We developed BEAR by employing an iterative process in which two investigators (JC and MZM) mapped constructs reported in the literature as determinants of behavioral change and acceptance of CDSSs (see Figure 1). In each iteration, both investigators mapped the constructs from a framework or a literature review into the emerging construct pool, starting with the Theoretical Domains Framework (TDF) [28,30] and the Unified Theory of Acceptance and Use of Technology (UTAUT) [34]. At the beginning of each iteration, the two investigators (JC and MZM) developed maps independently contrasting the information in the articles with the definitions in the construct pool. After that, the investigators discussed differences in their maps and agreed on modifications to the pool. These modifications encompassed the following: the inclusion of new constructs, in addition to agreeing to their definitions; changes in construct labels or definitions; and changes in the grouping of constructs into domains. The emerging framework was progressively documented in two files: one contained the definitions and the other contained the map to the original constructs (see Multimedia Appendices 1 and 2).

The first iteration comprised the mapping of two well-established frameworks: the TDF [28,30] and the UTAUT [34]. The TDF, proposed by Michie et al in 2005 [30] and revised by Cane et al in 2012 [28], comprises 84 theoretical constructs included in classic psychological theories about behavior change. The UTAUT, on the other hand, was proposed by Venkatesh et al in 2003 [34] to integrate the eight predominant models at the time about technology acceptance. The UTAUT comprises eight constructs that influence the regular use of a technology, directly or indirectly. The selection of these frameworks as a starting point was guided by the authors’ previous experiences with the evaluation of medical informatics interventions.

Subsequent iterations comprised the mapping of constructs identified through a literature review of recent aggregative studies presenting determinants of the acceptance of CDSSs. To obtain the initial pool of references, we queried Scopus, MEDLINE (Medical Literature Analysis and Retrieval System Online), Embase, CINAHL (Cumulative Index of Nursing and Allied Health Literature), and PsycINFO. Figure 2 presents the search strategy used in Scopus. We constructed equivalent searches for the other databases.

Two investigators (JC and MZM) screened the initial reference pool and evaluated their titles and abstracts. For a reference to be selected, it needed to fulfill all of the following inclusion criteria: (1) address the topic of CDSSs, (2) employ literature review methods to obtain its results, and (3) include, among its results, determinants of the acceptance of CDSSs. Exclusion criteria were limited to the following: (1) the publication was not a research article (eg, abstract, viewpoint, commentary, editorial, or protocol) and (2) the full text was not in English or Spanish.

We limited the search strategy to articles published since 2014; we were working under the assumption that, although recent, the aggregate studies found would cover the relevant literature published before that year. That assumption was validated by documenting the period covered by each included review.

The investigators (JC and MZM) then extracted the constructs and their definitions from the full text of each article (see Multimedia Appendix 1). In cases where definitions or descriptions were not explicitly stated, the investigators reviewed the full text of the cited articles, including supplemental materials. Furthermore, in cases where other frameworks were used to organize the review findings, those frameworks were included in the mapping exercise with their own iterations.

The grouping of constructs into domains was initially informed by the organization and definitions in the base frameworks (ie, TDF and UTAUT). Later on, new constructs identified from the literature reviews were contrasted with the domain definitions to choose their locations. In some cases, this process resulted in the creation of new domains or in changing previous definitions.

Finally, preliminary versions of the framework were discussed with the other authors (ZLL, SLKG, and RDB) and other colleagues, resulting in the refinement of construct labels, definitions, and grouping.

The mapping process comprised 13 iterations, corresponding to 10 reviews and four frameworks. The initial literature search identified 584 references. After removing duplicates, 405 references passed through the screening process (see Figure 3). Of these, 23 were selected for full-text assessment; however, text from six of these articles was not available. After reviewing the 17 available articles, seven were excluded. Finally, 10 articles met all inclusion criteria [35-40,42,45]. Only one article was excluded based on language. Table 1 shows the characteristics of the reviews included in this study and any associated frameworks [29,35-45].

The 10 reviews included in the mapping exercise (see Table 1) synthesize 219 studies from 1995 to 2018; these studies include participants from several populations of health professionals (ie, nurses, general practitioners, specialists, pharmacists, residents, laboratory technicians, physical therapists, medical assistants, medical students, paramedics, psychologists, and social workers). Additionally, two frameworks were included: the CFIR (Consolidated Framework for Implementation Research) [29], which was used in Ross et al [37], and the HOT-fit (Human, Organization, and Technology-fit) framework [41], which was used in Kilsdonk et al [38] and Van Dort et al [44].

Table 2 presents the constructs and domains included in the proposed framework [28,34,46-50]. The mapping of each construct into the sources is included in Multimedia Appendix 2. BEAR comprises 156 constructs arranged into 22 domains. Domain definitions are included in Table 2, whereas definitions for each construct are included in Multimedia Appendix 1.

