Study Overview The PERFORM-RL study was a 12-week randomised crossover trial conducted to evaluate the impact of app-based personalised exercise prescription on user satisfaction and exercise intensity, using reinforcement learning (RL) as the key mechanism for customisation. This study sought to determine whether RL-generated exercise sessions could improve user satisfaction and increase exercise intensity compared to non-personalised, generic sessions. The findings are expected to contribute to the growing field of digital health interventions, particularly in the area of scalable, app-based exercise programmes designed to enhance user engagement and long-term adherence.
Study Objectives The primary objective of the study was to compare user satisfaction between RL-generated exercise sessions and generic sessions. The hypothesis was that RL-driven personalisation would result in higher user satisfaction, as measured by the Physical Activity Enjoyment Scale (PACES-8). The secondary objective was to evaluate the effect of RL on exercise intensity, as measured by the Borg Rating of Perceived Exertion (RPE) scale. It was hypothesised that participants would exercise at a higher intensity in the RL condition due to the tailored nature of the sessions, which aligned with their preferences and fitness levels.
Study Design
This was a randomised, assessor-blinded crossover trial. Each participant completed two different conditions during the 12-week study:
Intervention Condition: Exercise sessions were personalised using a reinforcement learning model that adapted session difficulty and content based on participant preferences, past performance, and feedback.
Control Condition: Participants completed generic, non-personalised exercise sessions. These sessions were pre-designed and did not adapt based on the user's feedback or preferences.
Participants alternated weekly between the intervention and control conditions, meaning that each week they experienced a different approach to exercise prescription. This design ensured that each participant acted as their own control, reducing variability and increasing the robustness of the findings. All exercise sessions were delivered through the i80 BPM app, developed by Samsung, which also facilitated data collection.
Participants Participants were recruited from Dublin, Ireland, and its surrounding areas via word of mouth and social media. A total of 62 participants (27 males, 42 females; mean age 42 years) completed at least one exercise session, with 559 sessions completed overall during the 12-week trial period. Participants were healthy, recreationally active adults aged 18 to 65 years. They were screened for eligibility using the "Exercise Preparticipation Health Screening Questionnaire for Exercise Professionals." Exclusion criteria included physical disability, severe cognitive impairment, or an inability to read and write in English.
Interventions Reinforcement Learning (RL) Intervention
The RL model employed by the i80 BPM app personalised exercise sessions based on user preferences, perceived exertion, and feedback from previous sessions. The RL framework used a decision-making agent, which selected exercises from a pre-existing database based on the user's fitness level, goals, and real-time feedback. The model's reward function aimed to maximise user satisfaction and session effectiveness by adapting the exercise difficulty and content in response to the user's evolving needs.
For example, if a participant reported high satisfaction and moderate intensity in a previous session, the RL model would suggest similar or slightly more challenging exercises for the next session. The RL system continuously learned from user interactions, adjusting the exercise prescription to maintain optimal engagement and effectiveness.
Control Condition
In the control condition, participants received generic exercise sessions that were not personalised or adapted based on user feedback. These sessions included a fixed selection of exercises and did not evolve over time, providing a baseline against which the personalised RL intervention could be compared.
Outcome Measures Primary Outcome: User Satisfaction
User satisfaction was assessed after each session using an abbreviated 8-item version of the Physical Activity Enjoyment Scale (PACES-8). This validated scale measures the extent to which participants enjoyed their exercise sessions. Scores range from 1 (low satisfaction) to 5 (high satisfaction), with higher scores indicating greater enjoyment and overall satisfaction with the exercise experience.
Secondary Outcome: Exercise Intensity
Exercise intensity was measured using the Borg Rating of Perceived Exertion (RPE) scale, which ranges from 1 (very easy) to 10 (maximum exertion). Participants were prompted to rate their perceived exertion after each session via the app, ensuring real-time capture of their experience. Heart rate data, collected using a Samsung Galaxy Fit 2 smartwatch, served as an additional measure of intensity, validating the subjective RPE scores.
Additional Outcome: Heart Rate
Heart rate was continuously monitored during each session using the Samsung Galaxy Fit 2 smartwatch. The heart rate data was automatically relayed to the i80 BPM app and used to track physiological responses to the exercise sessions.
Data Collection and Management Data were collected through the i80 BPM app, which recorded user satisfaction, exercise intensity, heart rate, and the duration of each exercise session. To ensure participant confidentiality, all data were anonymised before analysis. Participants were assigned a unique identification code, and all sensitive documents were stored securely, with only authorised personnel granted access.
Sample Size and Power Calculations Based on a previous feasibility study, the sample size was estimated to be a minimum of 40 participants to detect a mean difference of 8 points on the PACES-8 scale with 80% power. A total of 69 participants were recruited to account for potential dropouts, and 62 participants completed at least one exercise session. The final sample size provided sufficient power to detect differences between the RL and control conditions.
Statistical Analysis Generalised Estimating Equations (GEE) were used to analyse the primary and secondary outcomes. GEE models were applied to account for the correlation of repeated measures within subjects. The dependent variables in the models were user satisfaction (PACES-8 scores) and perceived exertion (Borg scale scores), while the independent variables were condition (RL vs control) and trial week. Covariates included age, gender, and baseline physical activity levels, with participant ID included as a subject effect.
