Objective

An automated health counselor agent was designed to promote both physical activity and fruit and vegetable consumption through a series of simulated conversations with users on their home computers.

Methods

The agent was evaluated in a 4-arm randomized trial of a two-month daily contact intervention comparing: a) physical activity; b) fruit and vegetable consumption; c) both interventions; and d) a non-intervention control. Physical activity was assessed using daily pedometer steps. Daily servings of fruit and vegetables was assessed using the NIH/NCI self-report Fruit and Vegetable Scan.

Results

Participants in the physical activity intervention increased their walking on average compared to the control group, while those in the fruit and vegetable intervention and combined intervention decreased walking. Participants in the fruit and vegetable intervention group consumed significantly more servings per day compared to those in the control group, and those in the combined intervention reported consuming more compared to those in the control group.

Conclusion

Automated health intervention software designed for efficient re-use is effective at changing health behavior.

Practice Implications

Automated health behavior change interventions can be designed to facilitate translation and adaptation across multiple behaviors.

Keywords: relational agent, embodied conversational agent, behavioral informatics, dialogue system, health behavior change intervention, physical activity promotion, walking promotion, diet promotion, fruit and vegetable consumption promotion, ontology

1.1 Intervention Design

The knowledge (“Artificial Intelligence”) in the system is encoded using two fundamentally different representations: declarative facts about the world and procedures the agent can follow during its dialog with users.

Declarative facts are themselves represented in two fundamentally different ways: epistemological knowledge about the kinds of things in the world that the agent can reason about—represented in the ontology (using OWL [22])—and specific facts about the specific behavior change theories and techniques the agent can use, and about a particular user the agent is talking to (using RDF [23]). The ontology encodes dictionary-like knowledge, such as the fact that the “Actions” the agent can take can be either “Therapeutic Actions” or “Non-therapeutic Actions” (such as saying “Hello!” to the user), and that the Therapeutic Actions available to the agent include “Therapeutic Dialog Actions” (talk therapy) and “Homework Actions” (homework assignments the user is asked to do outside of the agent counseling sessions). Examples of specific facts would be that asking the user to “Try looking for information about exercise in newspapers and magazines.” is an example of a Homework Action for physical activity promotion, that this action is appropriate for users in the Contemplation stage of change, and that the specific user the agent is talking to right now is in Contemplation.

Procedural knowledge is encoded in a hierarchical task decomposition language (CEA-2018 [24]), that is based on a theory of the structure of dialog [25]. CEA2018 represents agent goals, and actions that the agent and the user can take during a given dialog, and various relationships that hold among these entities. For example, initial goals of the agent may be to hold a counseling dialog with a user with the intention of increasing a health behavior. This goal can be matched to a “recipe” that describes one way the agent can satisfy the goal, for example by initiating a dialog, greeting the user, conducting some “small talk”, conducting the core part of the behavior change counseling, saying farewell to the user, and terminating the conversation, in that order. Each of these steps can give rise to more refined goals that are matched to other recipes. For example, the core counseling dialog for a user in the Action stage of change may consist of reviewing their progress since the last conversation, conducting a problem solving dialog about specific barriers to change, and setting goals as a new homework assignment. Ultimately, the actions in the most refined recipes will consist of specific things the agent and user can say in the conversation.

Note that much of the structure of a counseling conversation, and much of the declarative knowledge, required by a counseling agent can be described independent of a specific behavior. The above examples work for either physical activity or fruit and vegetable promotion.

The declarative and procedural knowledge is interpreted and turned into an interactive conversation for a particular user by a dialog interpreter (DTask [26]). Together, the ontology and the procedural knowledge are a unique new set of AI methods for organizing how an agent can counsel a user for a series of health behavior change conversations.Figure 1 shows an excerpt of a conversation generated using this approach.

During the development of our intervention we first designed and implemented the physical activity promotion part of the system, then extended it to also promote fruit and vegetable consumption. Through the use of the knowledge representations described above we were able to build the second system re-using 98% of the actions, 98% of the recipes, and 14% of the agent utterances, representing 22% re-use by source lines of code. The second system was built in 9% of the calendar time and 4% of the person hours required to develop the initial exercise promotion system [1].

