Every year, hundreds of randomized controlled trials (RCTs) evaluating the effects of exercise in diverse population groups are published. This research has demonstrated the medicinal capacities of exercise for numerous chronic diseases.1 Consequently, exercise is considered an essential component in the management plans of many healthcare interventions.1-3 Regular exercise has also been shown to have wide-ranging health benefits for non-pathological groups. There is evidence to suggest that a sedentary lifestyle may be an even stronger predictor of mortality than smoking, hypertension, and diabetes.4 National and international campaigns, advertisements, and public health guidelines have been developed to increase public awareness of the benefits of exercise 5 6. Despite this, inactivity (and the increased risk of morbidity and mortality associated with it) remains a significant public health concern.7 8 This raises the question: if people know that regular physical activity and exercise are good for them, why are so many inactive?
Even in RCTs, low levels of adherence to exercise sometimes belie the treatment effect to such an extent that results in comparison to a non-exercising control group are statistically insignificant.9 10Adherence to exercise, which is defined as the degree to which the target intensity and volume are achieved,11 12 is likely to be worse among the general population, who are not enthusiastic volunteers in a research study and who are being closely supervised by a research team.13
Many theories and models have been proposed to explain why adherence to exercise is suboptimal.14 15 A recent umbrella review identified a number of key factors for improving adherence to exercise.16 Among the 14 factors that were identified, individualization of the exercise program, making sure that it integrates easily into participants’ daily living schedule, continually monitoring and providing feedback on progress (and adapting the exercise program accordingly), ensuring that users have an active role in goal setting and using technology to deliver the exercise intervention were deemed to be important in improving adherence.16Patient education was also deemed to be crucial to increase self-efficacy, enhancing the knowledge about what they can do and what they can change to improve their overall health.17 18
Mobile technologies (e.g., smartphone apps and wearable activity trackers) are a cost-effective, scalable way of delivering exercise programs that incorporate these factors to users. Over 60% of adults worldwide own a smartphone, with worldwide penetration rates highest in the U.S (where >80% of the population use a smartphone).19 In addition to being able to deliver interventions through wireless internet and messaging connectivity, smartphones have in-built tools like GPS, inertial measurement units and cameras that can objectively measure several exercise parameters 20-22. Smartphones also have powerful computation and communication capabilities that enable the use of machine learning and artificial intelligence to individualise each user’s exercise program.
One machine learning technique called reinforcement learning (RL) is a particularly promising approach to exercise individualisation and adaptation. In RL, a decision making agent takes specific actions that result in preferable states in its environment.23 With every decision made and action taken, the system environment transitions to a new state. The agent then receives either positive or negative reinforcement from that environment. The mapping of state to action is called the policy; the goal of RL is to learn an optimal policy that maximizes the amount of rewards it receives over time.
For example, in the environment of exercise prescription, the agent might be incorporated in a smartphone app. A user’s self-reported satisfaction or their engagement (i.e., more frequent use) with the agent would be potential state variables. The action space that the agent can modify might include the different parameters of exercise (e.g., the number of sets, reps or the duration of rest periods in a resistance training program) or the types of exercises being recommended. Finally, the reward function could measure the discrepancy between high and low levels of user satisfaction or engagement with the app housing the system. RL is particularly suited to the environment of exercise prescription because these decisions are typically made sequentially: the agent generates an exercise session for a user, the user completes that session and reports their satisfaction with it and the next session is modified accordingly. However, while several frameworks have been proposed to automate exercise prescription based on user demographics, fitness levels, engagement behaviours and preferences,24-27 trials evaluating fully computerised exercise prescription with smartphone apps are lacking.28-31
Therefore, the aim of the ‘Augmenting Exercise prescription with Artificial Intelligence’ trial (AEAI) is to test a smartphone app that generates adaptive exercise regimes, incorporating RL to personalise the composition of exercises within sessions based on user satisfaction. We will compare user satisfaction between sessions generated by the RL system to (1) [insert brief description of Samsung here], (2) a control condition that generates session randomly. The primary outcomes for this aim will be users’ satisfaction with exercise, which will be defined by an abbreviated 8-item version of the Physical Activity Enjoyment Scale (PACES).32 33 Our primary hypotheses are that users will report higher satisfaction and demonstrate greater adherence to exercise programs generated using RL, compared with sessions that are generated randomly or using predefined (i.e., non-personalised) templates.
Our secondary hypothesis is that users will demonstrate higher levels of engagement with the smartphone app when exercise sessions are generated with an RL approach. Measures of engagement will include time spent using the app. ‘Behavioural engagement’ will be quantified through log file analysis of participants’ app-usage meta-data. ‘Physiological engagement’ will be quantified based on biometrics during exercise, such as participants’ average heart rate during each workout and the total number of calories they burn.
