented in a streamlined layout, using tabular format, which makes understanding and comparison easy to the reader. Hence, this work can help inform future work related to the development and evaluation of technology-based health promoting and coaching systems, and also the research focused on exploring ways to overcome the underlying challenges of user abandonment and lack of engagement with these systems.

V. LIMITATIONS
The process of analyzing the applications regarding the realtime criteria was challenging, as some papers did not fully or explicitly detail the architecture of the systems. Such doubts were solved by consensus between the researchers. More advanced and elaborated types of personalization might have been used in systems not considered in this review, such as those with feedback delays due to more complex processing on the cloud. However, this was not the purpose of this review, which focused on analyzing the personalization mechanisms used specifically in real-time systems. We did not take a comprehensive review of BCTs, as none of the authors was experienced on such procedure. Instead, we extracted the theories reported in the systems' description. The authors consider taking a more extensive analysis on this matter in a future publication, by considering the Behavior Change Taxonomy [84] or the CALO-RE Taxonomy [85] to classify these apps. Statistically significant conclusions were not drawn, given that in the majority of cases evaluation was conducted through pilots and/or small scale trials, and only one study assessed long term effects. Additional results could have been obtained by taking into consideration specific journals, specific conference proceedings, grey literature, other databases, paid publications or even unpublished work. To reduce the chances of missing relevant papers we have searched both journals and conference proceedings from 7 different databases across multiple fields. Given the particular focus of this literature review, and our adoption of the same inclusion criteria as op den Akker et al., it might have led to the exclusion of some studies using well known commercially available fitness trackers. Chiefly this is due to the strict inclusion criteria in this review regarding real-time coaching that relies on direct communication between sensor and the coaching app/interface (without online synchronization), which is not easily achieved when integrating with commercial activity trackers. Studies that focused on coaching apps that involved delayed access to sensor data via a web API were excluded due to the lack of realtime feedback. Furthermore, there might have been other apps relevant to this review, not available in the literature, that could be found in the app stores. For example, apps such as Google Fit have recently introduced real-time feedback and coaching features. However, these have not yet been reported/evaluated in the scientific literature and so are not included in this review.

VI. CONCLUSION
In this work, we reviewed the most recent contributions on real-time physical activity coaching applications that used personalization strategies. From our findings, it is clear that these systems are not referring to the theory and practice in the field and, in most cases, are using the more simple forms of personalization. There is still limited evidence addressing the gaps highlighted in prior research, which include the lack in exploration of Adaptation and the more advanced forms of personalization, Self Learning and Context Awareness; the lack in proper evaluation of the effects of particular personalization strategies and overall system effectiveness; and the lack of design foundations on behavior change theories. Besides, we believe that future work should consider the model and suggestions proposed by op den Akker et al. in 2014; explore the use of gamification to increase engagement; and leverage the end-user insights for system design and personalization.

ACKNOWLEDGMENT
This study was performed in collaboration with Salumedia Tecnologias S.L.U., the University of Seville and The Insight Centre for Data Analytics from the University College Dublin.
Authors contributions: Francisco Monteiro-Guerra led the scoping review, participating in the screening, data extraction and analysis stages, and wrote the core of the manuscript. Octavio Rivera-Romero contributed to the screening, data extraction and the analysis, and supported with manuscript organization and writing. Luis Fernandez-Luque and Brian Caulfield contributed to the conceptual design, framing and manuscript revision.