This is an open-access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/2.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work, first published in JMIR mhealth and uhealth, is properly cited. The complete bibliographic information, a link to the original publication on http://mhealth.jmir.org/, as well as this copyright and license information must be included.
A dramatic rise in health-tracking apps for mobile phones has occurred recently. Rich user interfaces make manual logging of users’ behaviors easier and more pleasant, and sensors make tracking effortless. To date, however, feedback technologies have been limited to providing overall statistics, attractive visualization of tracked data, or simple tailoring based on age, gender, and overall calorie or activity information. There are a lack of systems that can perform automated translation of behavioral data into specific actionable suggestions that promote healthier lifestyle without any human involvement.
MyBehavior, a mobile phone app, was designed to process tracked physical activity and eating behavior data in order to provide personalized, actionable, low-effort suggestions that are contextualized to the user’s environment and previous behavior. This study investigated the technical feasibility of implementing an automated feedback system, the impact of the suggestions on user physical activity and eating behavior, and user perceptions of the automatically generated suggestions.
MyBehavior was designed to (1) use a combination of automatic and manual logging to track physical activity (eg, walking, running, gym), user location, and food, (2) automatically analyze activity and food logs to identify frequent and nonfrequent behaviors, and (3) use a standard machine-learning, decision-making algorithm, called multi-armed bandit (MAB), to generate personalized suggestions that ask users to either continue, avoid, or make small changes to existing behaviors to help users reach behavioral goals. We enrolled 17 participants, all motivated to self-monitor and improve their fitness, in a pilot study of MyBehavior. In a randomized two-group trial, investigators randomly assigned participants to receive either MyBehavior’s personalized suggestions (n=9) or nonpersonalized suggestions (n=8), created by professionals, from a mobile phone app over 3 weeks. Daily activity level and dietary intake was monitored from logged data. At the end of the study, an in-person survey was conducted that asked users to subjectively rate their intention to follow MyBehavior suggestions.
In qualitative daily diary, interview, and survey data, users reported MyBehavior suggestions to be highly actionable and stated that they intended to follow the suggestions. MyBehavior users walked significantly more than the control group over the 3 weeks of the study (
MyBehavior is a simple-to-use mobile phone app with preliminary evidence of efficacy. To the best of our knowledge, MyBehavior represents the first attempt to create personalized, contextualized, actionable suggestions automatically from self-tracked information (ie, manual food logging and automatic tracking of activity). Lessons learned about the difficulty of manual logging and usability concerns, as well as future directions, are discussed.
ClinicalTrials.gov NCT02359981; https://clinicaltrials.gov/ct2/show/NCT02359981 (Archived by WebCite at http://www.webcitation.org/6YCeoN8nv).
In 2010, the World Health Organization (WHO) attributed 63% of deaths to noncommunicable diseases that are largely preventable [
A rapid rise has occurred in the development of mobile phone apps and wearable devices to address diet and physical activity. While empirical data is lacking for some commercial apps and sensor-based technologies [
Although these methods show promise, they continue to fall short by not providing context-specific, relevant, personalized help at the moment when the individual needs it to make healthier choices. The science of how to present daily physical activity and dietary intake data back to users also has been at a suboptimal state. To date, feedback has been limited to one of three categories: (1) overall numeric summaries [
Given these observations, MyBehavior was created to address some shortcomings of current mHealth interventions. MyBehavior uses a machine-learning model—multi-armed bandit (MAB)—to automatically create contextualized and personalized suggestions based on the individual’s physical activity and dietary intake data collected solely from a mobile phone. Moreover, MyBehavior is one of the very few mHealth apps designed on the basis of established behavioral theory. As such, the system reflects and incorporates the contemporary state of the behavioral science knowledge about how to foster healthful change. Based on effective behavior change principles, MyBehavior provides low-effort suggestions that request small changes to users’ existing repeated behaviors. To the best of our knowledge, MyBehavior is the first mHealth app that encourages healthy behavior change by automatically providing low-effort suggestions based on the user’s context and personal information.