Methods and analysis
The Standard Protocol Items: Recommendations for Interventional Trials checklist was used when devising this protocol.34 This study is a randomised, controlled, single-centre crossover trial with three conditions and a primary endpoint of user satisfaction after each exercise session. The trial period will be 12-weeks and exercise prescription will be achieved using a smartphone app. The order in which participants complete each of the three conditions (condition #1; condition #2; condition #3) will be determined using block randomisation with a 1:1:1 allocation. Participants will be randomised automatically through our secure server during installation of the app, following the provision of informed consent.
The exercise sessions that are generated by the app will be altered according to each condition (condition #1; condition #2; condition #3) on a weekly basis; each participant will be administered condition-generated exercise sessions for 1-week. Each of these 1-week cycles will be comprised of three workouts of approximately 30-minutes in duration, containing 10 exercises. The only difference between the workouts within each condition will be that the specific exercises that are recommended; all other parameters of exercise will be consistent. After a 1-week cycle has elapsed, each cluster of participants completing the intervention under each of the three conditions will crossover to an opposing condition, with each crossover marking the start of a new 1-week cycle, regardless of whether participants actually completed the sessions within that cluster.[CD1] This design is illustrated in Figure X.
[Insert Figure X here.]
In summary, all participants will be subjected to weekly cycles of the three app conditions, cycling back-and-forth for a total of four weeks for each of condition #1; condition #2; condition #3-generated exercise sessions over a 12-week trial period. This design will optimise the efficiency of the crossover design to evaluate whether the sessions generated retrospectively based on participants’ satisfaction can be leveraged to prospectively increase participants’ satisfaction,35 with the added advantage of maximising the amount of data that is captured for each condition should there be a high number of dropouts or lowering levels of adherence as the trial progresses.
Recreationally active males and females who fulfil the study’s inclusion criteria will be recruited. Inclusion criteria are that volunteers should be healthy, recreationally active men and women between 18 and 65 years of age. For the purposes of the experiment, ‘recreationally active’ will be defined based on participants’ self-reporting that they engage in less than or equal to twice a week of aerobic activity for a total of 80 minutes at moderate intensity.36 We will exclude individuals with an inability to exercise due to physical disability or motor impairment, who have severe cognitive impairment or an inability to read and write in English.
Prospective participants will be recruited from the catchment area population of University College Dublin (Dublin, Ireland). Recruitment will be achieved through word of mouth and outreach on social media whereby the nature of the research (including its aims, objectives and the experimental procedure) will be communicated informally to prospective recruits on platforms like twitter, Facebook and Instagram. Prospective recruits who express an interest in participating will be screened and subsequently familiarised with the smartphone app.
Upon relaying their interesting in involvement, prospective participants will be invited into our research centre in University College Dublin for a baseline study visit and to obtain informed consent. Following this, participants will be guided in providing researchers with the International Mobile Equipment Identity (IMEI) for their smartphone device, which will be required to install the smartphone app. Once installed, participants will be guided through account registration and setup.
Following this, participants will be asked to complete a series of baseline survey measures relating to their demographics (age, sex etc) and information about current mobile technology familiarity and utilisation. These are detailed below. Thereafter, participants will be automatically randomised by our secure server (using a block randomisation with a block size of 3) into the order they will complete conditions #1, #2, #3 over each of the next 12-weeks.
During the first week of enrolment (the familiarisation week), participants will be informed of the features of the app and how the workouts that are recommended to them will vary; they will be permitted to discuss this with investigators during the course of the study. They will be informed that if they experience any difficulties using the app that they can communicate this with the research assistants, who will then liaise with the research team and software developers to address any errors. If exercise data are not being collected at the projected levels, research assistants may need to contact participants to ensure that the app is working and that they (the participant) understands how to use it. These steps will make it unfeasible to blind the researchers.
During an initial feasibility trial, members of the public worked with us to evaluate the sessions generated by the RL system, the mobile app user interface and were asked to assess the burden and time commitment of the study as part of a user-centred design approach. During this feasibility study, participants also completed an online survey to establish their current activity levels, their experience with difference health and fitness smartphone applications, how they currently exercise and their self-efficacy. Participants were then asked to use the app over an 8-week period and were asked to provide feedback on their experience. Analysis of participants’ feedback and the app usage logs was undertaken by the project team and the app was further iterated based on this feedback. The version of the app that will be used in the current study has integrated all participants’ feedback from this initial feasibility trial.
Our primary outcome, user satisfaction, will be determined after each workout and will be collected via the smartphone app. Satisfaction will be determined using an abbreviated 8-item version of the Physical Activity Enjoyment Scale (PACES).32 33 In addition, usability will be assessed with the Systems Usability Scale37 and open-ended qualitative questions about their opinions of the app.