1.2 Evaluation Study

In the rest of this paper, we describe the results of a randomized pilot study designed to provide preliminary evaluation of three behavioral interventions developed using the approach described above. We have used the ontology-based design approach to develop an animated conversational agent that plays the role of a health counselor that can promote both physical activity (ACT) and fruit and vegetable consumption (DIET) through a series of simulated conversations with users on their home computers (Figure 2). This approach and agent were evaluated in a 4-arm randomized trial of a two-month daily contact intervention comparing: a) ACT; b) DIET; c) ACT+DIET (both interventions); and d) CONTROL (a control group provided only with pedometers for tracking physical activity). The behavioral goal of the ACT system is to motivate sedentary subjects to perform the minimum recommended amount of physical activity: 30 minutes per day of moderate or greater activity [27,28]. The behavioral goal of the DIET system is to motivate subjects who are not regularly eating at least 4-1/2 cups of fruits and vegetables a day to reach this level [28]. The goal of the ACT+DIET system is to achieve both behavioral goals.

The pilot study has the following hypotheses.

• H1. Subjects receiving ACT will significantly increase their physical activity (PA) behavior, compared to non-intervention CONTROL subjects.

• H2. Subjects receiving DIET will significantly increase their intake of fruits and vegetables compared to non-intervention CONTROL subjects.

• H3. Subjects receiving ACT+DIET will significantly increase their PA behavior compared to non-intervention CONTROL subjects.

• H4. Subjects receiving ACT+DIET will significantly increase their intake of fruits and vegetables compared to non-intervention CONTROL subjects.

2.1 Measures

Physical activity was assessed at T₀ and T₁ using the International Physical Activity Questionnaire (IPAQ), which has been validated in several prior studies[29,30]. Steps walked was objectively assessed using Omron HJ-720ITC pedometers given to all study participants. While the pedometer assess walking specifically, the IPAQ assesses overall physical activity. Servings of fruit and vegetables were assessed at T₀ and T₁ using the NIH/NCI Fruit and Vegetable Scan (FVS) [31]. The FVS is a ten-item instrument that includes assessment of portion size.

Participants were weighed at T₀ and T₁ using a calibrated scale.

Satisfaction was assessed via three single scale measure items (Table 4).

2.2 Recruitment & Eligibility

Subjects were recruited from an online job posting site. Eligibility criteria for this study included: 1) age 18 or older; 2) have a home computer with Internet connection; 3) are able to walk unaided; 4) understand spoken and written English; and 5) are in Precontemplation or Contemplation Stages of Change with respect to both (a) the current DHHS/ACSM guidelines for PA (30 minutes a day of moderate-or-greater intensity PA on five or more days per week) [27,32]; and (b) current DHHS guidelines for daily consumption of fruits and vegetables (4-1/2 cups a day)[28]. Exclusionary criteria include: 1) have a medical condition that would make increasing PA level a health risk, assessed using the PA Readiness Questionnaire (PAR-Q)[33]; 2) are on a prescribed diet; and 3) have any other member of the household enrolled in the study.

2.3 Procedure

Following screening, consent, and baseline measurements in our laboratory, participants were provided with software to install on their home computer and given a brief tutorial on its use. Participants in the CONTROL group used their software only to upload their pedometer steps; participants in all other groups were also instructed how to interact with the conversational agent, and asked to conduct one conversation per day with the agent at home. All participants were provided with pedometers and instructed in their use. Following this, individuals participated in the study from home for the next two months. At the end of this time, participants returned to our laboratory to collect final outcome measures and engage in a semi-structured interview about their experience.

Table 1.

3.1 Longitudinal Analysis of Daily Pedometer Steps

Figure 4 shows pedometer steps by study participants, averaged per 15 day interval, by study group. We assume that any day in which less than 100 steps are recorded means that the participant did not wear the pedometer. We treat the step counts for this day as missing data. We fit a mixed-effect model including effect of study day (time), study condition, and an interaction. This model assumes a simple linear trend: the main effect of condition is the difference in intercepts between conditions, while the interaction is the difference in slopes. We assume here that the differences between groups in the first time period is a result of differences in baseline walking behavior between the groups prior to the start of the study (e.g., the effects of poor randomization), and thus we focus our analysis on significant changes in step counts over the duration of the study. A series of likelihood-ratio tests against reduced models indicates that the interaction effect is significant, but neither main effect is significant. The full model is shown inTable 2. Results indicate that the ACT group increases their daily walking faster than CONTROL, but the DIET group is slower, and the ACT+DIET group is somewhere in the middle. Thus, there is some evidence that multiple interventions might interfere with each other.

3.2 International Physical Activity Questionnaire (IPAQ)

The IPAQ data is highly skewed, and is transformed (Yeo-Johnson transformation) prior to analysis. A one-way ANOVA was used with the transformed IPAQ scores as the dependent variable, and study condition as the independent variable. There are no significant differences among conditions, F(3,107)=1.07, p=0.367.

