Personalized Exercise Recommendations for Chronic Pelvic Pain Using Reinforcement Learning
Trial statusRecruiting
Trial phaseNot applicable
Trial typeInterventional
Biological sexFemale
Age18-55
SponsorIcahn School of Medicine at Mount Sinai
About this trial
WorkoutCPP is a pilot study evaluating the feasibility of a personalized exercise recommendation system for individuals with chronic pelvic pain disorders (CPPDs). The study uses reinforcement learning (RL), a type of artificial intelligence that adapts recommendations over time based on each participant's reported pain levels, symptom burden, and exercise compliance. Participants receive daily exercise recommendations that alternate between standard, non-personalized guidance and personalized, RL-generated recommendations across four 2-week phases, allowing within-person comparison of outcomes under each condition. The primary hypothesis is that an RL-based adaptive recommendation system is feasible to deliver in a CPPD population.
Eligibility criteria
This trial does not accept healthy volunteersQualifiers
Self-reported CPPD (e.g., endometriosis, adenomyosis, fibroids, etc.) based on clinician diagnosis
Aged 18-55 years.
Ownership of an iOS or Android smartphone.
Willingness to self-track daily symptoms, exercise activities, and self-management behaviors using a smartphone research app.
Disqualifiers
Absolute contraindications to PA (e.g., recent myocardial infarction, complete heart block, acute congestive heart failure, unstable angina, or uncontrolled severe hypertension, BP ≥180/110 mm Hg).
More than two "Yes" responses on the Physical Activity Readiness Questionnaire (PAR-Q) (16) without physician clearance.
Major life events expected during the next 10 weeks (e.g., pregnancy, planned surgery, or extended travel likely to interfere with participation).
Current or planned pregnancy within the next 6 months.
Trial design
Design model
Crossover
Treatments tested in this trial
Reinforcement Learning (RL)-Based Personalized Exercise Recommendations
BehavioralDaily exercise recommendations (using type, intensity, and duration) are generated by a contextual bandit reinforcement learning agent, based on the implementation described in Meier et al. 2023. Recommendations are personalized using each participant's initially generated list of exercises based on their physical ability and resources available, as well as contextual daily factors including pain symptoms, prior exercise compliance, and their feedback to the previous exercise recommendation.
Generic Exercise Recommendation
BehavioralParticipants receive exercise recommendations from a standardized, set list of exercise recommendations that are based on USDHHS physical activity guidelines (Piercy et al., 2020). Recommendations are not personalized based on participant contextual information and do not adapt over the course of the study.
Treatment groups
45 Participants
are divided into 2 treatment groupsGroup A: RL-based personalized phaseExperimental treatment 1 intervention
Group B: Standard (Generic) Exercise ArmActive comparator 1 intervention
Trial outcomes
Primary outcomes
1
Exercise Recommendation Adherence Rate
Exercise recommendation adherence rate is calculated by the proportion of daily exercise recommendations completed over the course of the intervention. A higher exercise adherence rate indicates that participants are completing their given exercise recommendations at higher frequencies.
Time frame
At 9 weeks at study completion2
Participant Retention Rate
Participant retention rate is the proportion of enrolled participants completing study participation until the end of intervention. A higher retention rate indicates that participants complete the 9-week intervention period at higher frequencies.
Time frame
At 9 weeks at study completionSecondary outcomes
1
Reinforcement Learning Agent Action Entropy Over Time
Entropy of the reinforcement learning (RL) agent's action probability distributions is calculated at each decision point throughout the study period. Entropy ranges from a minimum of 0 (the agent selects a single action with certainty) to a maximum of log(K), where K is the number of distinct recommendation actions available to the agent at that decision point (maximum entropy occurs when the agent assigns equal probability to all available actions). Decreasing entropy over time indicates increasing model confidence and more consistent personalized recommendation patterns. Higher entropy indicates greater uncertainty in the agent's decision making. This outcome will be evaluated for the entire duration of the intervention based on the daily data.
Time frame
At 9 weeks at study completionSponsors and contacts
Click on the lead sponsor to view all of their trials.
Icahn School of Medicine at Mount Sinai
Lead sponsor
Hasso Plattner Institute, Potsdam, Germany
Collaborator
Source ClinicalTrials.gov