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 JMIR Research Protocols, is properly cited. The complete bibliographic information, a link to the original publication on http://www.researchprotocols.org, as well as this copyright and license information must be included.
Physical activity has established health benefits, but motivation and adherence remain challenging.
We designed and launched a three-arm randomized trial to test artificial intelligence technology solutions to increase daily physical activity in cancer survivors.
A single-center, three-arm randomized clinical trial with an allocation ration of 1:1:1: (A) control, in which participants are provided written materials about the benefits of physical activity; (B) text intervention, where participants receive daily motivation from a fully automated, data-driven algorithmic text message via mobile phone (
The study runs for 5 weeks: a one-week run-in to establish baseline, followed by 4 weeks of intervention. Data for study outcomes is collected automatically through a wearable sensor, and data are transferred in real-time to the study server. The recruitment goal is 42 participants, 14 in each arm. Electronic health records are used to prescreen candidates, with 39 participants recruited to date.
This study aims to investigate the effects of different types of intelligent technology solutions on promoting physical activity in cancer survivors. This innovative approach can easily be expanded and customized to other interventions. Early lessons from our initial participants are helping us develop additional advanced solutions to improve health outcomes.
Retrospectively registered on July 10, 2017 at ClinicalTrials.gov: NCT03212079; https://clinicaltrials.gov/ct2/show/NCT03212079 (Archived by WebCite at http://www.webcitation.org/6wgvqjTji)
There is consistent evidence that identifies poor dietary choice and physical inactivity as major contributors to death in the US and worldwide [
Recent advances in hardware technologies, statistical methods, big data processing, and cloud-based computing have resulted in artificial intelligence technologies that may offer efficient, low cost and potentially scalable solutions to deliver individualized behavioral interventions to at-risk populations. Further, the development of intelligent technology solutions via ecological momentary assessment provides unique opportunities for behavioral intervention at the individual level, which may increase adherence and promote long-term lifestyle change after the intervention is completed. To test the utility of such technology, our team constructed artificial intelligent agents (IA) to help cancer survivors become more active throughout their daily routines, via a technology-driven clinical trial designed to deliver affordable, scalable, easy to adopt, individualized behavioral interventions. Currently, the IA approach is being tested in a three-arm randomized trial: The Physical Activity by Technology Help (PATH). The trial compares the two technologies—a voice activated intelligent agent and a more traditional intelligent text messaging—compared to a self-driven traditional behavior change (control). The aim of this pilot study is to assess the preliminary effect of a 4-week intervention by different technological approaches to increase daily physical activity, defined as walking (eg, 10,000 steps per day), among overweight and obese cancer survivors, with a special interest in under-represented African American women, in the state of Maryland. In addition, we developed an innovative approach to identify study candidates from review of Epic Electronic Medical Records (EMR). While this IA is specific to cancer survivors, the methods and technology design may be replicated in the general population.
This is a three-arm single center pilot randomized trial with allocation ratio of 1:1:1. The arms are: (A) control, in which participants are provided with written materials about the benefits of physical activity; (B) text-messaging intervention, where participants receive multiple motivational messages from an automated, data-driven algorithmic text message program via mobile phone (
Physical Activity by Technology Help (PATH) study design.
Physical Activity by Technology Help (PATH) schedule of enrollment, interventions, and assessments.