The objective of this study was to evaluate a new behavior change technology—MyBehavior—using a mixed-method approach as suggested by others [
The outcomes from this study will be used to further refine the features and messages of MyBehavior to optimize its effect on physical activity and dietary intake.
To evaluate the feasibility of MyBehavior, a small 3-week, two-group randomized control trial (RCT) was conducted. The team that supervised the trial included the builders of the MyBehavior app and authors of this paper. This team recruited participants through advertisements placed around the Cornell University campus. In the advertisement, we invited participants to test a new mobile app to help them stay on track for physical activity and food intake. Recruitment was restricted to participants who owned an Android mobile phone and had an interest in fitness. Prior to the study, the investigators arranged face-to-face meetings with the participants and acquired their informed consent. Participants also completed a brief survey to provide demographic data and information about their prior experience with mobile technologies and weight loss/fitness apps. All participants attended a training session, where they installed MyBehavior on their primary mobile phone and received basic instructions, including how to enter their gender, height, and weight and how to set up a weekly weight goal (ie, lose weight, maintain weight, or gain weight). During the first week, users received a daily summary of their activities and food intake. This baseline week was intended to resemble many modern mobile health apps [
After the first week, the experimenters conducted an in-depth, semistructured interview with participants about their experience to date and then randomized participants into control and experimental groups. A random number generator was used for randomization. Assignment was single blind, as the study participants did not know their condition, while experimenters had full knowledge about the assignments.
We provided MyBehavior’s personalized context-sensitive suggestions to the experimental group, while the control group received generic prescriptive recommendations generated from a pool of 42 suggestions for healthy living, such as “walk for 30 minutes” and “eat fish for dinner.” A certified fitness professional created these generic suggestions after following National Institutes of Health resources [
This study was approved by Cornell University Institutional Review Board (1302003617) and a protocol was registered retrospectively at ClinicalTrails.gov (NCT02359981).
We recruited 18 participants, 17 of whom completed the study. Of the 17 participants, there were 13 students (76%), 4 professionals (24%), 8 females (47%), and 9 males (53%), and all were between the ages of 18 and 49 (mean 28.3, SD 6.96, lower quartile [q25]=22, median [q50]=26.3, upper quartile [q75]=36). All participants reported low-to-moderate levels of physical activity. The majority of participants were experienced mobile phone users—9 participants (53%) had previous experience using a food diary, and 6 participants (35%) had previously kept an exercise log. After the randomization, participants in the groups were similar in terms of level of active lifestyle and experience with using mobile-based self-management tools. Our sample size was determined based on earlier literature [
Flow of participants in the MyBehavior trial.
MyBehavior is comprised of five key modules: (1) physical activity tracking, (2) food logging, (3) life-log generation, (4) physical activity and food clustering, and (5) suggestion generation.
MyBehavior uses the accelerometer and the Global Positioning System (GPS) sensor inside the mobile phone to continuously keep track of an individual’s physical activities. A number of statistical features (eg, mean, variance, zero-crossing rate) are extracted from the sensor data and a machine-learning model—Gaussian Mixture Model (GMM) [
Users select food items from a database and enter the consumed quantity to get the corresponding calorie intake. The United States Department of Agriculture (USDA) [
MyBehavior provides several features to make the food logging experience easier for users. First, users can take photos of their food. These photos serve as a memory aid when users are prompted to input their food information at 9:30pm every night (see
MyBehavior app screenshots: (a) taking photo of a food item, (b) searching for foods from the USDA database, and (c) Life Log, a chronological list of activity and food log events.
