This section describes in detail the results from (a.) UCD workshop 1 (iteration 1), (b.) UCD workshop 2 (iteration 2), and subsequently, (c.) the overall design considerations based on the workshops to develop working research prototype of eCoach for personalized activity recommendation generation.
Results from Workshop 1
Iteration 1: Scenario Co-Design with Four Dialogue-Lab Zones
From UCD workshop 1 we identified end-users and their context and followed by, developed a concept (user requirements) based on the focus group discussion to answer the identified research questions as described in Additional file 2-5.
Despite the initial briefing by our team about the motivation of a health eCoach app, all groups suggested that eCoaching must be user-friendly, accessible, effective, evidence-based, predictive, transparent, and accurate. In goal settings, end-users told that goals could be open, flexible, adjustable, specific, measurable, attainable, relevant, sharable, and real-time. One group suggested considering cultural aspects, social traits, and individual preferences regarding coaching or motivating. The other two groups suggested the inclusion of key performance indicators (e.g., an overall health index computed by combining several health parameters to forecast health status and update the timely progress indicator) in automatic lifestyle coaching.
All groups suggested that long-term goals must consist of multiple short-term goals, daily goals must be different from long-term goals, and personal preference based. An individual will be motivated if rewards, performance comparison, constructive motivational feedback, and personal preferences are incorporated in eCoaching. One group suggested including gamification, mood assessment, and iconography to convey feedback without requiring much cognitive involvement of the user. The other two groups ideated to consider a progress evaluation graph or report, fitness status evaluation, goal comparison, timing feedback, reminder design, and high-level contextual information in feedback generation to motivate participants in self-management.
Discussion: Concept Design for Personalized Recommendation Generation
The discussion opened a broad scope for the eCoach system to promote a healthy lifestyle. Usability, credibility, and effectiveness were identified as essential factors to determine the performance of an eCoach system. According to the discussion, the needed data collection for activity, nutrition, and habit is necessary without burdening the participants. Personalized goal setting, health risk prediction, goal evaluation, and evidence-based contextual real-time tailored recommendation generation are essential features for health eCoaching. Goals must be intelligent, customizable, personalized, and context-driven in goal settings. Iterative recommendation generation based on health status adjustment, reminder design, adjustable preferences, progress evaluation, rewarding, realistic feedback generation, and an appropriate information visualization may motivate participants to self-monitor and manage their goals. Recommendation generation can be combined with personalized mood assessment feedback to determine the satisfaction level of participants. The eCoach app must exhibit beyond state-of-the-art innovation to be better than existing apps to manage individual behavioral change. This workshop helped to refine the questionnaire set in the eCoach prototype design and development for meaningful, personalized recommendation generation.
End-user’s remark on personalized recommendation generation –
“My FITBIT scares me a bit, because it constantly tells me that I sleep too little. It is perceived as annoying bullying and I cannot set up that I do not want all this feedback. My experience is that I like to see that I have been active from week to week, and I probably think that I am more conscious, and that it motivates me to make the right choices".
We created a basic initial eCoach prototype for personalized activity coaching given by the participant’s discussion and design to capture the high-level plan for goal management and tailored recommendation generation in activity coaching and interactions predicted across the four dialogue-labs zones. Researchers involved in UCD workshop 1 created an eCoach prototype over the next month using data and objects of the workshop. The prototype was further modified based on the outcome of Workshop 2.
Results from Workshop 2
Iteration 2: Scenario Co-(re)Design with Three Dialogue-Lab Zones
We started the workshop with a group discussion focusing on preference(s) and motivation. The selected topics were – goal settings, response and coaching, and interaction type.
