The next evolution, or revolution, of computerized health interventions, mobile technology and health (mHealth), appears to be underway. Mobile phones have achieved rapid and high penetration. There are over 285 million wireless subscribers in the USA alone [4], and an estimated 67.6% of adults worldwide own cell phones [5]. Approximately 75% of US high school students own a mobile phone [6]. In contrast to the initial Internet digital divide which limited the reach of computerized health behavior interventions for those in lower socioeconomic groups, mobile phone use has been widely adopted across socioeconomic and demographic groups and appears greater among those populations most in need of these interventions [6, 7]. The penetration rates in developing countries, where wireless technologies have leapfrogged the wired computer infrastructure, have produced considerable excitement in the global health community to reach and follow individuals who were previously unreachable via traditional communication channels [8, 9].

Compared to Internet interventions delivered to desktop and laptop computers, mobile interventions have the capacity to interact with the individual with much greater frequency and in the context of the behavior. As sensing technologies integrated with the mobile phone via Bluetooth or other data transmission process continue to develop, health behavior change interventions can be delivered based not only on self-reports and time/location parameters but also on psychophysiological state, social context, activity level, and behavior patterns [10]. The availability of these rich, complex, and frequent data inputs provides the potential to deliver health behavior interventions tailored not only to the person’s baseline characteristics but also to his/her frequently changing behaviors and environmental contexts [11].

To better leverage the potential of mobile technologies for health behavior interventions, health behavior theories and models need to be able to guide the development of complex interventions that adapt rapidly over time in response to various inputs. Existing health behavior theories and models have served for many years as guides to intervention development and delivery, and these theories have been the basis not only for face-to-face counseling interventions but also mass media and social marketing efforts. Models such as the Health Belief Model [12], Theory of Planned Behavior [13], Social Cognitive Theory [14], the Transtheoretical Model [15], and Self-Determination Theory [16] have served as the basis for many of the eHealth web and desktop/laptop computer interventions and have informed how interventions can be tailored to the individual’s baseline status.

As intervention developers take full advantage of mobile technologies, health behavior models will be required to guide not only tailored adjustments at intervention initiation but also the dynamic process of frequent iterative intervention adjustments during the course of intervention. The content and timing of a specific mobile phone intervention delivered via voice, text, resident application, mobile web, or other modality can be driven by a range of variables including (a) the target behavior frequency, duration, or intensity; (b) the effect of prior interventions on the target behavior; and (c) the current context of the individual (time, location, social environment, psychophysiological state, etc.). Such interventions require health behavior models that have dynamic, regulatory system components to guide rapid intervention adaptation based on the individual’s current and past behavior and situational context. Some have argued that current health behavior models are inadequate even for low-tech interventions [17], but the predominately linear and static nature of these models severely limits their ability to guide the dynamic, adaptive interventions possible via mobile technologies.

To evaluate the theoretical basis and adaptive nature of mobile interventions, the current literature on mobile technology health behavior interventions was reviewed. For this review, we define “mobile technology” as computer devices that are intended to be always on and carried on the person throughout the day (i.e., during normal daily activities). Mobile phones are the prototype device within this space, and most of the interventions reviewed used mobile phone platforms, but precursors to current smart phones and text messaging systems such as personal digital assistants (PDAs) and two-way pagers also are consistent with this definition. Excluded from this definition are devices that are portable (e.g., laptops, netbooks, iPads) but are not intended to be carried on the person throughout the day. Although technically mobile, we also excluded interventions that involved phone-based counseling, whether live or by interactive voice response (IVR), that could be delivered similarly via landline (i.e., phone-based counseling via mobile phone). Prior to the advent and widespread use of multipurpose mobile devices, special purpose devices were developed for the delivery of health behavior interventions [18, 19], but these pioneering prototypes often lacked the portability, connectivity, input–output interface, and/or computing power of current multi-purpose devices and are not included in this review. This review is also limited to the use of these mobile devices for delivering health behavior interventions, not assessments. There is a substantive literature on the use of mobile devices for health behavior assessment (e.g., ecological momentary assessment), and self-monitoring alone has been shown to be an effective health behavior intervention [20], but we excluded from this review mobile programs that involved assessment input only or that used mobile device output solely for the purpose of encouraging input (i.e., reinforcing recording adherence).