3.3 Fruit & Vegetable Consumption (FVS)

Figure 5 shows mean change in FVS scores (T1 – T0) by study condition.

Model selection indicated that the simplest model is best, in which baseline FVS alone is used as a covariate in an ANCOVA, with final FVS as the dependent measure and study condition as the independent variable (Table 3). Results show a significant difference by condition, F(3,103)=4.52, p=0.005, with the DIET intervention doing the best. A Tukey post-hoc analysis indicates that the only significant pairwise difference, adjusting for multiple comparisons, is DIET vs. CONTROL, with ACT+DIET vs. CONTROL near significant.

3.4 Weight

Weight changed little for participants over the two months of the study, with no significant differences among groups, F(3,105)=1.09, p=0.374.

3.5 Satisfaction

Self-report satisfaction scores are shown inTable 4. There were no significant differences among the three intervention groups on these measures.

3.6 Qualitative Results

Semi-structured interviews were conducted with 20 participants from the ACT, DIET, and ACT+DIET groups. Participant responses to semi-structured interview questions were transcribed from the videotape and common themes were identified [34].

When asked what they like best about the agent Karen (the conversational agent inFigure 2), 7 people (35%) said that they like Karen because she was nice, personal, and was someone they could relate to, and four people (20%) said that they liked how Karen would show them their progress with the steps chart every time they had a conversation with her. Four people (20%) said that they like Karen because she would not discipline them if they did not meet a goal or log onto the system, while three people said that they liked being held accountable by Karen.

When asked if the agent helped them to achieve their goals, 10 people (50%) agreed that the agent did help (e.g., “I think it moved me closer to eating better. I am more motivated and I am thinking more about trying to get 5 servings of fruits and veggies every day.”). Four people (20%) said they believed it was a combination of using the pedometer and talking to Karen that helped them reach their goals. When asked what about the agent helped them the most, 3 people said that the constant reminder by Karen was what helped them the most.

4.1 Discussion

The evaluation study demonstrated that, while the individual health behavior change interventions performed as expected—with ACT resulting in superior walking behavior and DIET resulting in improved fruit and vegetable consumption, both relative to a non-intervention control group—the combined, multi-behavior change intervention did not result in significant improvements relative to controls. There are several possible reasons for this. One possible explanation is that our approach to intervening on multiple behaviors is to alternate which behavior was discussed on sequential logins, thus participants in the ACT+DIET group received half the number of physical activity counseling sessions as those in ACT and half the number of diet counseling sessions as those in DIET. Another possibility is that the two interventions actually interfered with each other, either on a psychological level (e.g., through distraction) or an instrumental level (e.g., taking more time to shop for and prepare fruit and vegetable dishes may have decreased time available for physical activity). Another explanation is that the combined intervention simply had smaller effect sizes compared to the individual behavior interventions, requiring more participants to reach significance. The trends inFigure 4 may also imply that the PA intervention was simply most effective for sedentary individuals (those with the lowest baseline step counts).

We have begun exploring solutions to address the problem of perceived repetitiveness mentioned by several participants [35]. In one study, we confirmed that simply increasing superficial linguistic and graphical variability during daily health behavior change counseling sessions with an automated coach agent leads to greater self-reported desire to continue using the system, as well as a significant increase in actual system usage. In a second study we demonstrated that having the coach agent tell personal “backstories” about its personal history (as if it were a person) led to significant increases in system usage compared to an identical agent that told the same stories as if they were about another user, demonstrating that users were more engaged by a more anthropomorphic agent.

Limitations of our study include the generalizability of our findings, especially given the very small convenience sample used and the relatively short duration of the intervention. We also acknowledge that our CONTROL group does not represent a true non-intervention control, since providing individuals with pedometers alone has been shown to result in increased physical activity. However, this should make any significant physical activity results stronger. Finally, we cannot determine whether the differences between groups during the first 15 days of the study were due to baseline differences or initial reactions to the study conditions, since we did not measure true pre-intervention baseline behavior.

4.2 Conclusion

Together with our earlier pilot study [1], we have demonstrated that automated health interventions can be developed using software representations that promote reuse, that the representations do indeed lend themselves to reuse, and that this approach results in interventions that are effective at changing health behavior.

4.3 Practice Implications

Adapting a computerized behavioral intervention developed for a research study (or for a different application) for large-scale deployment by an institution or public health department typically requires significant modifications to the original system, and these modifications often represent an insurmountable barrier to dissemination [36]. The use of open source software, designed using public standards and computational ontologies, with re-use in mind, significantly lowers the cost and complexity of adapting these systems for practice.