MyBehavior generates as a “life log,” a chronological list of activity and food log events, as shown in
In the first stage, activity predictions, which happen every 1 second, are aggregated every minute and labeled automatically. In the second stage, the contiguous activities having the same label are combined into a single entry. For example, if a user is stationary for 50 minutes, MyBehavior will generate a “stationary” activity entry into the life log with a 50-minute duration. Other common life events include a sequence of different activities that happen within a short time interval (ie, 15 minutes). An example might be the following: walk to the bus stop, wait for a few minutes, ride the bus, and walk to the office after exiting the bus. In this example, MyBehavior generates a “mixed” activity entry in the life log (eg, taking the bus from home to work) by combining the multiple activity sequences that happen within a 15-minute window.
To enable suggestion generation delivery to the experimental group only, MyBehavior used the life logs to cluster similar physical activities and similar food items. The food similarity matching process follows a simple logic—food is clustered based on similar food ingredients. For example, MyBehavior will detect if a user is repeatedly having high-calorie burgers with similar ingredients and form the cluster “burger” that groups together the same or similar types of burgers.
Regarding clustering physical activities, manually tracked activities are clustered based on the type of activities similar to food clustering. Automatically tracked activities, tagged with location information, are clustered by places they occur. Clusters are found using unsupervised machine-learning techniques to identify similarity. As indoor localization is often accurate up to 150 meters, any stationary activities that fall within 150 meters of each other are clustered together. For example, a user’s stationary activities in the office are typically in close proximity to each other and, as such, MyBehavior clusters these office locations into a single unit that represents the user’s stationary behavior in the office. Walking and running activities are more difficult to cluster because MyBehavior needs to determine whether two activity trajectories look similar and happen at a similar location. To group similar walking or running events, MyBehavior uses an algorithm derived from the literature on handwriting recognition [
Clusters generated from user activities: (a) locations where user A stayed stationary, (b) location traces for user B where he walked around his office, and (c) walking traces of user A from his house to the bus stop.
After clustering user behaviors, MyBehavior uses an
MyBehavior
For stationary activities, the recommender engine suggests small changes, such as walking 3 minutes for every hour spent stationary. The right-most image in
When generating food suggestions, a separate set of suggestions is created based on the exploit-explore strategy. First, MyBehavior distinguishes between meals and snacks. Then it takes into account both intake frequency and calories similar to the physical activity suggestions. Thus, a user’s frequent healthy low-calorie meals are exploited and are encouraged to be continued. During exploration, a random selection of infrequent low-calorie meals/snacks from the past is suggested. Here, the expectation is that users will take up some of these infrequent meals and make them frequent in the future.
At the start of every day, MyBehavior generates 10 food and 10 activity suggestions. Of these, 90% are from the users’ most frequent activities (ie, exploit) and 10% are from the users’ infrequent behaviors (ie, explore). This split of 90% exploit and 10% explore was heuristically chosen based on previous literature [
Screenshots showing recommended suggestions for exercise and food: (a) physical activity suggestions made by MyBehavior, (b) food suggestions made by MyBehavior, (c) physical activity suggestions for a different user, and (d) physical activity suggestions for the same user as in (a), but at a different point in time.
First, we used a suggestion-rating survey to evaluate user
The daily diary and the in-depth, semistructured interviews measured participant feedback regarding the suggestions. For the daily diaries, we queried (1) whether they looked at MyBehavior’s suggestions, and (2) whether they made or wanted to make any changes after seeing the suggestions. The semistructured interviews covered users’ general overall experience with MyBehavior and the quality of the suggestions. Specifically, we inquired about awareness, behavior change, and of any software improvement they would like to see. In addition, in the interview, we asked clarifying questions that explained quantitative results observed from the data.