For this review, a Medline search of “mobile phone,” “cell phone,” “text messaging,” “personal digital assistant,” “PDA,” “palmtop,” “handheld computer,” and “pager” published through June 2010 was conducted. Articles meeting the definition described above were selected, and additional articles were identified from references of these initially selected articles. A considerable number of these articles were proof-of-concept reports that described technical development, an ongoing study, or reported only usability/satisfaction data. To insure review of fully functional mobile interventions, we narrowed the selection criteria to studies that reported some form of clinical outcome data, with or without randomized controlled conditions. Nearly all of the articles identified meeting these criteria fell into one of four major health behavior areas: smoking, weight loss (including diet and exercise interventions), treatment adherence (i.e., adherence to medical recommendations including both taking medication and attending appointments), and chronic disease management. Therefore, we limited this review to mobile health behavior interventions in these four areas. Others have reviewed subsets of the research reviewed here for the purposes of estimating effect sizes (e.g., [21, 22]). In contrast to meta-analytic reviews of effect size, the purpose of this review is to assess how theory has been employed in the development of mobile health behavior interventions and how the interactive and adaptive potential of these interventions may require new dynamic systems theories of health behavior.

In addition to voice, text, and data inputs, researchers have begun to use video input via camera features on these phones. Although video input has been primarily used for telemedicine purposes [78, 79], health behavior researchers have begun to use cameras for dietary assessment [48]. Another significant set of input features are automated sensors, either resident on the smartphone (e.g., accelerometers and GPS) or connected via Bluetooth. Much of the mobile health behavior interventions to date have focused on transmitting data from glucometers and spirometers, but an expanding variety of sensor technologies can be used to assess physiological states, movement, location, and other variables using mobile phones.

These early mobile health behavior interventions also have only begun to leverage the output functionality of these phones. As with input, most of the output delivery to date has been via text messaging. A few studies have used other output functions such as audio (e.g., [35]) or video (e.g. [26]), but device-resident or wireless Internet applications on smartphones provide a rich array of possible data and display outputs (e.g., progress charts, animation, videos, games) that could be used to deliver health behavior interventions.

Most apparent from this review, however, is the limited use of the rapid two-way interaction of inputs and outputs that could be used to deliver just-in-time health behavior interventions. Disease management and weight loss mobile health interventions have relied predominately on input while adherence and smoking cessation mobile health interventions have relied predominately on output. Among applications relying on input, many employed a user-initiated or unprompted input procedure despite a considerable ecological momentary assessment literature supporting the use of prompted inputs [80]. Following input, some applications provided tailored intervention responses, but others failed to deliver even a confirmation message, despite the potential importance of this feature for encouraging continued input. Of the applications that predominately pushed out intervention messages, a few encouraged confirmation of receipt or some other reply response (e.g., [43]), but most did not. None of the adherence interventions made just-in-time adjustments, but most of these interventions were single prompt appointment reminders with no longitudinal basis for adjustment, and none obtained the input necessary to make treatment adjustments. Most of the smoking and weight loss mobile interventions provided some form of “just-in-time” response based on a prior input, but the frequency or intensity of interactivity varied substantially among these interventions from requests for additional text message assistance common among the smoking interventions (e.g., [23, 27]) to regular intervention adjustments based on dietary and exercise inputs [34, 43]. The disease management interventions also typically provided just-in-time intervention adjustments, but most relied on manual adjustments by a health care professional to do so. Although more complex intervention adjustments might be best reserved for health professional judgment, standard treatment algorithms can be used to automate, with 100% treatment fidelity, many of these treatment adjustments, greatly improving scalability, and reducing professional time and costs.

A simple reminder intervention, however, is an excellent example for the necessity of theory to guide intervention. Text messaging reminders to attend appointments, take medications, exercise, etc. appear relatively straightforward and are consistent with the “cue to action” component of many health behavior theories, but the history of health behavior theories cautions us about assuming that behavior change is straightforward or that innovative modes of delivery alone are sufficient to produce behavior change. In the 1950s, mobile neighborhood tuberculosis screenings were innovative, but the added salience and convenience did not substantially increase screenings. Researchers studying uptake of these neighborhood screenings learned that the perceived threat (severity and susceptibility) of tuberculosis, the perceived benefit of being screened, and the perceived barriers to getting screened contributed to a person obtaining these screenings. These findings led to the Health Belief Model [85] and the various health behavior theories and models that followed. Therefore, a greater reliance on health behavior theories to guide mobile technology intervention development, even for apparently simple interventions, should result in interventions that address more comprehensively the potential mechanisms of behavior change, resulting in more effective interventions.