• 1.Bickmore T, Schulman D, Sidner C. A Reusable Framework for Health Counseling Dialogue Systems based on a Behavioral Medicine Ontology. J Biomed Inform. 2011;44:183–197. doi: 10.1016/j.jbi.2010.12.006. [DOI] [PMC free article] [PubMed] [Google Scholar]
• 2.US Dept of Health and Human Services. Physical activity and health: a report of the Surgeon General. Atlanta, GA: U.S. Department of Health and Human Servies, Centers for Disease Control and Prevention, National Center for Chronic Disease Prevention and Health Promotion; 1996. [Google Scholar]
• 3.Bouchard C, Stephens T. Physical activity and health: international proceedings and consensus statement. Champaign, IL: Human Kinetics; 1994. [Google Scholar]
• 4.Stewart A, Haskell W. Endurance exercise and health-related quality of life in 50–65 year old adults. Gerontologist. 1993;33:782–789. doi: 10.1093/geront/33.6.782. [DOI] [PubMed] [Google Scholar]
• 5.LaPorte R, Adams L, Savage D, Brenes G, Dearwater S, Cook T. The spectrum of physical activity, cardiovascular disease and health: an epidemiologic perspective. Am J Epidemiol. 1984;120:507–571. doi: 10.1093/oxfordjournals.aje.a113911. [DOI] [PubMed] [Google Scholar]
• 6.Young D, Haskell W, Jatulis D, et al. Associations between changes in physical activity and risk factors for coronary heart disease in a community-based sample of men and wormen: The Stanford Five-City Project. Am J Epidemiol. 1993;138:205–216. doi: 10.1093/oxfordjournals.aje.a116849. [DOI] [PubMed] [Google Scholar]
• 7.US Dept of Health and Human Services. The Surgeon General's Report on Nutrition and Health. Washington, DC: 1998. [Google Scholar]
• 8.McGinnis J, Foege W. Actual causes of death in the United States. J Amer Med Assoc. 1993;270:2207–2212. [PubMed] [Google Scholar]
• 9.Knekt P, Reunanen A, Javinen R, et al. Antioxidant vitamin intake and coronary mortality in a longitudinal population study. Am J Epidemiol. 1994;139:1180–1189. doi: 10.1093/oxfordjournals.aje.a116964. [DOI] [PubMed] [Google Scholar]
• 10.Gillman M, Cupples L, Gagnon D, et al. Protective effect of fruits and vegetables on development of stroke in men. J Amer Med Assoc. 1995;273:1113–1117. doi: 10.1001/jama.1995.03520380049034. [DOI] [PubMed] [Google Scholar]
• 11.Gaziano J, Manson J, Branch L, et al. A prospective study of consumption of carotenoids in fruits and vegetables and decreased cardiovascular mortality in the elderly. Ann Epidemiol. 1995;5:255–260. doi: 10.1016/1047-2797(94)00090-g. [DOI] [PubMed] [Google Scholar]
• 12.Key T, Thorogood M, Appleby P, et al. Dietary habits and mortality in 11,000 vegetarians and health conscious people: results of a 17 year follow up. BMJ. 1996;313:775–779. doi: 10.1136/bmj.313.7060.775. [DOI] [PMC free article] [PubMed] [Google Scholar]
• 13.Rimm E, Ascherio A, Giovannucci E, et al. Vegetable, fruit, and cereal fiber intake and risk of coronary heart disease among men. J Amer Med Assoc. 1996;275:447–451. doi: 10.1001/jama.1996.03530300031036. [DOI] [PubMed] [Google Scholar]
• 14.Ness A, Powles J. Fruit and vegetables and cardiovascular disease: a review. Int J Epidemiol. 1997;26:1–13. doi: 10.1093/ije/26.1.1. [DOI] [PubMed] [Google Scholar]
• 15.Joshipura K, Ascherio A, Manson J, et al. Fruit and vegetable intake in relation to risk of ischemic stroke. J Amer Med Assoc. 1999;282:1233–1239. doi: 10.1001/jama.282.13.1233. [DOI] [PubMed] [Google Scholar]
• 16.Li R, Serdula M, Bland S, et al. Trends in fruit and vegetable consumption among adults in 16 US states: Behavioral Risk Factor Surveillance System, 1990–1996. Am J Public Health. 2000;90:777–781. doi: 10.2105/ajph.90.5.777. [DOI] [PMC free article] [PubMed] [Google Scholar]
• 17.Krauss R, Eckel R, Howard B, et al. AHA Dietary Guidelines: revision 2000: A statement for healthcare professionals from the Nutrion Committee of the American Heart Association. Stroke. 2000;31:2751–2756. doi: 10.1161/01.str.31.11.2751. [DOI] [PubMed] [Google Scholar]
