Evaluating PEARL - a Personalized Exercise Assistant Using Reinforcement Learning (Walkmate Study)

Evaluating PEARL - a Personalized Exercise Assistant Using Reinforcement Learning (Walkmate Study)

To evaluate the impact of two personalized nudging strategies delivered as pop-up notifications via the Fitbit app on user step count. Specifically, to personalize the following parameters of the pop-up notification system: message content, and timing (hr of the day).

Phase 1 [Model calibration]: Recruit up to 1,000 Fitbit users for a 4 week pilot study. Nudge content and timing will be varied randomly in order to collect training data to prime the RL architecture prior to Phase 2.

Phase 2 [Performance evaluation]: Recruit up to 12,000 Fitbit users for a 60 day study in which they are randomized evenly between the following 4 arms:

[Control] No nudges Randomly selected nudges from the custom nudge library, delivered at constant time and frequency Behavior science (BS)-only nudge agent PEARL agent

Phase 3 [Micro-randomized Trial (MRT) & LLM Feasibility]: Recruit up to 6,000 Fitbit users for a 60 day study in which they are randomized evenly between the following arms:

Behavior science Micro-randomized Trial (BS-MRT): Once per day users will be randomized across the below factors, and receive a message written by a behavior scientist (the same messages from Phase 1 & 2).

COM-B Theme: 6 themes, and 1 control (no nudge) Time of day: 3 timeframes, and 1 control (no nudge) Large Language Model Micro-randomized Trial (LLM-MRT): Once per day users will be randomized across the below factors, and receive a message written by a large language model.

COM-B Theme: 6 themes, and 1 control (no nudge) Time of day: 3 timeframes, and 1 control (no nudge) Large Language Model + Reinforcement Learning (LLM-RL): Once per day, a reinforcement learning model will run to select the optimal COM-B theme and time of day (the same RL model from Phase 2 Arm 4). The nudges will be selected from a repository that is written by a large language model (LLM).

Primary Purpose:

Phase 1: Model calibration Phase 2: Evaluate performance on step count Phase 3: Micro-randomized Trial (MRT) & LLM feasibility

Inclusion Criteria:

• Age 22-60 years
• [For Phase 2 & 3] Daily average steps 1 month prior to recruitment is less than 8000
• Owns a Fitbit device & smartphone with Fitbit app
• Willing to share a month of step count data prior to enrollment data
• Wear a Fitbit device regularly (user-reported)

Exclusion Criteria:

• Advised by a doctor to avoid exercising or limit physical activity
• Current employees of Alphabet
• [For phase 2] Have participated in Phase 1 of the study
• [For phase 3] Have participated in Phase 1 or 2 of the study