This is an open-access article distributed under the terms of the Creative Commons Attribution License (https://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work, first published in the Journal of Medical Internet Research, is properly cited. The complete bibliographic information, a link to the original publication on http://www.jmir.org/, as well as this copyright and license information must be included.
Chronic pain is a globally prevalent condition. It is closely linked with psychological well-being, and it is often concomitant with anxiety, negative affect, and in some cases even depressive disorders. In the case of musculoskeletal chronic pain, frequent physical activity is beneficial. However, reluctance to engage in physical activity is common due to negative psychological associations (eg, fear) between movement and pain. It is known that encouragement, self-efficacy, and positive beliefs are effective to bolster physical activity. However, given that the majority of time is spent away from personnel who can give such encouragement, there is a great need for an automated ubiquitous solution.
MyBehaviorCBP is a mobile phone app that uses machine learning on sensor-based and self-reported physical activity data to find routine behaviors and automatically generate physical activity recommendations that are similar to existing behaviors. Since the recommendations are based on routine behavior, they are likely to be perceived as familiar and therefore likely to be actualized even in the presence of negative beliefs. In this paper, we report the preliminary efficacy of MyBehaviorCBP based on a pilot trial on individuals with chronic back pain.
A 5-week pilot study was conducted on people with chronic back pain (N=10). After a week long baseline period with no recommendations, participants received generic recommendations from an expert for 2 weeks, which served as the control condition. Then, in the next 2 weeks, MyBehaviorCBP recommendations were issued. An exit survey was conducted to compare acceptance toward the different forms of recommendations and map out future improvement opportunities.
In all, 90% (9/10) of participants felt positive about trying the MyBehaviorCBP recommendations, and no participant found the recommendations unhelpful. Several significant differences were observed in other outcome measures. Participants found MyBehaviorCBP recommendations easier to adopt compared to the control (
In the pilot study, MyBehaviorCBP’s automated approach was found to have positive effects. Specifically, the recommendations were actualized more, and perceived to be easier to follow. To the best of our knowledge, this is the first time an automated approach has achieved preliminary success to promote physical activity in a chronic pain context. Further studies are needed to examine MyBehaviorCBP’s efficacy on a larger cohort and over a longer period of time.
Chronic pain is defined as pain that persists despite the resolution of injury or pathology [
A particularly common form of chronic pain is of a musculoskeletal nature, which affects 1 in 10 adults globally. This form of chronic pain is also a leading cause of disability, with 28% reporting limitations in movement due to the condition [
Despite the benefits of physical activity, adherence to regular and sustained physical activity is low [
Low adherence to physical activity is further compounded by the need to self-manage. Typically, in day-to-day life settings, there is no care provider present to offer encouragement and guidance [
While the utility of physical exercise for chronic pain is well known, it has also been found that introduction of new exercise tasks is more successful when small changes are made to current daily activities [
We developed MyBehaviorCBP, a mobile phone app that operationalizes various strategies to address psychological barriers of chronic pain. MyBehaviorCBP uses machine learning on sensor data and self-reported physical activity logs and automatically generates physical activity recommendations based on an individual’s past behavior. This strategy of persuasion has been shown to be effective in MyBehavior, our predicating system designed for general populations [
In this paper, we report on a formative study on the use of MyBehaviorCBP and present results from a 5-week pilot study among individuals with chronic back pain (N=10). Since the MyBehaviorCBP automated suggestion generation approach is being tested for the first time in the chronic pain context, we investigate the feasibility and acceptability of the approach before an expensive randomized controlled trial. Prior works have recommended small pilot trials (N>4) for novel mHealth apps to investigate early evidence of acceptance and use demonstrating the intervention is affecting the intended outcomes and document lessons learned, if any, for future improvements [
We conducted a 5-week within-subject study on 10 individuals with chronic back pain. The first week of the study served as a baseline period where participants familiarized with the app. No physical activity recommendation was given in the first week. The next 2 weeks were a control phase where 7 suggestions were randomly chosen every day from a pool of suggestions. This pool of suggestions was created by a fitness expert according to the US National Institutes of Health guidelines for healthy living [
During each day of the control and experimental phases, participants filled out a short in-phone survey in the evening. The survey asked about the ease of following recommendations, how many recommendations they followed, and their emotional state. In addition to the daily surveys, participants completed a Web-based exit survey after the study. The exit survey asked about the helpfulness of the recommendations, what future changes they would want to see, and whether they would recommend this app to other people with chronic back pain.