• 18.Bickmore T, Gruber A, Picard R. Establishing the computer-patient working alliance in automated health behavior change interventions. Patient Educ Couns. 2005;59:21–30. doi: 10.1016/j.pec.2004.09.008. [DOI] [PubMed] [Google Scholar]
• 19.Bandura A. Social cognitive theory: an agentic perspective. Annu Rev Psychol. 2001;52:1–26. doi: 10.1146/annurev.psych.52.1.1. [DOI] [PubMed] [Google Scholar]
• 20.Marcus B, Simkin L. The transtheoretical model: applications to exercise behavior. Med Sci Sports Exerc. 1994;26:1400–1404. [PubMed] [Google Scholar]
• 21.Miller W, Rollnick S. Motivational Interviewing: Preparing People for Change. second ed. New York: Guilford Press; 2002. [Google Scholar]
• 22.OWL Web Ontology Language Overview. (Accessed athttp://www.w3.org/TR/owl-features/.)
• 23.Resource Description Framework (RDF): Concepts and Abstract Syntax. (Accessed athttp://www.w3.org/TR/2004/REC-rdf-concepts-20040210/.)
• 24.Task Model Description (CEA-2018) Consumer Electronics Association. (Accessed athttp://www.w3.org/2005/Incubator/model-based-ui/wiki/ANSI/CEA-2018.)
• 25.Grosz B, Sidner C. Attention, Intentions, and the Structure of Discourse. Computational Linguistics. 1986;12:175–204. [Google Scholar]
• 26.Bickmore T, Schulman D, Shaw G. DTask & LiteBody: Open Source, Standards-based Tools for Building Web-deployed Embodied Conversational Agents. Proceedings of the Intelligent Virtual Agents Conference; Amsterdam, Netherlands. 2009. [Google Scholar]
• 27.Pate RR, Pratt M, Blair SN, et al. Physical Activity and Public Health: A Recommendation From the Centers for Disease Control and Prevention and the American College of Sports Medicine. J Amer Med Assoc. 1995;273:402–407. doi: 10.1001/jama.273.5.402. [DOI] [PubMed] [Google Scholar]
• 28.U.S. Department of Health and Human Services and U.S. Department of Agriculture. Dietary Guidelines for Americans 2005. Washington, DC: U.S. Government Printing Office; 2005. [Google Scholar]
• 29.Craig C, Marshall A, Sjöström M, et al. International physical activity questionnaire: 12-country reliability and validity. Med Sci Sports Exerc. 2003;35:1381–1395. doi: 10.1249/01.MSS.0000078924.61453.FB. [DOI] [PubMed] [Google Scholar]
• 30.Hagströmer M, Oja P, Sjöström M. The International Physical Activity Questionnaire (IPAQ): a study of concurrent and construct validity. Public Health Nutr. 2006;9:755–762. doi: 10.1079/phn2005898. [DOI] [PubMed] [Google Scholar]
• 31.Thompson F, Subar A, Radimer K, et al. Performance of two new cognitively enhanced fruit and vegetable short assessment forms; Proceedings of the Fourth International Conference on Dietary Assessment Methods; 2000. [Google Scholar]
• 32.Marcus BH, Selby VC, Niaura RS, Rossi JS. Self-efficacy and the stages of exercise behavior change. Res Q Exerc Sport. 1992;63:60–66. doi: 10.1080/02701367.1992.10607557. [DOI] [PubMed] [Google Scholar]
• 33.Chisolm D, Collis M, Kulak L, Davenport W, Gruber N. Physical activity readiness. B C Med J. 1975;17:375–378. [Google Scholar]
• 34.Taylor SJ, Bogdan R. Introduction to Qualitative Research Methods. John Wiley & Sons; 1998. [Google Scholar]
• 35.Bickmore T, Schulman D, Yin L. Maintaining Engagement in Long-term Interventions with Relational Agents. International Journal of Applied Artificial Intelligence. 2010;24:648–666. doi: 10.1080/08839514.2010.492259. [DOI] [PMC free article] [PubMed] [Google Scholar]
• 36.Vinson C, Bickmore T, Farrell D, et al. Adapting Research-tested Computerized Tailored Interventions for Implementation in Practice. Translational Behavioral Medicine: Practice, Policy, Research. 2011;1:93–102. doi: 10.1007/s13142-010-0008-9. [DOI] [PMC free article] [PubMed] [Google Scholar]