According to group-1, goals can be generic as well as personalized. In our eCoaching, personalized goal management will be more meaningful than the existing market apps. They addressed that goal setting is an essential aspect in eCoaching. Goals can be set up by a doctor, a nurse, or a person. Thus, a contextual consideration is necessary for the eCoach design and development. As suggested, goals must be broken down into more detailed, specific goals linked to the more significant life priorities social and competitive perspectives. Group 2 indicated that motivations could be - user-based, situational-based, and environmental-based. An evidence-based personalized recommendation generation strategy will be very relevant for our eCoaching. According to group 2, selection of appropriate target group, presentation of data, selection of the device, type recommendations, and innovative motivational feedback presentation are essential in our automatic activity coaching with the choice of feedback generation frequency. Group 3 highlighted that interaction type is highly related to “user type” and their “emotional state or perspective”. The interaction design in our eCoaching must be two-way, adaptable, ubiquitous, easy to comprehend and visualize, accessible, customized, and personalized.
Discussion: Preference Settings for Personalized Recommendation Planning
From UCD workshop 2 we gathered end-user feedback on the personal preference settings (goal settings, response type for coaching, and interaction type) for personalized recommendation generation and visualization in a health eCoach app based on the focus group discussion to address the RQ5 and its sub-questions. In this workshop, we narrowed the scope of holistic behavioral coaching for managing body weight to only activity coaching to reduce sedentary time.
In goal settings, the goals can be personalized or generic. The generic goals in activity coaching can be general activity guidelines set by the WHO [6,7]. In contrast, athletes or obese or overweight people who want to stay active or reduce their weight to a normal range can set the goal, differing from WHO’s generic guidelines. The personalized activity goals can be multiple types (e.g., weight reduction, staying active, body fat level, proper sleeping) and need prioritization. Besides the selection of goal types, goal setting is also essential. A question may arise who will set the goals: a doctor or a trainer or the person itself! However, it depends on the context. Goal scoping in context is also an essential factor in effective coaching. Therefore, it should be broken down in a more readable, detailed, and specific way to link to the purpose. Besides, in successful goal management, social or community perspective (e.g., doing activities together) and/or competitive views (e.g., ranking, rewards) should be addressed. Overall, goals shall be “SMART”: specific, measurable, attainable, relevant, and time-bound.
Motivation is the desire to take action to achieve a goal. It is a critical factor in setting and achieving goals. Motivation is one of the driving forces behind human behavior. It includes the desire to continue working towards meaning, a purpose, and a life worth living. In eCoaching, motivation is an essential factor in daily life activities. Motivation differs from person to person based on the context (e.g., feedback generation to motivate a blind participant differs from a non-blind or color-blind participant). Participants can be encouraged with personalized, evidence-based, and contextual recommendation generation and its purposeful presentation (e.g., graphs, selection of colors, contrasts, visual aspects of movements, menus, adjustable with device type). Charts can produce a visible reflection of time-bound activities; however, app developers should consider the device’s battery usage.
Interaction is an action that occurs due to the mutual influence of two or more objects. The concept of two-way effects is essential in interaction, not one-way causal effects. The factors associated with a good interaction design are – two-way interaction (e.g., having a diologue), ubiquitous interaction (e.g, interaction at home, outside, office or in running or walking), opportunistic (e.g., triggered automatically), adapted to the situation (e.g., former activity the user was doing at the same place, time frame, adaptive in some way based on user’s instructions (e.g., visual, audible)) or, interaction preferences (e.g., user needs to see anything, only hear something, feel something, emotional needs, understanding (e.g., complexity), and motivated), visualization of graphs (e.g., what will you use the graphs/voice for?), frequency of interaction (e.g., hourly, twice/thrice per day, per day, weekly, bi-weekly, monthly), accessibility (e.g., voice, chart or graph, text to speech, text), situation awareness (e.g., situation awareness, multimodal interaction), usable and accessible following the international standards, culturally adapted following the cultural conventions, error reduction by design (e.g., redundancy), and personality (e.g., type of user and their action). Notification generation and presentation are a part of interaction and can be persistent or not. In notification design, a balance should be maintained between relevance, persistency, and disruption.
End-user’s remark on motivation –
“I wish to expend 7*X (X>0) calories per week. I can spend more than X calories on a day when I am highly motivated. Then, it would be nice if the system saves the extra calorie expenditure in a virtual energy bank that I can expend on a lower motivated day or treat myself to my favorite food (e.g., a chicken burger).”