Our objective was to develop a framework, grounded in the literature about determinants of behavioral change and technology acceptance, that would be useful to researchers investigating the implementation of CDSSs as a strategy to foster the uptake of evidence-based recommendations.

The idea of BEAR originated in our search for a conceptual framework to guide our research in the use of CDSSs as a strategy to implement clinical practice guidelines. From the beginning, we realized that the effectiveness of such an approach would be mediated by aspects of behavioral change and technology acceptance. We found part of the guidance we were looking for in the TDF [28,30] and the UTAUT [34]. These frameworks provide constructs that address both aspects of the phenomenon, although at a higher level than the one we were seeking, particularly on the side of technology acceptance. For example, UTAUT includes the concept of facilitating conditions, defined as “the degree to which an individual believes that an organizational and technical infrastructure exists to support the use of the system” [34]. However, that definition is not enough to identify what specific facilitating conditions are missing, which we consider a necessary step for the development of interventions. This need for further detail is what led us to an iterative process by which we incorporated the findings of recent literature reviews about determinants of the acceptance of CDSSs.

To capture the variability in the phenomenon of implementing CDSSs, we sought to include constructs that are specific enough to facilitate the identification of what is different between one implementation experience and another. For example, we added constructs to represent four quality aspects for the system’s output: accuracy, completeness, specificity, and timeliness. However, recognizing that it is unlikely that we have identified every relevant concept, we also included general constructs, in this case output quality. Along the same lines, most domains include a construct that shares the domain’s label. In cases where a domain label corresponds to a group of concepts (eg, memory, attention, and decision processes), we do not include a general construct but only those representing each constituting concept.

The decision to include both specific and general constructs led us to make two more decisions about the framework’s structure: (1) to have only one grouping level (ie, domain and construct) and (2) to include each construct only once, inside the domain where, in our opinion, the construct fits better. With these decisions, we sought to control complexity while maintaining detail. The resulting domains sort the constructs thematically; that is, constructs included in a domain represent determinants that could influence the concept represented by the domain, instead of particularizations of that concept.

BEAR is not a parsimonious framework. We believe this is both a strength and a limitation. On the one hand, we expect that the level of detail facilitates the identification of actionable determinants; on the other hand, using the whole framework could be difficult, particularly in quantitative-oriented projects. However, for most projects, using every construct in the framework is not necessary or even advisable.

Our recommendation is to use the framework at the domain level during the initial stages of research design, particularly when discussing scope. Later, in qualitative-oriented studies, BEAR could be utilized to develop data collection guides for interviews, focus groups, or observations. This could be done at the domain level, in the case of exploratory studies, or based on selected constructs. During analysis, BEAR could serve as an initial coding schema, either at the domain or construct levels, raising the researcher’s awareness of determinants and supporting the identification of categories.

In quantitative-oriented studies, besides informing decisions about scope and research questions, BEAR could support the search for theories and measurement tools. In both cases, we recommend reviewing the definitions and references provided in Multimedia Appendix 1.

In the next section, we present two examples of how we have used BEAR in our research.

We are currently using BEAR in a qualitative-oriented project whose objective is to identify barriers and facilitators to the use of a CDSS to support the implementation of a case-finding recommendation included in the Colombian chronic obstructive pulmonary disease (COPD) clinical practice guideline.

According to the recommendation, suspicion of airflow obstruction could be established by the identification of specific risk factors and symptoms. Once the suspicion is established, spirometry should be ordered to confirm the limitation in the airflow [51].

In the study, we explained the recommendation to primary-level physicians and asked them to use a CDSS, implemented as a mobile app, during their patient encounters. The CDSS asked a series of questions about the patient’s clinical history and current symptoms. Using this information, the system applied a decision algorithm to establish the suspicion of COPD, in which case it recommended that the participant order a spirometry test to discard or confirm the diagnosis.

After 2 months using the system, we interviewed the participants to explore their experiences applying the recommendation and using the system. The guide used in these interviews was developed using BEAR at the domain level. Table 3 presents selected questions included in the guide. Designing the interview around BEAR’s domains allowed us to explore both the behavioral change and the technology acceptance aspects of the intervention.

Data collected in the interviews were analyzed thematically using BEAR’s constructs as the initial coding schema. During the initial analysis, the transcript below—adapted from the original data in Spanish—was coded under the following constructs: patient–health professional relationship, diminished eye-contact, and patient’s preferences.

Later in the analysis, the review of the content in these codes led to development of two categories: perceived loss of attention and negotiating the use of the device with the patient. The first category refers to the way the patient seems to interpret and resent that he has lost the physician’s attention when the latter is using the computer. The second category refers to the way physicians prevent complaints about themselves when using the phone during the encounter by telling the patients what they are using the phone for and sometimes including them in the process of using the system. The relationship between these categories allows us to recognize differential effects over the patient–health professional relationship of CDSSs implemented as mobile apps and as desktop applications.