Regarding the user’s intention of following MyBehavior’s suggestions, we gathered ratings for suggestions on a secure website and analyzed the data using RStudio. Since the ratings were in ordinal scale, we used a nonparametric Mann-Whitney U test [
We measured behavior changes by analyzing activity and dietary logs for statistical significance using MATLAB (MathWorks, Inc) statistics toolbox and RStudio. For each user, we computed median walking length and calories per food item. We considered medians across entire weeks over other central measures since they are less susceptible to spurious noise or outliers (eg, occasional intake of very-high-calorie food or atypical, unusually lengthy walk). We did not report changes in running and manually logged exercises in the data analysis as they often require higher effort and are tough to change within the 3 weeks of the experiment. In our analysis, we first considered the number of positive changes. A
Finally, face-to-face, semistructured interviews were audio recorded and transcribed. Interview transcripts and daily diaries were then broken down into themes using thematic analysis [
A total of 17 participants completed the 3-week study, yielding almost 2.1 million recorded physical activity instances, amounting to more than 8000 hours of physical activity. During the same period, participants labeled nearly 850 images of food with annotations.
In the suggestion-rating survey, the experimental group (mean 3.4, SD 1.2, q25=2.75, q50=3, q75=4), with MyBehavior suggestions, intended to follow personalized suggestions more than the control group (mean 2.5, SD 1.6, q25=1, q50=2, q75=4) intended to follow the generic suggestions. A nonparametric Mann-Whitney U test [
For walking, 78% (7/9) of participants in the experimental group (
Qualitative data from daily diary and face-to-face interviews largely supported this quantitative result. However, we also observed some important subtleties. First, participants in the experimental group described the activity suggestions to be actionable and relevant to their lives. Control group participants appreciated that the generic suggestions reminded them of good habits. However, they often faced problems incorporating the suggestions into their daily lives. The following quotes were taken from the daily diaries of participants.
Those suggestions are quite good, which reminds me not to sit too long in one place.
The exercise suggestions made me want to do some more activities and be less stationary. Seeing how long I have been stationary and the low frequency of activity made me want to make a change.
Try to get up from my desk more often...added “walk" notes to my calendar.
I did some walking where I normally walk. The app now shows I walked there 26 times. The app makes me feel that I can do it again since I have done the same walk many times.
The suggestions encourage me to do/plan exercises for the near future...It reminds me that some foods are better than others.
They seem like good generic suggestions. The kind you would read...as tips in a health magazine or some such...
Some MyBehavior users reported that even the nonfrequent explore suggestions were actionable and expressed interest in acting on them. For instance, experimental group participant #7 said the following in his/her daily diary:
I saw a walk to my nearest bus stand listed. Normally, I drive my car to go to my office. But looking at the extra walking I got while going to the bus stop makes me think about doing it often and making it a habit.
Results from interviews also revealed that participants at various stages of active lifestyle reacted to suggestions differently [
Finally, on a few occasions, MyBehavior suggestions were hard to follow or did not reflect user preferences. For example, one user reported in the interview that he used to play soccer with his friends but his friends recently moved to a new location. He could no longer play soccer, which MyBehavior was suggesting. In addition, often user-preferred activities are not top MyBehavior suggestions. For instance, one user preferred to swim even though she did not do it often. Finally, experimental group participant #8 (subject 8 in
Box plots showing the distribution of walking lengths for the experimental group (a) and for the control group (b) over the 3-week study. We joined the medians of distributions and showed each trend as a thick green line (increasing trend) or red line (decreasing trend) for walking length.
For caloric intake, 78% (7/9) of participants in the experimental group showed positive trends (green lines in
In qualitative feedback, similar to physical activity suggestions, experimental group users found the suggestions to be more actionable and reported to make more changes compared to control group users who found the suggestions to be hard to work on. This feedback is illustrated in the following quotes from participants’ daily diaries.
The pictures of my meals are very useful to keep track of what I've been eating in the past. People tend to forget about their habits, but pictures in this case are a nice way to bring your food history in front of your eyes.
The suggestions remind me that some foods are better than others.
It recommends me to eat stuff that I don't have at home.
These suggestions don't take into account my dietary restrictions.
Similar to activity explore suggestions, MyBehavior users often found the explore suggestions to be actionable.
I just wanted to see what it was...These ones [explore suggestions] seemed to pick up some "good" food habits.