Given the prevalence of chronic pain, invitation of the study was sent via the Wellness Center and retiree mailing lists from Cornell University. Recruitment was restricted to participants with a history of chronic back pain (≥6 months in duration) and willingness to use MyBehaviorCBP on an Android mobile phone, either their own or one provided by the study. Further inclusion criteria were having some reasonable level of outdoor movement (eg, traveling to and from work), not being significantly housebound, having a basic level of mobile phone proficiency, being between ages 18 and 65 years, and being fluent in English. Exclusion criteria, determined during an initial interview, were the need of mobility aids; having had joint replacement, arthrodesis, or limb amputation; having a learning disability; or being pregnant, but no subject fell into these categories. Eligible participants were invited for a face-to-face session where informed consent was acquired and instructions for using the app were provided (
Participant flow diagram for MyBehaviorCBP pilot study.
The MyBehaviorCBP app comprised 2 modules: routine behavior recognition module and recommendation generation module.
The first stage of MyBehaviorCBP is to log the physical activities of an individual with a combination of movement sensors (geolocation and accelerometer) along with manual input. Similar recurring activities are then grouped together to find routine behaviors. Specifically, activity states such as walking, running, stationary, and in-vehicle are automatically tracked using movement sensors within the phone; these activities are also tagged with the geographical location [
Visualization of a user’s movements over a week: (a) heatmap showing the locations where the user is stationary everyday, (b) location traces of frequent walks by the user, and (c) location traces of frequent walks by another user.
In the interest of consistency, we will refer to each of these multifaceted clusters as a “behavior” in the remainder of this paper. Note the clustering process is determined for each participant separately without using data from other participants. Furthermore, the clustering is carried out in the phone, and no location data is exported to the cloud, minimizing privacy risks.
Once the tracked data are grouped into different behaviors, the app then uses a sequential decision-making algorithm (multi-armed bandit or MAB [
Most frequent and repeated behaviors are prioritized. In doing this, we aim to exploit the fact that participants are familiar with these frequent behaviors and they likely have a higher level of mastery or sense of self-efficacy toward undertaking those actions [
Less intensive and energetic actions are prioritized. For example, walking is prioritized over running or gym exercises. This factor is considered to promote easier or perceived as easier suggestions, which may be more compelling in situations when there is fear or anxiety of contemplating exercise [
Newly generated suggestions are based on the continuation of small changes made to a user’s existing repeated behaviors. As suggested in Singh et al [
Suggestions will be uniquely contextualized to each user. Contextual information such as road or place names (
In addition to the main tenets listed above, a further requirement is the need for the system to be adaptive and future proof. Since the suggestions are generated when the app is being used and data acquired, the system only has an account of the user’s past behaviors and the suggestions that have been actualized. This information is incomplete to inform what may happen in the future (eg, an ineffective suggestion from the past may become effective at a later point in time and vice versa). Thus, the system needs to have the capacity to adjust over time and adapt if necessary. Within the MAB framework, principles from the reinforcement learning (RL) branch of artificial intelligence are used, and this learning paradigm is designed to address the task of being continually adaptive. In this context, the RL agent can take a sequence of decisions in an environment to reach a predefined objective where each subsequent decision is based on the success or failure of the previous decisions.
One can consider the MyBehaviorCBP system as an RL agent as follows: let
One exception to the above equation for
At the end of each day
The MyBehaviorCBP system intends to encourage more physical activity over a sustained period of time. However, given the early stage of the technology, a 5-week pilot study was conducted. The goal of the pilot was to investigate the feasibility of MyBehaviorCBP, which was measured by 3 factors: use, acceptability, and early efficacy. In addition, we report lessons learned for future improvements [
Use was measured by how frequently study participants opened the app, recorded from the phone log. Acceptability, a more complex quantity, was measured by traingulating a variety of self-reports that focused on intention and behavior toward the recommendations. We specifically measured perceived easiness, which indicates the actionability of the recommendations [
Number of times the app is accessed was analyzed using the simple descriptive statistics of mean and standard deviation. The acceptability and early efficacy outcomes are less straight forward to analyze because data points from the same subject being likely correlated and different subjects having different baseline conditions at the start of the study (eg, different levels of physical activity and type of chronic back pain) [
MyBehaviorCBP’s personalized suggestions for 2 users.
Different outcome measures captured in the MyBehaviorCBP pilot study and their purposes.