Alert fatigue is a common problem for clinicians who use technology designed to improve patient safety. Evidence-based strategies to overcome alert fatigue are lacking, especially in the intensive care unit (ICU). There is an evidence gap, as discussed in a recent review and guidance document [52].

The goal of the project in this use example is to provide effective strategies for the management of alert fatigue in the ICU. The behavioral change of interest is increasing clinician responsiveness to CDSS alerts provided during patient care (eg, ordering medications). Formative research needs to be completed to understand the barriers and facilitators to clinicians’ responsiveness to alerts. To meet these goals, a mixed methods approach was applied using a survey and in-depth interviews conducted with critical care clinicians. Questions were developed based on BEAR at the domain level. A sample of selected questions is provided in Table 4.

Out of 122 constructs identified in the literature review, 52 (42.6%) mapped to the TDF (see Multimedia Appendix 2). This supports our initial assumption that, in the context of CDSSs, behavioral change and technology acceptance are interrelated. Since the TDF was selected as a source from the beginning, it is not a surprise that both its constructs and structure had a substantial influence on the resulting framework. However, the TDF’s constructs emerged from the mapping process with some modifications. In some cases, these changes corresponded to the integration of constructs (eg, the TDF’s skills, ability, and competence, whose definitions in the American Psychological Association Dictionary of Psychology [48] are similar). In other cases, we changed the construct definition to facilitate its interpretation in the context of clinical practice change. For instance, the TDF’s definition for modelling—“In developmental psychology the process in which one or more individuals or other entities serve as examples (models) that a child will copy” [28]—was changed to “The process in which one or more individuals or other entities serve as examples (models) for a person to copy.” We believe these alterations do not substantially change the meaning of the affected TDF constructs, but rather improve their applicability.

The CFIR [29] and HOT-fit framework [41] were also part of the mapping process. However, several of their constructs were not included in the resulting framework due to differences in the scope. Whereas BEAR deals with behavioral change and technology acceptance from the individual's perspective, the CFIR considers the implementation as a whole, integrating other perspectives (ie, those related to the organization, the government and health system, and the implementation project [29]). In some cases, those perspectives intersected. For example, the CFIR’s inner setting—implementation climate construct represents an organizational characteristic that influences the individual’s behavior. However, in other cases (eg, the CFIR’s process—reflecting and evaluating construct), we did not identify a direct influence over the individual. The same happened with the HOT-fit framework’s organization domain [41]. We recognize that the level of influence of a particular construct over the individual’s behavior could be a matter of debate—indeed, we had several discussions about it during the mapping process—thus, we used our better judgment. For information about the mapping of specific constructs, we refer the reader to Multimedia Appendix 2.

This is certainly not the first attempt to apply technology acceptance models in health care [31,53]. The majority of these attempts have tried to adapt the Technology Acceptance Model (TAM) [33], one of UTAUT’s eight contributing frameworks [34]. BEAR has similarities and differences with these works. On the one hand, BEAR attempts to cover a wide range of possible determinants, but it does not make statements about the magnitude of their influence on each other or the individual’s behavior. In other words, BEAR does not attempt to state a theory about technology acceptance. Instead, BEAR is meant as an exploratory tool that allows for the identification of determinants that could be articulated into hypotheses and potentially form the basis of interventions. We hope that the study of those hypotheses in different health care contexts results in a future theory that is able to explain and predict practice change in the context of CDSSs.

On the other hand, literature reviews have found that the TAM has low explanatory power in health care environments [31,53]. The authors of these studies attribute this lack of fit to professional differences between health professionals and other workers [53] and to the fact that the TAM does not completely incorporate the emotional and cultural aspects of health care decision making [31]. Our interpretation of these findings is that behavioral change determinants operate differently in health care in comparison to other work environments. If that is correct, it could be expected that bridging the gap between behavioral change and technology acceptance brings forward the missing pieces in the puzzle.

Finally, our objective of this study was to develop a conceptual framework, not a theory. A theory serves as an explanation of a phenomenon that in many cases allows for the prediction of outcomes. However, our objective was not to predict an outcome or the relative weight of each determinant in the explanation of an outcome, but to synthesize and organize all potential determinants reported in the reviewed literature. That is why we do not state any conclusion about the relationship between specific constructs, besides grouping them into domains to facilitate organization and presentation, nor their relative contribution to the success in the implementation of CDSSs. In this sense, BEAR is akin to other determinant frameworks [54], such as the TDF [28,30] or the CFIR [29], rather than a theory, such as the UTAUT [34].