Study’s activities | Study period | |||||||||
Enrollment | Allocation | Postallocation | Close-out | |||||||
0 | ||||||||||
Eligibility screen | X | |||||||||
Informed consent | X | |||||||||
Allocation and wearable installation | X | X | ||||||||
Alexa Installation | X | |||||||||
Control | X | X | X | X | ||||||
Coachtext | X | X | X | X | ||||||
MyCoach | X | X | X | X | ||||||
Individual characteristics | X | X | ||||||||
Baseline physical activity | X | |||||||||
Outcomes | X | X | X | X | ||||||
Analysis and reporting | X |
Participant inclusion criteria are: (i) adult cancer survivors residing in the State of Maryland who have had one or more of the cancers of interest (breast, prostate, colon, lung, cervical, oral, melanoma); (ii) completed active cancer treatment (surgery, chemotherapy and/or radiation) at least three months prior to enrollment, with the exception of anti-hormonal therapy; (iii) overweight or obese (BMI 25 and above) status; (iv) internet access at home via Wi-Fi; (v) access to an Android or Apple smartphone; (vi) ability to perform low-intensity daily steps (walking), with physician approval; (vii) less than 150 minutes per week of physical activity reported during the previous four weeks; (viii) willingness to wear an accelerometer; and (xi) willingness to consent and accept randomization. Due to disparities of cancer survivorship among minority groups [
Exclusion criteria are: (i) reported engagement in, or more than, 150 minutes per week of physical activity during the previous four weeks (The Godin-Shephard Leisure-Time Physical Activity Questionnaire) [
Participants are recruited using both passive and active strategies. In the
Once eligibility is determined, an IRB-approved consent designee obtains an IRB-approved written informed consent form (ICF) in the outpatient clinic or at the study office. A signed and dated IRB-approved ICF is documented in a secure participant study file prior to initiating study-related procedures. As part of the consent process, the study team 1) informs each participant of study procedures and requirements and allows sufficient time for the individual to decide whether to participate in the study; 2) answers questions about the details of the study; and 3) ensures that the ICF is approved by the IRB when an amendment to the study protocol is made. Participants are free to withdraw consent for participation in the study at any time, without affecting their current or future treatment.
The investigators designed this study to compare: (i) a dialogue Intelligent Agent, using data driven Interactive Digital Voice Assist on Amazon Alexa/Echo speaker (
Participants are randomized with an allocation ratio of 1:1:1 to the control self-driven arm (Group A),
Those assigned to the control arm (A) receive educational materials and are advised to increase physical activities to 10,000 steps per day. Participants in the control arm receive National Cancer Institute (NCI) printed educational materials providing summary evidence about the benefits of exercise for cancer survivors [
Participants in the
preferred name>, you were too inactive yesterday. You got <#step_count_yesterday steps>. Your goal is 10,000. How about taking a stroll today with <dog_name>. Aim for 10 minutes or more. This will not only be good for you, but is essential to <dog_name's>good health
The third arm is
The
Alexa, I’m planning to go out for a walk. Ask my coach if I need sunscreen.
Alexa, ask my coach, “How is my progress so far?”
Alexa, I’m not making good progress; ask my coach for health tips to help me become more active.
The study utilized Amazon’s prebuilt user-machine interaction data to track participant utilization of
Both interventions (B) and (C) have personalized and tailoring components. Each participant completes a questionnaire about personal family status, habits, pet ownership, work status, and preferred schedule to wake up, eat, work, and relax. All of this information, plus the individual physical performance data, are used to personalize and tailor each individual experience with both interventions. All participants are motivated to increase their physical activities to at least 10,000 steps per day regardless of the arm assignment. Participants in all arms are able to keep the Fitbit Charge 2 HR as a gift from the study team at the end of the trial.
The study team designed each intervention for a specific population, cancer survivors, to best fit their needs, thus the current intervention design is only suitable for such a target population. However, the innovative study design, operations, and technology development make this study easily adaptable for the more general population with slight design modification(s).
Due to the nature of the study interventions, study participants are not blinded. Data for the principal outcomes, however, are collected using sensors. Therefore, the study data analyst is blinded during data collection and during analysis.
The aim of this study is to assess the effectiveness of different technological approaches in increasing physical activity among overweight and/or obese cancer survivors captured by wearable sensors (Fitbit Charge 2 HR). The primary outcome is the percent change in daily steps from the 1-week run-in to the end of the 4-week intervention period. The data for the primary aim is collected in real-time and transferred simultaneously to the computational server of the Johns Hopkins University via an application programming interface (API) that directly transmits the data from the user’s wearable device to the server. Participant wear-time is validated using the heart rate sensor readings on a minute-by-minute basis.