While mobile technology applications of health behavior interventions should be guided by our current health behavior models, it is important to acknowledge that these current models appear inadequate to answer many of the intervention development questions likely to arise as interventions better leverage the interactive capabilities of mobile technologies. Some have argued that these theories are inadequate even for traditional interventions [17], but they are particularly limited at informing just-in-time intervention adaptations. Dunton and Atienza have noted that the increasing availability of time-intensive information (i.e., longitudinal data obtained at an intensive frequency) allows for the intra-individual tailoring of interventions but that current health behavior theories are based on delineating between, not within-person differences [86]. Boorsboom and colleagues have argued that between-person theories do not imply, test, or support causal accounts valid at the individual level [87]. As a result, these theories and models have been used with considerable success to tailor health behavior interventions based on pre-intervention factors, but typically have not been used to adapt the intervention to the individual over the course of the intervention [88, 89]. As these interventions become deliverable on mobile devices, the content of interventions can be adapted for each individual not only initially but also over time based on his/her prior outcome data, prior responses to specific intervention outputs, current environmental and/or social context, and a range of other variables that might influence the optimal intervention based on the current state of the individual (i.e., ecological momentary interventions) [90].

In addition to adapting the content of the intervention over time, mobile technology applications of health behavior interventions also have the ability to adapt the timing of the interventions based on these same data. For example, following a prompt to exercise, how long should the program wait before prompting again and how should the timing and content of the follow-up prompt be tailored based on prior pattern of responses to prompts? In this review, only a few studies adjusted the timing of the interventions, usually during the transition from behavior change to maintenance (e.g., [23, 40]), but these timing adjustments were set a priori across all participants and were not dynamically determined at the individual level.

The development of time-intensive, interactive, and adaptive health behavior interventions via mobile technologies demands more intra-individual dynamic regulatory processes than represented in our current health behavior theories. These theories do not preclude the possibility that some conceptual components change over time, and some concepts such as reciprocal determinism in Social Cognitive Theory are explicitly dynamic in nature [14]. Moreover, a rich behavior change process–outcome literature describes dynamic interactions in face-to-face behavioral interventions that could be applied to behavioral interventions via mobile technologies [91]. Adoption of dynamical system models for mobile health behavior interventions does not require that our current health behavior theories and models be discarded, but the predominately static, linear nature of these theories appears to be a poor fit with the intra-individual dynamics of future mobile technology interventions.

Control systems engineering [92] may provide the dynamical system models needed to transform our current health behavior theories into the dynamic theories required for the time-intensive, interactive, and adaptive health behavior interventions delivered via mobile technologies. Control systems engineering examines how to influence dynamical systems to achieve a desired outcome. These dynamical systems are multivariate, time-varying processes, often nonlinear in nature, in which variables that can be manipulated (e.g., system input variables) lead to changes in outcomes of interest (e.g., system output variables) adjusting for exogenous effects (e.g., disturbance variables). Control engineering approaches have been proposed for adaptive interventions to determine, for example, when and how much to change the frequency of family counseling to prevent the development of conduct disorder in children [93].

There are a number of practical advantages of using dynamical systems to inform adaptive, time-intensive interventions delivered via mobile technologies. Control systems engineering principles and procedures are mature and provide a robust computational framework for modeling, simulation, and systematic decision making over time. These control system principles and procedures fit well with the rich longitudinal data and real-time intervention adaptations potentially available via mobile technologies. Interventions based on theories enriched by control system models also provide engineers and health behavior researchers with a common language for collaborating on new interventions delivered via mobile technologies [94].

Perhaps more importantly, these feedback control systems have the potential to reshape health behavior theories and improve our understanding of human behavior. Feedback regulatory processes are core to neurobiology. Basic neuronal processes are time-intensive, adaptive interactions of excitatory and inhibitory synaptic processes that have been modeled at the single neuron level [95]. The communalities between brain circuitry and control system engineering principles are so compelling that some brain processes are described via engineering feedback systems (e.g., [96]). Since the brain appears to regulate itself and other organs using feedback control processes, it seems reasonable that the brain also regulates human behavior via similar feedback control processes. Optimal feedback control systems have been applied to basic sensorimotor control [97] and to reinforcement learning models [98]. Therefore, the application of feedback control systems to current health behavior theory concepts provides not only potential theoretical models that are amenable to the time-intensive, interactive nature of mobile health behavior interventions but also theoretical models that are more consistent with the putative neurobiological and environmental processes that regulate these behaviors.