“The app should generate contextual recommendations to motivate. Example – I am highly interested in soccer, and the app knows it. While I am walking or running, the app can track if any soccer event is progressing nearby and can recommend me with a message like if you walk or run X kilometers, then you have a chance to enjoy an exciting soccer game.”
End-user’s remark on feedback generation –
“Daily feedback would be better instead of every minute or hour.”
“Personalized activity recommendation should be presented in the form specialized graph or chart based on activity, goal setting and goal achievement to motivate participants.”
“Feedback could be internal or external. Internal feedbacks should be generated through the device or eCoach app. External feedbacks can be generated from external sources.”
End-user’s remark on interaction –
“Graphs: for someone without academic background or low graphical literacy, how do they understand? May other forms of interaction from the eCoach app.”
“Think of presentation of graphs: understand the level of literacy when having visual text. It is widespread. It is important to think very clearly about various questions. What is the goal of the graph? What type of information is needed? How can they adapt of different levels of literacy (e.g., visual numeric literacy)? Is it possible to have different shapes and forms and screen sizes?”
“One notification every one hour may be too disruptive.”
We presented the initial activity eCoach prototype in UCD workshop 2. We received feedback from participants in three dialogue-lab zones to modify it further to improve the quality of goal settings, motivational status visualization from self-monitoring, and personalized feedback and recommendations. The overall design and modular implementation of the ProHealth eCoach prototype is described in the following sub-section.
Design and Development of the ProHealth eCoach Prototype
Here, we describe the high-level design consideration for the ProHealth eCoach prototype. The UCD workshop 1 has given an overview of the necessary data to be collected from the participants relevant to our research’s goal. UCD workshop 2 helped preference setting, recommendation generation, and its visualization. Our developed ProHealth eCoach app for personalized activity coaching consists of the multiple modules described in Table 2, and the corresponding data considered for prototype design is shown in Table 3. The software architecture of the ProHealth eCoach app development is depicted in Figure 2. Please refer to the video (see Additional file 6) to see the demonstrators in action.
On a conceptual level, the activity eCoaching framework consists of – a. high-level components (e.g., activity monitoring, sleep monitoring, monitoring based on self-reports) and b. low-level components (e.g., step prediction, sleep trend analysis, determination of good goal, effective feedback generation for behavioral motivation) as depicted in Figure 3.
Participants can select single or multiple high-level component blocks for eCoach-based self-monitoring and recommendation generation. In the framework, a semantic ontology can be used to transform distributed, heterogenous health and wellness data (e.g., sensor, self-reported questionnaire) into meaningful information, including health state prediction . We have considered activity monitoring based on time-series data processing with Long Short-Term Memory (LSTM) networks . Here, we presented activity prediction as a set of numbers or intervals and used its visualization for motivational purposes. However, the usability study and the efficacy evaluation of the eCoach app for behavioral motivation is the future scope of research. In our design consideration, the eCoach system has access to contextual weather data, activity sensor data, and questionnaire data. The overall modularized eCoach app design and its implementation is described below, addressing ideas and concerns.
The log-in has been kept as simple and secure as possible. We have planned to collect person-related and activity data without personal identity disclosure. Only authorized users can access the eCoach system. Each participant has been provided with a unique user identifier (UUID), and they will be able to access the system with personal email-id and modifiable password. The system is further protected with the “eduVPN” network. Activity data can only be shared with the researchers to create meaningful information out of raw data. Sharing data through social media or any other means is prohibited by NSD rules. The simple log-in interface of the eCoach app is depicted in Figure 4.
Data Collection with eCoach System Prototype
The data collection has been divided into four parts –
- Activity data collection with wearable Bluetooth enabled (BLE) low energy activity device,
- Questionnaire for daily weight reporting (to analyze over a period of time whether activity coaching has an impact or not!), feedback (or survey), and the reporting of technical problems (without personal identity disclosure) during study in progress,
- Personal preference settings (goal-settings, response, and interaction), and
- Contextual weather data collection with OpenWeather representational state transfer (REST) application programming interface (API) against API Key validation.