Finally, users reported manual food logging to be time consuming in the interview. However, they also reported that this manual process made them more aware of their foods. Consequently, control group participants reported making dietary changes without personalized suggestions.
Box plots showing the distribution of food calories for the experimental group (a) and for the control group (b) over the 3-week study. We joined the medians of distributions and showed each trend as a thick green line (increasing trend) or red line (decreasing trend) for median food calorie intake.
To our knowledge, MyBehavior is the first system to automatically provide personalized suggestions that relate to users’ lifestyles. In the quantitative results, MyBehavior users demonstrated superior behavior changes compared to the control group. Qualitative measures from the face-to-face interviews and the daily diaries confirmed that the suggestions indeed were perceived to be personalized to their lives. This concordance of superiority in both quantitative behavior change and qualitative user perception makes MyBehavior’s automated health feedback approach very promising and provides support for longitudinal studies and future investigations into automated personalization approaches.
Specifically, in our evaluation, users rated that they could follow MyBehavior personalized suggestions more than the control condition suggestions. Results also revealed a significant change in walking behaviors for MyBehavior users. In qualitative measures, users reported MyBehavior activity suggestions to be more actionable. Interestingly, although users qualitatively reported the dietary suggestions to be more actionable, dietary behavior changes were not found to be different between the groups. This finding could be due to the manual-logging nature of food intake being sufficient for behavior change alone. The manual process of food logging might produce self-awareness and reflection. Indeed, past research demonstrates that simple logging can improve one’s food consumption behavior [
Nonetheless, MyBehavior explores a unique space for health feedback. Earlier studies in this domain predominantly focused on overall behavior [
Despite this promising direction, the automated data-driven personalization approach of MyBehavior brings its own challenges. Manual logging of food and exercise, in addition to automated logging, are necessary for proper functioning of MyBehavior. Qualitative interviews revealed that manual food and exercise logging were often burdensome. Future iterations could use crowdsourcing-based semiautomated approaches to decrease the burden of manual food journaling [
An important limitation is the short-term and small-scale nature of the study, which makes it difficult to make definitive conclusions. However, the study helped us to identify the potential efficacy of MyBehavior and pinpoint design improvements for future deployments. Indeed, Klasanja et al [
Despite these limitations, this pilot study demonstrates the potential of using automated personalization for actionable health feedback. As we move into an age where increasingly more people are tracking their health with mobile and other technologies, we believe MyBehavior’s automated technique holds great potential to provide feedback that can be used to improve health outcomes at scale.
MyBehavior is the first mobile health app that can encourage healthy behavior change by automatically providing low-effort suggestions based on the context and personal information of users. The pilot user study demonstrated the feasibility and acceptability of MyBehavior. Users considered MyBehavior’s personalized, contextualized suggestions to be more actionable and to require less effort to implement than generic prescriptive suggestions. Preliminary evidence of behavior change shows that a high percentage of MyBehavior users did more physical exercise, yet the potential impact on eating behaviors remains unclear. The addition of more human control over the suggestions and providing easier logging mechanisms for food and exercise were identified as key areas of improvement. Future directions should include addressing the identified shortcomings of the system and testing its effectiveness in fostering health behavior change in a larger longitudinal trial.
Generic suggestions used in the control group.
A video showing how MyBehavior works.
CONSORT-EHEALTH checklist V1.6.1 [
Centers for Disease Control and Prevention
Fogg Behavior Model
Gaussian Mixture Model
Global Positioning System
multi-armed bandit
Metabolic Equivalents of Task
lower quartile
median
upper quartile
randomized control trial
United States Department of Agriculture
World Health Organization
We acknowledge Emiliano Miluzzo, Erin Carroll, Fabian Okeke, Jean Costa, Steven Voida, and Chantelle Farmer for their early contributions to this research. This work was supported by Intel Science and Technology Center for Pervasive Computing.
None declared.