Data collection methods and description of outcome measure | Purpose of outcome measure | ||
Record of how many times the app is opened | Use | ||
Number of minutes spent walking per day | Early efficacy | ||
Number of minutes spent in nonwalking exercises per day | Early efficacy | ||
Perceived easiness: How easy did today’s suggestions seem after reading them? (Likert scale: 1=I could never do these suggestions to 7=I could always do these suggestions) | Acceptability | ||
Intention: How many suggestions did you want to follow today? (integer value between 0 and 7) | Acceptability | ||
Behavior: How many suggestions did you follow today? (integer value between 0 and 7) | Acceptability | ||
Pain level: Please indicate your pain level today. (Likert scale: 0=no pain to 10=extreme pain) | Early efficacy | ||
Did receiving suggestions from your phone help you to be more active? (multiple choice: not helpful, somewhat helpful, very helpful) | Acceptability | ||
How likely are you to recommend this app to another person with back pain? (multiple choice: not likely, somewhat likely, very likely) | Acceptability | ||
What changes do you think could be made to the app that would make it more effective in helping you be more active? (open-ended) | Future improvement |
The type of intervention is considered as a fixed effect, and we coded the intervention type as 0 and 1 for control (ie, the static suggestions generated by experts) and experimental phases (ie, MyBehaviorCBP suggestions), respectively. Coded this way, the intervention coefficient would represent the relative improvement of the outcome measure of MyBehaviorCBP over the control. When we included time (as day within the study) as a fixed effect, it was found to be not significant. Also, we tested the study participant identity as a random effect and found it to be significant in likelihood ratio tests (
Over the 5-week study with 10 participants, the mean number of times the MyBehaviorCBP app was opened is 106.9 during the control and experiment phases (
In the exit survey, the participants reacted positively about MyBehaviorCBP recommendations, with 2 of 10 participants finding MyBehaviorCBP recommendations very helpful and 8 of 10 finding MyBehaviorCBP recommendations somewhat helpful. No participant reported the recommendations unhelpful. All participants (10/10) reported that they would likely recommend the app to other people with chronic back pain.
The acceptability of MyBehaviorCBP was also measured using (1) self-reported rating of easiness of the recommendations, (2) how many recommendations the participants wanted to follow, and (3) how many recommendations the participants actually followed. The results of the statistical analysis are reported in
The number of self-reported recommendations followed and wanted to follow, however, had important differences for different emotional states in the day.
Number of times a day MyBehaviorCBP app was accessed.
Summary of differences between control and MyBehaviorCBP as collected from survey and physical activity logs.
Outcome measure | 95% CI |
–2logL | AICa | BICb | LRc | |||
How easy were the suggestions | 0.42 | <.005 | 0.2 to 0.6 | 0.25 | 817.5 | 879 | 894.6 | 0.009 |
# of suggestions followed | 0.46 | <.005 | 0.2 to 0.7 | 0.11 | 4795 | 4809 | 4839 | 0.01 |
# of suggestions wanted to follow | –0.2 | .02 | –0.5 to –0.1 | –0.2 | 4795 | 4809 | 4839 | 0.002 |
Walked (minutes/day) | 4.9 | .02 | 0.8 to 8.9 | 0.31 | 2123 | 2131 | 2144 | 0.009 |
Exercised (minutes/day) | 9.5 | .31 | –6.3 to 21.8 | 0.03 | 2986 | 2993 | 3008 | 0.01 |
Pain level | –0.19 | .24 | –0.5 to 0.14 | 0.17 | 1160 | 1168 | 1183 | 0.001 |
aAIC: Akaike information criterion.
bBIC: Bayesian information criterion.
cLR: likelihood ratio test between the fitted models compared to unconditional mean models [
Mean and standard deviations of acceptability measures.
Means of several outcome measures for different emotional states.
Mean several outcome measures for different preliminary efficacy outcomes.
From
Participants provided qualitative feedback in the exit survey, which provided further insights about the quantitative results and gave directions for future changes. For instance, when we asked participants to compare control group recommendations with MyBehaviorCBP, participants reported that they liked the personalization of MyBehaviorCBP. They also mentioned MyBehaviorCBP recommendations were more actionable and easier, and they were more likely to succeed if they tried the recommendations.
I really liked the personalization. I thought it was a nice touch. Suggestions were more specific and tailored, which for me made them more relevant and likely for me to use them.
...most of the suggestions were fairly easy; at least the ones that involved walking.
Because the suggestions of MyBehaviorCBP were based on my own chosen activities, I was much more likely to follow them.
I liked them more because it seemed more likely that I could do them—I was more likely, in my mind, to succeed.
Again, because the suggestions were based on my activities, they felt more feasible. I didn't have to take the extra step of thinking about how I might get the right tools (eg, bike) or where I can do the suggested exercise.
Other than changing the sitting behavior, I liked the fact that they seemed more do-able.
In addition, some participants liked the specificity of the recommendations and how they could be carried out in a smaller piecewise manner.
...they were location specific, smaller chunks of time.
...more detailed explanations/suggestions, based on past exercises logged, and having the location helped, too!
Regarding the control phase suggestions, some participants struggled with their nonpersonalized nature and how they needed to plan ahead to execute them.
...which I wanted to do the longer suggestions in version 2 [ie, control phase], unless I scheduled or planned it, I couldn't do most of them.