We used the MOX2 medical-grade (CE certified) accelerometer-based low energy activity sensor for continuous monitoring [38,39]. The device flawlessly measures and transfers high-resolution activity data, such as activity intensity, weight-bearing, sedentary, standing, low physical activity (LPA), medium physical activity (MPA), vigorous physical activity (VPA), and steps for every minute. The collected data is well suited for physical activity classification (LPA, MPA, VPA) and posture detection (sedentary, (such as sitting or lying), standing, and weight-bearing). The recommended wear locations of the device are thigh, hip, arm, or sacrum. We used the publicly downloadable Android MOX2 mobile app to capture individuals’ activity parameters into the smartphone’s download folder. We then used our developed eCoach app to periodically transfer the activity data to the eCoach backend server tagged with the unique user-id, following the android secure file access policy. Participants had the following two options to upload their activity data from their smartphone to the remote eCoach server – automatic (to upload data automatically after every regular interval) or manual (if automatic data upload fails due to technical problems). The personal health, wellness, and questionnaire data are sent from eCoach app to remote eCoach server via a REST API (HTTP POST) to store them in a Postgres database in line with General Data Protection Regulation (GDPR) and Norm for information security and privacy in health (NORMEN) guidelines. No disclosable personal identifier has been collected with the questionnaire, complaint, or feedback (survey) data.
The MOX2-5 activity sensor is a 3-dimensional accelerometer with a 25-100 Hertz sample rate (dimensions 35 x 35 x 10 mm). Its sensitivity is 4mg/LSB. Maastricht Instruments had developed it. It is dust and waterproof gives a battery backup for seven days, and is built with a rechargeable “Lithium Ion125 mAh”. The current version of the MOX2-5 activity sensor is not suitable for classifying activities into the following detailed activity classes: cycling, swimming, rowing, and skiing. Therefore, the participants must report them manually as questionnaire data in the latest version of the eCoach app. The MOX2 sensor-based and questionnaire data collection interfaces of the eCoach app are depicted in Figure 5 and Figure 6. The daily weight reporting data will help to analyze if the regular physical activities or behavioral motivations impact gradual weight change. It can be a helpful direction in obesity and overweight case study with eCoaching.
We have designed interfaces for the questionnaires to collect personal preference data, such as goal setting, response, and interaction (see Figure 7). There are two goal types – system-defined general goals for staying active following the guidelines of WHO and person-defined goals (as athletes might want to get coached towards specific training goals). The duration of the goal period can be 4-12 weeks or more based on personal preferences. The goal-setting can be short-term (e.g., daily, weekly) or long-term (e.g., bi-weekly, monthly). The eCoach system should encourage end-users to reach their long-term goals with the generation of tailored recommendations and the achievement of short term goals.
In our eCoach app, we have considered the following pre-selected default values for the preference settings and the graphical user interface (GUI) design are depicted in Figure 7.
- Goal Type: Generic or Personalized
- Goal Period: 4 weeks
- Response Type: Representation of steps, VPA, MPA, LPA, sedentary bouts, future step prediction and interval prediction value
- Interaction Mode: Graph, Text, Audio
- Interaction Frequency: Regular interval, Daily, Weekly
- Interaction Medium: Text (e.g., push notification), Audio
All the preference and physical activity data are recorded in a relational database using semantic annotation. Individuals are always allowed to view and update their preference data. A hybrid (data and rule-driven) health state monitoring component is responsible for analyzing physical activity progress and followed by the generation of recommendations to reach personal activity goals (see Figure 8).
Monitoring and Recommendation (Feedback) Visualization
The app keeps track of an individual number of steps, duration of VPA, MPA, and LPA (in minutes per day), and sedentary bouts until the monitoring period gets over. Participants can actively monitor or track the number of exercises they have performed over the day or week based on their preferences. They will have the option to see their historical performances as well. At the end of the eCoaching session, they can report notes on their satisfaction with using the app. In UCD workshop 2, end-users showed interest in simplified metrics. Therefore, the eCoach app provides numerical feedback on the activity performed on simplified graphs. Here, feedbacks are of two types to motivate participants – indirect visual feedback and direct (e.g., textual pop-up notification generation). The participant receives daily as well as cumulative feedback at the end of the session to view their progress towards the goal.
