According to the County Health Rankings [1], variation in health can be accounted for by health behaviors (30%), clinical care (20%), social and economic factors (40%), and physical environment (10%). Increasing evidence shows that lifestyle-related behaviors such as diet, exercise, sleeping, emotion, and smoking play an essential role in people’s health. Chronic diseases caused by unhealthy behaviors and habits are among the leading causes of mortality [2]. Some of the chronic diseases, for example, type 2 diabetes, could be lifelong and bring a heavy burden to the patients and their family. The only way to prevent many chronic diseases is to change unhealthy lifestyles, for example, diet and physical activity.

With the potential for low cost and high scalability for chronic disease prevention, in the past decade, digital health interventions (DHIs) have been widely discussed by government stakeholders, clinicians, and researchers [3]. Designing and deploying DHIs are challenging due to the complexity of human behavior, which could be affected by individuals’ motivation, emotion, ability, social environment, and physical environment. Therefore, DHI design could accordingly require theories and practice from several disciplines, including phycology, public health, behavioral science, human-computer interaction, and so on. The interdisciplinary nature of DHIs calls for a comprehensive framework to guide the development, evaluation, and report.

As DHIs are expected to change human behavior, behavioral theories can serve as the development foundation. It has been shown that theory-based behavior change interventions are more effective than others [4,5]. Nevertheless, behavioral theories could also be ignored [6] or misused [7]. Although behavioral theories allow the explanation and prediction of behavior, they lack the guidance for translating into operational techniques.

The behavior change technique (BCT) taxonomy [8] and persuasive system design (PSD) principles [9] are two widely used taxonomies in DHI research [10-13]. These taxonomies not only inform DHI design but also enable precise reporting, which is favored by systematic reviewers. Although derived from different philosophies, BCTs and PSDs have some common elements. However, they are used separately in many DHI studies. To benefit from both, we combined the BCT taxonomy and PSD principles into a more comprehensive taxonomy in the light of the behavioral intervention technology (BIT) model [14].

In this paper, we aim to put the puzzles together and build a holistic framework to aid DHI researchers to design, evaluate, and report their studies. In short, our contributions include the following:

All the three reviewed works above mention behavioral theories, but only IDEAS explicitly integrates behavioral theories into the steps of the development process. Behavioral theories refer to the social-psychological theories of behavior change, which explain and predict human behavior. As depicted by Sutton [18], each of the behavioral theories specifies a small number of cognitive and affective factors as the proximal determinants of behavior (see Figure 1). These factors are called constructs in behavioral science [7]. We will use this term to refer to the fundamental components of behavioral theories in the rest of the paper.

Glanz et al [19] illustrated the most frequently used behavioral theories published before 2010: the social cognitive theory (SCT) [20], the transtheoretical model of behavior change (TTM) [21], the health belief model (HBM) [22], and the theory of planned behavior (TPB) [23]. Davis et al [24] also identified 82 behavioral theories, among which the most frequently used theories are TTM, TPB, SCT, the information- motivation-behavioral skills model, HBM, the self-determination theory [25], the health action process approach (HAPA) [26], and the social learning theory [27]. Based on different assumptions of human behavior, these behavioral theories can be grouped into continuum theories and stage theories [28].

Continuum theories assume people’s behavior is caused by a set of variables, for example, intention and skills. Except for TTM, all other mentioned theories fall into this group. Based on the behavioral model integrating several existing ones [28], we present a hypothesized continuum model as shown in Figure 1. The constructs in black are borrowed from the integrated behavioral model in [28]. Planning (shown in red in Figure 1) is specified as a mediator of the intention-behavior relationship in HAPA [26,29,30]. The habit (shown in green in Figure 1) has been found to moderate the effects of planning on behavior change [31].

Stage theories assume people change their behavior in a process including several stages. The factors pushing people from one stage to the next are believed to be different. Therefore, the strategies at each state should be adapted accordingly. For example, Figure 2 shows the stages and strategies of TTM, which is adapted from [32]. This model divides the behavior change process into five stages, namely precontemplation, contemplation, preparation, action, and maintenance and relapse prevention. Depending on the stage of change, different strategies could be applied accordingly to make the intervention effective.