In addition to the primary outcome, the study will examine the following secondary outcomes: 1) total number and duration of activity bouts (defined as 3 minutes or more of uninterrupted activity); 2) fragmentation indices (counting transitions between active/inactive periods) and variability indices (counting the magnitude of transitions between different levels of physical activity); 3) daily diurnal patterns of activity that model patterns in minute-by-minute profiles; and 4) weekly and daily trends of total steps. Levels of physical activity intensity will be defined using each participant’s personalized heart rate reserve.
The Johns Hopkins School of Medicine IRB reviewed the protocol, information technology (IT) solutions, safety, and security of the data collection & storage, and approved the study (# IRB00113882). The Sidney Kimmel Comprehensive Cancer Center at Johns Hopkins (SKCCC)/The Clinical Research Review Committee (CRC) also reviewed the protocol and found that this study poses minimal risk to the participants, and assigned Minimum Risk category to this study. SKCCC/CRC recommended forming Data and Safety Monitoring Board for this study & SKCCC/CRC will conduct random audits. To ensure the safety and security of participant information, the
As the primary outcome variable is the 4-week change from the baseline in the average number of steps per day, the primary analysis will use a two-sample t-test to compare each active intervention arm with the control arm based on intent-to-treat strategy. A Bonferroni correction will be used to adjust for multiple comparisons.
For the secondary outcomes (activities bouts, fragmentation, and patterns), a linear regression model will be fitted for the 4-week change from the baseline in the number of steps and duration of activity bouts each day. The independent variables include the baseline number of steps, study arm (modeled by two dummy variables for the two active treatment arms, and the control arm is treated as reference group), and any potential baseline confounding factors. Poisson regression will be used to analyze the total number of activity bouts, fragmentation indices, and variability indices. The independent variables include baseline number of steps, study arm, and any potential baseline confounding factors as the independent variables. In all analyses, data will be properly transformed to fit the model assumption before regression analysis. Although missing data is not anticipated to be a challenge, small amount of non-wear time for charging, bathing, etc. will have minimal effects on the primary study aim since it depends on the average steps across the intervention period. Since Stratified Permuted Block Randomization is being employed, we will consider minimization to balance interventions simultaneously over several prognostic factors to ensure equal distribution across arms. To date nearly 80% of the target population has been enrolled, and only one participant stopped wearing the tracker one day a week for religious reasons.
This study designed, developed, and launched a clinical trial to investigate the effect of novel emerging technologies to assist cancer survivors to become more active in their daily lives. The study uses innovative technology and methods to apply autonomous artificial intelligent agents as an intervention. This is an affordable, scalable, and easily deployed personalized and tailored intervention that has the potential to transition from the research setting to general use. It is important to offer a scalable and affordable solution that can assist cancer survivors to become more active. Given the wide general utility of the technology and applicability of the outcome of interest, this approach may also be applicable to the general population with slight modification of the intervention and testing in a larger population study.
The trial operation is minimalist in design and effort. Recruitment activities and enrollment were designed and developed with a focus on low cost and efficiency. The study team developed innovative prescreening in the EMR environment, with remote intervention deployment and data collection processes. The entire trial is operated by the PI and the study coordinator. Data collection processing is autonomous and no interference from the study team or participants is needed. The data flow on a minute-by-minute basis to a structured database for real time analysis allows decision support for the artificial intelligent agent. The same data will be used subsequently to assess the effect of the intervention on the primary outcome. The trial’s recruitment is currently active and eligible consented patients have been enrolled.