...I received the suggestion to ride a bike, but that's currently simply not possible, logistically.
However, one participant did not like MyBehaviorCBP recommendations and wanted more variety.
There was very little variation in the suggestions during the final 2 weeks—almost everyday was walk slightly farther and play tennis for a few minutes more... In the 2 to 3 weeks, there was a greater variety of things to try and I tried a few novel suggestions.
Participants also asked for the following features: (1) a reminder system to plan in the morning and notifications in the moment, (2) adapt suggestions based on weather or weekend/weekday, and (3) better insight to relate high pain days and activity level, etc.
It would be helpful to have reminders and suggestions pop up in the morning or at other chosen times. This could be optional and set by the user.
Maybe adding an alarm or something, to say “here, you should go do this thing now.” I think if I had something bugging me to get up and take a short walk, for example, I would be more likely to do it than just looking at a list of things I might do.
If it could ask me to rank the things I enjoy doing and then download weather data for the following days. This could suggest times when I have performed these tasks in the past and also match it with weather predictions. “You played tennis last Tuesday in the afternoon for 90 minutes. How about from 2 to 4 today when the weather will be clear and 85.”
Maybe a tally at the end of each week regarding days unable to exercise, based on back pain.
Finally, one participant wanted to use the app even after the study, and mentioned the following:
I liked this app and look forward to possibly using it permanently in the future.
To the best of our knowledge, MyBehaviorCBP is the first mobile app to provide automatically generated data driven physical activity recommendations in the chronic pain context. We conducted a pilot study to examine the feasibility of the approach. In the study, we found participants used the MyBehaviorCBP app 1 or more times a day. Furthermore, we observed early indication of acceptance and efficacy in both the qualitative and quantitative data. For instance, in the daily surveys, participants perceived the tracked data-based recommendations to be easier to follow. In the qualitative feedback on the exit survey, participants were positive to successfully complete MyBehaviorCBP recommendations. This means participants likely had a greater sense of self-efficacy toward MyBehaviorCBP-generated suggestions. According to protection motivation theory, higher self-efficacy may cause the recommendations to be carried out despite the presence of fear in chronic pain [
From
The current MyBehaviorCBP system is a variant of a prior system, MyBehavior [
Despite the similarities between the two systems, the effect sizes in
Over the years, a variety of mobile apps have been proposed for chronic pain self-management, with some apps aiming at prescribing cognitive behavioral components [
One limitation is the small number of study participants and relatively short study length. However, MyBehaviorCBP is an early stage technology. It is difficult to acquire resources to conduct efficacy trials with unproven technology on a potentially vulnerable chronic pain population. As a result, the purpose of this pilot study was to inform feasibility and acceptability. In our future work, we will use the lessons learned in this pilot study to conduct longer term studies on a larger population and also to include specific back pain outcomes such as the Oswestry Low Back Pain Disability Questionnaire [
Another limitation is that MyBehaviorCBP does not fully address the question of whether even moderate exercise can have adverse consequences. If there is a short-term pain flare, then moderate exercising can temporarily increase pain and MyBehaviorCBP should not recommend exercising during a pain flare. However, it is not clear whether exercising has any long-term adverse effect on pain. Some prior work [
In this paper, we presented the acceptability and feasibility of MyBehaviorCBP, a data-driven physical activity recommender system for chronic pain. We found preliminary evidence of increased walking activity; a few key areas of improvements have been also identified. In future work, we will incorporate these improvements and run a randomized controlled trial. If efficacy is demonstrated, then a technology like MyBehaviorCBP could have great promise because it is an automated system with no second person involved (eg, a physiotherapist). Also, all the data processing of MyBehaviorCBP is kept inside the phone which allows the app to preserve user privacy. Such automated and privacy-preserving features imply that MyBehaviorCBP has few barriers to scalability.
Akaike information criterion
Bayesian information criterion
likelihood ratio
multi-armed bandit
metabolic equivalent of task
mobile health
Photographic Affect Meter
reinforcement learning
This project is funded by the Translational Research Institute for Pain in Later Life at Weill Cornell Medical College and the National Institute of Aging. MR is funded by the National Institute on Drug Abuse (NIDA P50 DA039838; PI: Linda Collins) and the National Institute on Alcohol Abuse and Alcoholism (NIAAA R01 AA023187; PI: Susan Murphy). We thank Elaine Wethington, Xiaoya Wu, Zhe Lin, and Minghao Li for their early input in this project.
TC co-founded and has equity interest in HealthRhythms Inc, which develops mobile phone–based systems for mental health. GG serves on the advisory committee of HealthRhythms. This company, however, does not have any commercial interest in the area of chronic pain.