MyBehavior: Persuasion by Adapting to User Behavior and User Preference

MyBehavior is a mobile application with a suggestion engine that learns a user's physical activity and dietary behavior, and provides finely-tuned personalized suggestions. To our knowledge, MyBehavior is the first smartphone app to provide personalized health suggestions automatically, going beyond commonly used one-size-fits-all prescriptive approaches, or tailored interventions from health-care professionals. MyBehavior uses an online multi-armed bandit model to automatically generate context-sensitive and personalized activity/food suggestions by learning the user's actual behavior. The app continually adapts its suggestions by exploiting the most frequent healthy behaviors, while sometimes exploring non-frequent behaviors, in order to maximize the user's chance of reaching a health goal (e.g. weight loss).

A dramatic rise in self-tracking applications for smartphones has occurred recently. Rich user interfaces make manual logging of users' behavior easier and more pleasant; sensors make tracking effortless. To date, however, feedback technologies have been limited to providing counts or attractive visualization of tracked data. Human experts (health coaches) have needed to interpret the data and tailor make customized recommendations. No automated recommendation systems like Pandora, Netflix or personalized search for the web have been available to translate self-tracked data into actionable suggestions that promote healthier lifestyle without needing to involve a human interventionist. MyBehavior aims to fill this gap. It takes a deeper look into physical activity and dietary intake data and reveal patterns of both healthy and unhealthy behavior that could be leveraged for personalized feedback. Based on common patterns from a user's life, suggestions are created that ask users to continue, change or avoid existing behaviors to achieve certain fitness goals. Such an approach is different from existing literature in two important aspects: (1) suggestions are contextualized to a user's life and are built on existing user behaviors. As a result, users can act on these suggestions easily, with minimal effort and interruption to daily routines; (2) unique suggestions are created for each individual. This personalized approach differs from traditional one-size-fits-all or targeted intervention models where identical suggestions are applied for groups of similar people or the entire population.

Control group participants received suggestions generated by the a nutritionist and exercise trainer. These suggestions didn't relate to user's life or their past behavior.

Experiment group participants received personalized suggestions from MyBehavior that relates their life and past behavior.

The intervention automatically provides personalized suggestions based on users behavior and user context. Suggestions relates to users life and how often they have done them in the past. Since the suggestions relate to users' lives, they are easy to follow.

A nutritionist and an exercise trainer jointly created 45 food and exercise suggestions based on guidelines posted by the NIH. These suggestions ask users to walk for 30 minutes or eat healthier foods. These suggestions however doesn't personalize to users daily behavior into account.

The primary outcome is to measure efficacy of MyBehavior suggestions. Efficacy will be measured in two dimensions (1) whether users intend to follow the automated suggestions from MyBehavior (2) effectiveness of automated suggestions in actual behavior change. User intentions towards following MyBehavior suggestions are measured using a 5 point likert scale. The investigators will ask users to rate whether they can follow the suggestions on an average day within a scale of 1-5 (1- I can't follow the suggestion, 5 - I can easily follow the suggestion). On the other hand, behavior change is measured from food (calories in per meal consumed) and activity (walking, running or exercise durations per day etc.) log collected using their smartphone. Regarding physical activity, how much physical activity users are performing will be compared across experiment conditions. Similarly, calorie consumption change in food will be used to compare dietary behavior change.

MyBehavior is the first system to provide health suggestions for food and activity automatically. Thus there are scopes of usability improvement on how to effectively present the automatically generated information to the user. Qualitative interviews at the end of study will be conducted to gather user experience of using MyBehavior. This interviews will help to build a better and more usable version of MyBehavior for future larger scale deployments.

Inclusion Criteria: In relatively healthy condition. Also, users must be interested in health and fitness. Exclusion Criteria: Individuals with physical disability and dietary problems are excluded.

Rabbi M, Pfammatter A, Zhang M, Spring B, Choudhury T. Automated personalized feedback for physical activity and dietary behavior change with mobile phones: a randomized controlled trial on adults. JMIR Mhealth Uhealth. 2015 May 14;3(2):e42. doi: 10.2196/mhealth.4160.