Intelligent Agents (IA) are not new. Stuart Russell and Peter Norvig researched this topic extensively in their leading textbook in Artificial Intelligence—Artificial Intelligence: A Modern Approach [
Artificial intelligence (AI) has promising wide applications in translational research. Enablers to use such technologies, including tools and developer kits, are widely available for interested researchers, while codes libraries are shared among most researchers and powerful platforms are widely used in the form of mobile devices and home speakers. The current study provides insight beyond the target outcomes by its efficient operation. The entire study is operated by 1-2 persons. This is possible because most of the study components, including data collection, have been automated or enhanced by technology solutions. Even the start of the intervention is automated. The study team also automated the instructional communication via scheduled automated emails based on each participant enrollment date. The written consent process is still in paper format, a requirement of the University IRB.
In summary, the current trial demonstrates a very effective and efficient model for intervention that holds the potential to substantially change the way interventions are applied, monitored, and analyzed . The study has been design, developed and launched within the projected cost and timeline. However, challenges still exist. The main challenge for adopting such technologies include understanding the technology’s limitations; how behavioral theories can be adapted to human-machine interaction; limited funding sources for such research; privacy concerns; and the limited aggregable performance metrics for such technologies. Developing such interventions requires in-depth knowledge of AI and their operations, specifically programming. Most of the available behavioral change theories have been developed without consideration of human machine interaction and the potential limitations of such interactions. There are also still concerns about the stability and confidentiality surrounding AI in general, particularly in a health care setting; thus, there is no standardization for adopting AI technologies as there is for traditional interventions. As a result, the design concept for adapting the health coach functions to the AI machine was not straightforward. The user experience aspect was not traditional since no interface currently exists, and voice detection was the main interaction method. It was difficult to orchestrate different components, devices, and algorithms and to develop a secure environment/platform for transferring data within the requirements of the IRB. As a result, it was necessary to build our own server and connect all devices and databases via secure APIs. Further, the study team also overcome the security concerns by building creative solutions including transferring the decision of data transmission and devices communication into the hands of the participants in a form of electronic permissions and unique credentials. Lessons from this study are already helping to develop a more advanced Intelligent Agent to tackle common chronic conditions that rely on self-management and coaching and provide evidence that advancement in technology can be an agent of change to improve health outcomes.
The study was approved as IRB #IRB00113882 on March 16, 2017. Recruitment activities were initiated in late April 2017. We started with targeted advertisements throughout Johns Hopkins oncology outpatient clinics. We also started active prescreening on Epic EMR to pre-identify participant prior to clinical encounter. The protocol went through a revision to refine the definition of physically active in the inclusion criteria using the American Heart Association definition of physically active. The recruitment activities are expected to last until March 2018. As of January 2018, we screened 75 patients and 39 are already enrolled in the study.
Dashboard for tracking user machine interaction.
Detailed peer review reports and responses by the scientific review committee at the SKCCC.
artificial intelligence
application programming interface
Clinical Research Review Committee
clinical research management system
electronic medical records
intelligent agents
informed consent form
Institutional Review Board
information technology
Johns Hopkins University
National Cancer Institute
natural language processing
Physical Activity by Technology Help
principal investigator
Sidney Kimmel Comprehensive Cancer Center
The study team would like to express their appreciation for the State of Maryland Cigarette Restitution Fund, and for continued support from Norma Kanarek, PhD, MPH, the Executive Director of the Maryland Cigarette Restitution Fund Research Grant at Johns Hopkins. The team would also like to thank Hadas Zeilberger for assisting in coding software. This study was funded by the Faculty Retention Award from the State of Maryland Cigarette Restitution Fund.
VS received research funding from Pfizer and Novartis. SM has served on scientific advisory boards for Quest Diagnostics, Sanofi/Regeneron, Amgen, and Akcea Therapeutics. He has received research support from the PJ Schafer Cardiovascular Research Fund, David and June Trone Family Foundation, American Heart Association, Aetna Foundation, Maryland Innovation Initiative, CASCADE FH, Nokia, Google, and Apple. The rest of the authors declare no potential conflicts of interest.