Piloting a Reinforcement Learning Tool for Individually Tailoring Just-in-time Adaptive Interventions

Reinforcement Learning (RL), a type of machine learning, holds promise for addressing the limitations of previous approaches to implementing JITAIs. Adaptive RL applications work by updating information about expected "rewards" (i.e., proximal outcomes) based on the results of sequentially randomized trials. To realize the potential of adaptive interventions to reduce health disparities in cancer prevention and control, mHealth interventionists first need to identify methods of using digital health participant data to continually adapt decision rules guiding highly tailored intervention delivery. This research team has developed a reinforcement learning model (AdaptRL) that reads in and analyzes user data (e.g., calories, weight, and activity data from Fitbit) in real-time, uses RL to efficiently determine which message a participant should receive up to 3 times per day, and creates a JITAI tailored to optimize daily behavioral goal achievement and weight loss for each participant. The objective of this study is to test the feasibility of using this reinforcement learning model in a pilot weight loss study.

Inclusion criteria

• Age 18-55 years
• Body Mass Index of 25-40 kg/m2
• English-speaking and writing
• Has a smartphone with a data and text messaging plan

Exclusion criteria

• Currently participating in a weight loss, nutrition, or physical activity study or program or other study that would interfere with this study
• Currently using prescription medications with known effects on appetite or weight (e.g., oral steroids, weight loss medications), with the exception of individuals on a stable dose of SSRIs for 3 months)
• Previous surgical procedure for weight loss or planned weight loss surgery in the next year
• Currently pregnant or planning pregnancy in the next 4 months
• Lost 10 or more pounds and kept it off in the last 6 months
• Report a heart condition, chest pain during periods of activity or rest, or loss of consciousness on the Physical Activity Readiness Questionnaire (PAR-Q; items 1-4). Individuals endorsing joint problems, prescription medication usage, or other medical conditions that could limit exercise will be required to obtain written physician consent to participate
• Pre-existing medical condition(s) that preclude adherence to an unsupervised exercise program, diabetes treated with insulin, history of heart attack or stroke, current treatment for cancer, or inability to walk for exercise
• Type 1 diabetes or currently receiving medical treatment for Type 2 diabetes
• Other health problems which may influence the ability to walk for physical activity or be associated with unintentional weight change, including cancer treatment within the past 5 years or tuberculosis
• Health or psychological diagnoses that preclude participation in a prescribed dietary and exercise program, including history of or diagnosis of schizophrenia or bipolar disorder, hospitalization for a psychiatric diagnosis in the past year, a current diagnosis of alcohol or substance abuse
• Report a past diagnosis of or receiving treatment for a DSM-5-TR eating disorder (anorexia nervosa, bulimia nervosa, or other diagnosis)
• Another member of the household is a participant or staff member in this trial
• Not willing to attend one study visit
• Not willing to wear a Fitbit every day
• Reason to suspect that the participant would not adhere to the study intervention
• Have participated in another study conducted by the UNC Weight Research Program within the past 12 months