Paper Summary - MyBehavior: Automatic Personalized Health Feedback from User Behaviors and Preferences using Smartphones

Original paper: https://dl.acm.org/doi/10.1145/2750858.2805840

As part of the coursework for Advanced Artificial Intelligence Masters course at United International University, I wrote a summary of a paper titled "MyBehavior: Automatic Personalized Health Feedback from User Behaviors and Preferences using Smartphones". With minor modifications, I am sharing the work, with the hope that some readers will find this useful. Enjoy!

In this paper, the authors build upon their previous work which they had termed as MyBehavior 1.0 [1]. It is a smartphone application that utilizes recommendation algorithms, paired with behavioral analysis, to provide personalized exercise and dietary suggestions to users for improving their health. The Multi Armed Bandit (MAB) algorithm has been employed to generate lifestyle suggestions related to the users’ health. The suggestions for changes in lifestyle are based on automatically learned factors like activity and diet patterns of the users. The goal is the maximization of calorie loss. Easy user adoption has also been considered by the authors. Furthermore, incorporation of the preferences of the individual users has been one of the primary motivations for this iteration of the authors’ work, which is termed as MyBehavior 2.0. A 14-week study has also been conducted to test the effectiveness of the developed application.

The authors had already implemented the user behavior data collection and mining modules in their previous work [1]. Using this mined behavior data, they had also previously developed a MAB algorithm-based system which suggested a mix of frequent (exploitation) and infrequent (exploration) health actions. The MAB algorithm is very much suited for this problem since limited data means that the algorithm should not have too many parameters. It is also computationally efficient.

To prepare the behavior data, the authors classified user activities like walking, cycling, stationary, etc. and clustered groups of food as well as different exercises which had been entered manually. Moreover, all this data had been further processed to gather location-based intelligence like places where the users are stationary for long periods of time, and common walking or running routes of the users (see figure below [3]). The cluster size of behaviors is employed by the authors to correspond to the frequency of behaviors. The frequency is used alongside the average calorie benefit (by multiplying them together) to develop the objective function for their optimization algorithm. This maps well to theories in behavior change. For instance, frequency of action and their relatedness to one’s life makes it more likely that the actions would be adopted by the users.

The suggestion strategy accounts for possible changes in user’s behavior over time by including suggestions from infrequent behaviors (exploration) in addition to the exploited frequent behaviors of the user (learned from the activities and consumption data). EXP3 strategy is used with exploration-to-exploitation split of 10%-to-90%. It is also important to note that MyBehavior keeps the suggestions for foods and exercises separate. Within food consumption, meals and snacks are also considered separately.

There were two key learnings from a 3-week pilot of MyBehavior 1.0 with 9 users. First, the users found it difficult and repetitive to manually log food consumption data. Second, the application did not allow for the users to add their own preferences for specific tasks. In MyBehavior 2.0, these features were developed to counter the aforementioned limitations. Easier logging was enabled by using a crowd-sourced solution, which allowed nutrition information to be returned from photos of food captured and submitted by the users. The simplified case derived by limiting categories of food to the most frequent 40 (according to data from a popular fitness app) ensured that the crowd-sourced labeling gave calorie estimates within reasonable error bounds. The users could also select from their past exercises in this new version of the app, which further reduced the data logging burden.

Second, MyBehavior 2.0 allowed users to remove and re-order suggestions, thus giving them more control and allowing them to customize the suggestions. The suggestions made by the algorithm focus on low-effort, while the user’s customization introduces a preference component. The decision theory technique called pareto-frontier is employed to find a balance between low-effort and preference.

To test the effectiveness of the newly developed features, a 3-week pilot was first conducted. Improvements in score (N=7 users) for both of the new features were noted in this pilot study. Next, a 14-week study followed for evaluating MyBehavior 2.0’s 1) efficacy, and 2) persisting effects beyond the initial “novelty” phase. The study design followed a “single case experiments” method where repeated data are collected from single users. In the study, the users went from a period of control condition to experiment condition after 2-4 weeks. The experiment phase lasted for 7-9 weeks. 16 participants were selected based on certain eligibility criteria. The users had a fairly balanced distribution in each of the categories of Gender, Age, Stage of behavior change before the study, and Previous experience with self-tracking (see table below [3]).

A baseline phase (involving no interventions) of 3 weeks was included to move beyond the “novelty effect”, where there can be increased activity even without interventions. The control phase followed next, where common prescriptive suggestions (curated from National Institute of Health resources) were made to the participants. Finally, in the last 7-9 weeks in the experiment phase, the users received MyBehavior suggestions (see figure below [3]).

For measuring outcomes, first, food and exercise calorie changes were used. A daily 5-question phone survey was also conducted to measure outcomes (see table below [3]). Statistical significance were reported using student t-tests, and effect size using Cohen-d. The outcome measures of number of suggestions followed, number of suggestions wanted, relatedness, walking/day (min), exercise/day (cal), each meal (cal) showed significant statistical difference between the control and MyBehavior. Thus, the authors have proven high efficacy of MyBehavior using rigorous statistical analysis. Moreover, the 2-4 weeks control phase is compared against the experiment phase’s last 3 weeks, in order to report the change beyond the novelty periods. The improvement due to MyBehavior over the control condition lasted beyond the novelty phase.

The current system guarantees privacy by only using the individual users’ own data in the analysis for health suggestions in the users’ own device. This research was done in 2015. In 2016 [2], Google developed Federated Learning technology to allow distributed learning by sharing of model weights only. This allows learning to take place using more data in a privacy-preserving manner. This work can be further extended by employing such modern privacy-preserving learning techniques.

?Personal Comments and Remarks: This has been a very useful paper for me to review and summarize. This is because I currently work for a healthcare technology company that focuses on using behavior change theories to support people living with chronic health diseases. I provide data analytics support to the Clinical Team and Health Coaches in order to help them prioritize the communication with people living with chronic diseases. This is necessary as there are around 40000 people using the platform and less than 20 Health Coaches. This paper is a great step towards providing automated and personalized care digitally, and at scale. An interesting area to explore would be to build a hybrid system that complements the work of human Health Coaches.

Refer to the figure above [3]. Since this research is strongly focused on behavior change theory, maybe it is best to avoid showing tempting food images in the app, as obese people may crave the high calorie meal when seeing a photo of it in the app. Consequently, they might slip from the "acting" stage to the "ready" stage (Trans-theoretical model (TTM)).

Acknowledgements: Thanks to Dr. Swakkhar Shatabda for picking such an exciting, engaging and interesting topic for the coursework.

References:

[1] M. Rabbi, A. Pfammatter, M. Zhang, B. Spring, and T. Choudhury. Automated personalized feedback for physical activity and dietary behavior change with mobile phones: A randomized controlled trial on adults. JMIR mHealth uHealth, 3(2):e42, May 2015.

[2] J. Kone?ny, H. B. McMahan, F. X. Yu, P. Richtárik, A. T. Suresh, and D. Bacon. Federated learning: Strategies for improving communication efficiency. arXiv preprint arXiv:1610.05492, 2016.

[3] M. Rabbi et al. MyBehavior: automatic personalized health feedback from user behaviors and preferences using smartphones.?Proceedings of the 2015 ACM International Joint Conference on Pervasive and Ubiquitous Computing. 2015.