Behavioral theories provide a toolbox to understand human behavior and explain the rationale behind interventions. However, their shortcomings should be noted before they are used. Hekler and colleagues [7] have pointed out three shortcomings of behavioral theories: (1) most behavioral theories explain only a small portion of variance in the outcomes they are trying to account for, (2) many behavioral theories, in their current form, are not falsifiable, and (3) there is a fragmentation and an overabundance of different theories. Therefore, DHI developers should not be limited to behavioral theories. With the emergence of DHIs, the existing behavioral theories can be further evaluated and improved [33]. Here we list some guidelines when using specific behavioral theories: [34] and [35] for the SCT, [36] for the HBM, [37] for the TPB, and [38] for the HAPA.

While behavioral theories can predict and explain human behavior, there is a gap between theories and operational interventions. Will self-monitoring increase self-efficacy for promoting physical activity? Will information about health consequences affect perceived advantages or disadvantages? Due to the high complexity of human behavior and health, one DHI may involve several techniques. The lack of a consistent taxonomy of DHIs will lead to poor replicability and low comparability of the results from related studies. Although there exists taxonomies to bridge the theory-intervention gap, the use of different taxonomies still hinders the understanding and contribution among communities. Therefore, we present the DHI taxonomy, a unified taxonomy taking advantage of two widely used taxonomies (the BCT taxonomy and PSD principles) in light of the BIT model.

BCTs are defined as observable, replicable, and irreducible components of an intervention designed to change behavior [8,39], for example, self-monitoring or goal setting. Abraham and Michie developed a taxonomy of BCTs that identified 22 BCTs and 4 BCT packages [8] and was later extended to a taxonomy containing 93 BCTs in 16 groups, called BCT taxonomy (v1) [39]. The BCT taxonomy has been used for informing intervention development and report [40,41] and identifying the effectiveness of BCTs [13,42-44]. It also provides a means to evaluate health and fitness apps [12,45-47] and wearables [48]. From the official website of the BCT taxonomy [49], we found a collection of 405 intervention studies with BCT coding. We show the word cloud of BCTs based on this collection in Multimedia Appendix 1. The top five used or tested BCTs are goal setting (behavior), instruction on how to perform a behavior, problem solving, information about health consequences, and action planning.

In related work, we have introduced the BIT model [14]. In terms of the intervention strategies in the BIT model, only some examples (ie, education, goal setting, monitoring, feedback, and motivation enhancement) were provided. We think the BCT taxonomy can serve as a strategy pool for the BIT model.

Aiming to create a conceptual framework that can be directly applied to persuasive system development, the PSD model describes 28 principles in four categories (supporting primary task, computer-human dialogue, system credibility, and social) as an extension of Fogg’s work on persuasive technology [50]. Table 1 describes the details of PSD principles. We found 16 principles that have the same or similar definitions with counterparts from the BCT taxonomy. For example, self-monitoring appears both in PSD principles and the BCT taxonomy. Tunneling (1.2) in PSD principles has the same meaning as the BCT “4.1 structure on how to perform the behavior” (refer to Multimedia Appendix 2 for more details). There are three PSD principles (ie, cooperation, competition, and recognition) whose counterparts could not be found from the BCT taxonomy; these can serve as a supplement to the BCT taxonomy.

Next, we present the diagram of our DHI taxonomy (see Figure 3). The blue part is the strategy part, whereas the green part is the characteristic part. We have just shown its strategy part, which includes 93 (the BCTs from the BCT taxonomy) plus 3 (cooperation, competition, and recognition from PSD principles) strategies. The other part of our DHI taxonomy corresponds to the characteristics. The BIT model described four characteristics (medium, complexity, aesthetics, and personalization). Inspired by the characteristics-related PSD principles (Table 1), we included social role and trustiness, in addition to the mentioned four from the BIT model, into the characteristic part of the DHI taxonomy.

We divided the PSD principles into two groups. The ones fitting the definition of the BCT were placed in the strategies group, while the others fell into the characteristics group. Personalization is one of the characteristics in the BIT model. We found that tailoring has a very close meaning to personalization according to their definitions in the PSD principles [51], whereas similarity and liking are also in line with the definition of personalization. Therefore, we regard tailoring, similarity and liking the same as personalization. Likewise, trustworthiness, surface credibility, real-world feel, and verifiability were merged to one characteristic as trustiness. By dividing the PSD principles and merging the overlapping principles, we hope our new taxonomy can reduce the confusion and difficulty of coding DHIs [6,10] (see Multimedia Appendix 3 for the complete list of elements in our DHI taxonomy).