One significant example of regulatory feedback for disease management processes is in diabetes where frequent adjustments to diet and insulin dose are determined based on blood glucose levels. Control engineering approaches have been applied to the glucose–insulin closed-loop system [99], and this application illustrates the potential of this approach to produce the desired outcome in systems that possess measurement error, model uncertainty, noise, and lagged effects. This work could serve as a basis for developing mobile technologies for diabetes management that take advantage of advanced closed-loop control system models to automate and better regulate blood glucose for insulin-dependent patients.

Detailed descriptions of applications of control systems engineering to other health behavior interventions such as substance use have been described elsewhere [93, 100], but to illustrate the application of control systems engineering to mobile health behavior interventions, we provide a simple model for intervening on smoking urges depicted in Fig. 1 which represents an individual’s smoking behavior in terms of a production–inventory model [101]. Using a fluid analogy, smoking urge is accumulated as an inventory that is built up by an “inflow” of negative emotional stimuli (e.g., sadness, irritation) and external stimuli (e.g., presence of others smoking, drinking coffee) and depleted by an urge “outflow” resulting from smoking events and the salutary effects of positive affect and dosages of intervention components such as medications and psychosocial interventions. In a control systems conceptualization, smoking urge (y(t)) is the controlled variable, while smoking activity represents a manipulated variable (u₁(t)) whose action serves to reduce the urge to smoke. Environmental stimuli and positive and negative affect represent external disturbance variables (d₁(t), d₂(t)) whose changes build up or deplete the smoking urge inventory. The dose of intervention components represents additional manipulated variables (u₂(t)) that can be selected adaptively by a set of decision rules or mathematical algorithms (referred to as the intervention controller in Fig. 1) based on assessed participant response. A differential equation drawn from the general principle of conservation of matter can be used to describe the relationship between these variables [102]

where K_d1, K_d2, K_u1, and K_u2 represent coefficients that quantify the rate of change in urge based on changes in manipulated variables (intervention dosages, smoking activity) and disturbance variables (negative and positive affect, environmental stimuli). The inherent dynamics of smoking behavior in Fig. 1 constitutes a closed-loop feedback system in which increases in smoking urge precipitate a decision to smoke (the relationship between urge and smoking activity is denoted by the “urge controller” box in Fig. 1).

The potential of mobile technologies to generate the intensive longitudinal data sets reflecting the behavior of a smoker over time and observed in different settings and contexts is important for building a dynamical system model such as the individual smoking behavior model illustrated above. Mobile devices can collect information in real time about the controlled, manipulated, and disturbance variables, as represented by sensor locations in Fig. 1. Model estimation techniques such as system identification [103] and functional data analysis [104] can be used to estimate the K_d1, K_{d 2}, K_u1, and K_u2 coefficients from these data. These modeling techniques also can be applied to reverse-engineer the functional relationship underlying the urge controller which reflects the actions of an individual smoker. With a dynamical systems model, it becomes possible to apply control systems engineering to develop algorithms that use real-time assessments and predicted responses from the model to adaptively decide on the timing and dose of the intervention components. Mobile technology is an enabler to advanced control algorithms such as model predictive control [105] that employ formal optimization methods to decide on current and future doses of intervention components while satisfying clinical practice preferences and restrictions [100]. The use of these control technologies in a smoking cessation intervention as shown in Fig. 1 parallels the closed-loop control system for diabetes management described previously. The optimization process leads to specific intervention decisions (u₂(t)) for providing behavioral strategies and/or medication prompts to replace smoking as a means of reducing smoking urge.

In addition to being sensitive to threshold (or mean) levels of key variables, a control system framework can adapt to and consider variability of key controlling or explanatory variables. Most theories of health behavior change to date have focused on mean levels of predictive variables, but with the advent of newer data technologies, we can develop models that focus on lability or stability of key controlling factors. For many health behaviors, maintaining a stable level or tighter range for some key variables may be as or more important than increasing or decreasing mean levels.

Novel dynamic health behavior theories and their application to health behavior interventions via mobile technologies will require empirical validation. Adaptive treatment methodologies has been applied to interventions that adjust over the course of weeks or months, not hours or days, but these methodologies can be applied to mobile technology interventions [106]. Other methodological considerations for adaptive interactive technologies applied to eHealth also may be appropriate to consider for mHealth interventions [107, 108]. The opportunity via mobile devices to collect intensive context- and time-dependent (longitudinal) data and to systematically vary intervention components enables researchers to test not only these components but also the theoretical concepts and dynamic models that underlie them.