Our secondary outcomes will relate to different aspects of user engagement. ‘Behavioural engagement’ will be quantified through log file analysis of participants’ app-usage meta-data, including the total time they spend in the app, the number of time- or user-dependent ‘hits’ or sessions on the app 38 39, the number of user inputs 40 41 and the number of click-throughs from one section of the app to another 40, for example from viewing an exercise to then completing a live demonstration of that exercise. ‘Physiological engagement’ will be quantified based on participants’ average heart rate during each workout and the total number of calories they burn. These will be derived from a Polar H9 HR sensor (Polar Electro Oy, Kempele, Finland), which we will ask participants to wear around their chest during each exercise session (each participant will be given a H9 sensor upon enrolment in the study). Physiological engagement will also be measured using a self-report rating of perceived exertion. Perceived exertion will be measured using the Borg scale and will be administered via the smart-phone app at the end of each session; participants will be cued to ‘rate your perceived exertion on a scale from 2 to 10, where 2 means “really easy” and 10 means “maximal exertion”’. See Table X for all outcomes and questionnaires.
[Insert Table X here.]
During each exercise session
After each exercise session
Physical Activity Enjoyment Scale-8
System Usability Scale
Total calories burned
Rating of perceived exertion
Total time spent on app
Number of sessions completed
Demographics (e.g., age, sex)
Mobile Phone Affinity Scale
International Physical Activity Questionnaire (IPAQ)
‘180BPM’ is a health and fitness smartphone application that uses machine learning to guide users through a catalogue of calisthenic exercises. Sessions delivered via 180BPM are comprised of a selection of from a catalogue of 291 calisthenic exercises. These exercises are differentially combined to create an X-item or XX-minute exercise session (a ‘workout’). Each workout is designed to be unique to the user, whereby the app leverages either condition #1or condition #2 built on a battery of user engagement metrics to inform exercise variation and/or progression within and between workouts, or condition #3 which generates sessions randomly. A selection of app screenshots are available in Figure X.
[Insert Figure X here.]
Reinforcement Learning algorithm
Will need this section to be written. Example here: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC7220419/#!po=20.4545
Statistical analysis plan[CD2]
Descriptive statistics will be used to summarise user demographics and scores on the Mobile Phone Affinity Scale.
Primary and secondary hypotheses
Analyses will be conducted on an intention-to-treat basis. We will use multiple imputation to handle missing data where appropriate (i.e., provided underlying assumptions are not violated). We will include participants in the analysis for primary and secondary outcomes that have at least 1-week of data available for each condition (i.e., 3-weeks total).
The primary outcome measure (user satisfaction) and the secondary outcomes measures (relating to behavioural and physiological engagement) will be evaluated at the session-level. Specifically, the unit of analysis for the outcomes that are measured during and after each workout is not the user (each of whom might contribute multiple observations to the analysis), but rather the exercise session itself. This includes the outcomes of scores on the PACES questionnaire, average heart rate, total calories burned and ratings of perceived exertion. Similarly, for outcomes that are measured weekly, including measures of behavioural engagement (total time spent on the app and number of sessions completed) and scores on the International Physical Activity Questionnaire, the unit of analysis is the condition.
For session-based outcomes that are measured during or after each session, Generalized estimating equations (GEE) will be used to assess differences in user satisfaction for each condition (condition #1; condition #2; condition #3) with average heart rate, total calories burned and participants’ perceived exertion included as covariates using an exchangeable correlation structure.
For condition-based outcomes that are measured on a weekly basis, a separate GEE will be used to assess differences for each condition (condition #1; condition #2; condition #3) in scores on the International Physical Activity Questionnaire with total time spent on the app and number of sessions completed included as covariates using an exchangeable correlation structure.
Each model will be corrected for dependent observations by including participants’ identifying code as a subject effect. The a priori p value for each analysis will be set at p < 0.05. All statistical analyses will be performed using the Statistical Package for the Social Sciences (SPSS) (IBM SPSS Statistics for Windows, version 24.0. Armonk, NY: IBM Corp.).
Planned subgroup analyses
We will conduct an exploratory subgroup analyses for our primary and secondary hypotheses for participants who exhibit higher levels of adherence with the intervention. Subgroup analyses will also be conducted for the subgroups of gender (males vs females), based on mobile phone affinity (as mobile phone use may influence the receptivity to, and ultimately the efficacy of, the intervention) and the physical activity levels at baseline.
Power calculations and sample size
A sample size of minimum 40 participants has been estimated based on the following factors; the sample size recruited as part of the initial feasibility study (we previously recruited 36 participants over an 8-week trial period), an estimated mean difference of 8 on the PACES-8 scale based on previous work for 80% power; the rule of thumb of at least 10 events for variable (or measures of user satisfaction, behavioural and physiological engagement).42 Controlling for 15% drop-out, we aim to recruit a total of 42 participants.
Patients will receive a compensation of a €50 gift voucher for participation for completion of the 12-weeks of follow-up, having completed at least one session each week.
We will submit study results for publication in peer-reviewed journals and disseminate our findings at international conferences. We will attempt to publish all findings in open-access journals where possible,. Curated technical appendices, statistical code and anonymised data will be made freely available from the corresponding authors on request using a data sharing platform (OSF.io).
Ethics and dissemination
The UCD Human Research Ethics Committee has approved this protocol (LS-XX-XX). The informed consent form for this study can be found in the online supplementary file.