In our activity eCoaching app, we have considered a hybrid health state monitoring component. During health state assessment, the module can predict the activity pattern of the participants (e.g., steps), automatically for the next “n” days (n>0) based on the temporal pattern in data. It can help participants to identify which kind of activities they should perform to reach their long-term goals. Temporal analysis on data (e.g., deviation in activities) helps to analyze the pattern in human activities and generate evidence-based tailored recommendations to motivate participants (e.g., comparative statistical analysis in activity data between weeks W1, W2, and W3 helps to determine if any deviation or improvement in performance or in which week the participant was more active). These recommendations can be contextual with the inclusion of weather information (e.g., tomorrow morning, the weather is sunny, and temperature is between 15–18-degree Celcius (C). Therefore, you can plan to walk for one hour or perform similar activities).
We have formatted activities in minutes per day or steps per day instead of calories which is inaccurate and difficult to understand for the users how calories relate to the activity goal. Moreover, for estimating future activity in terms of "steps" based on time-series monitoring data processing using LSTM, we focused on probabilistic interval prediction rather than abstract point prediction. A prediction interval gives an interval within which we expect to remain with a specified probability.
A prediction interval can be written as,
Where, “c” depends on the coverage probability and in one-step interval prediction its value is 1.96 (95% prediction interval where forecast errors are normally distributed). is the estimation of the standard deviation in the h-step forecast distribution (h>0). However, LSTM implementation, calculation of residual errors in temporal step data, and h-step prediction interval calculation is beyond the focus of this paper. By default, we have used c = 1.96. However, participants can choose the value of “c” up to 1.28 (80% interval).
In UCD workshop 2, end-users agreed to visualize their activity intensity simplified and briefly. Therefore, we had not considered infantile animations, sound like feedback when goal achieved as they might cause unnecessary interruptions. We have prioritized weekly performance evaluation rather than daily performances as participants can be active and maybe less active on the next day. A balance of activities must be maintained to achieve the short-term weekly goals to reach long-term monthly goals. We have shown a sample recommendation visualization screen in Figure 9. In the figure, daily step count has been represented with a target daily step count based on the goal settings and LSTM-based interval step prediction. The nature of step prediction is dynamic and depends on the steps achieved.
The recommendation module generates personalized and contextual recommendations based on the predicted health state. Recommendations can be direct (for example, pop-up notifications or alerts) or indirect (for example, activity status visualization). Instant notifications can contain two types of messages: (a.) formal To-Do (for instance, “You need to complete another 1500 steps in the next three hours to reach your daily goal”) and (b.) informal motivational notifications (e.g., “Good job, Keep going! You have achieved targeted steps.”). In the activity eCoaching framework, the messages are annotated in a semantic ontology. To inform the user about activity in progress, we have used the indirect approach for recommendation visualization, and to give direct instant notifications, we have considered pop-up text alerts. The participants can select the notification frequency as part of the app preferences. By default, we have considered activity notifications every three hours between 8 AM and 11 PM; however, the user can modify that. These notifications are timely alerts. It will help participants to stay on the right track either with motivational messages or with activity improvement suggestions. Notifications have been kept short, understandable, and positive. We have depicted sample push notification generation screens in Figure 10.
We have considered a simple emoji and a textual message to represent individuals' short-term (e.g., daily weekly) and long-term (e.g., bi-weekly, monthly) goals. We have used three emojis to classify individual progress to reach personalized goal into three groups – well done (😊) [10 credit points], up-to-the mark (😐) [5 credit points], must be improved (☹) [0 credit point]. All the credit points can be reimbursed against “Food bank”, as "reward" means that the user can eat a bit more, if he has trained more. We will decide to offer a list of potential food items in the “Food bank” against weekly accumulated credit achieved. It is a motivation to do more activity. In the future, we will enhance the reward generation with demographic clustering and profile ranking methods to motivate participants. We have depicted a sample personalized weekly reward generation in Figure 11.