The systematic literature review followed an a priori defined protocol as registered in PROSPERO (reference number 42016037954) and reporting is consistent with the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) statement [18]. Inclusion and exclusion criteria are outlined in Textbox 1.

A systematic search of the following databases was undertaken: Cumulative Index to Nursing and Allied Health Literature (CINAHL), Cochrane Database of Systematic Reviews (CDSR), Cochrane Central Register of Controlled Trials (CENTRAL), Cochrane Library (including Database of Abstracts of Reviews of Effects [DARE] and Health Technology Assessment [HTA] databases), Database of Promoting Health Effectiveness Reviews (DoPHER), Embase, MEDLINE, PsycINFO, Trials Register of Promoting Health Interventions (TRoPHI) and Web of Science (Social Science Citation and Science Citation Index). All databases were searched from 2000 until March 2016. Reference and citation searching were also undertaken. The searches were performed by an experienced Librarian at the Norwegian University of Science and Technology (NTNU). The search strategy included subject indexing terms and free-text terms for title, abstract, and keyword searching. The search terms were grouped into 3 concepts: (1) back pain, (2) digital interventions, and (3) self-management. The search terms were selected with reference to previous systematic reviews of interactive digital interventions for hypertension [19,20] and asthma [21,22] and after discussion with the review team. The full version of the search terms used, including specifications on use of title, keywords, or abstract screening, is documented for the example of MEDLINE in Multimedia Appendix 1.

All identified citations from the searched databases were uploaded to Distiller software (Evidence Partners). An integrated duplication detection tool was used to identify duplicates. All suggested duplicate pairs were screened for correctness by one reviewer (LS). Title and abstract screening was performed for each article by two independent reviewers from four (LS, BN, MM, NS). Disagreement between the two reviewers resulted in inclusion of the citation to full-text screening. Full-text screening was similarly performed by two independent reviewers from four (LS, BN, MM, NS), assessing the eligibility of the citation. Any disagreement was resolved through discussion mediated by a third reviewer (PJM).

Similar to the study selection process, data extraction was performed independently by two of four reviewers (LS, BN, MM, NS) using the Distiller software. Discrepancies in data extracted were considered by LS by revisiting the original paper to adjudicate on appropriateness and discussed and finalized with BN where required. Data were systematically extracted on study settings (country, inclusion and exclusion criteria, recruitment and participation numbers); study population (baseline characteristics such as age, gender, ethnicity, duration of symptoms, comorbidities); description of the intervention (details on the key components, characteristics, and underlying theoretical concepts); and outcome measures (time-points for outcome assessment, choice of primary outcomes, included secondary outcomes and effects, if any, noted as well as attrition rates, where available).

Our primary and secondary outcomes of particular interest are outlined in Textbox 2. These outcomes were a priori defined as of interest, however all outcomes reported were included in the data synthesis. For this review, pain-related disability was of special interest, as it measures a construct of the physical functioning domain, which has been recommended as a core domain in LBP research by several authors and guidelines [23-25].

The methodological quality of all included studies was assessed using the Cochrane Collaboration tool for assessing risk of bias in randomized trials [26]. Two reviewers independently assessed selection bias (allocation concealment and randomization procedure); blinding of participants, personnel, and outcomes assessors; completeness of data; selective outcome reporting; and other potential biases. Any disagreements were resolved through discussion by the two independent reviewers (BN, PK). Papers were not excluded from the study on the basis of quality.

The study population, intervention components, outcomes, and characteristics of the included studies were narratively described. In our protocol we stated that we would conduct a meta-analysis if included studies were sufficiently homogeneous; however, due to the heterogeneity of identified studies, meta-analysis was not possible. Quantitative results from all outcomes reported in the completed RCT studies were described as either favoring the intervention group, no difference between groups, or favoring the control group. The outcomes reported in Textbox 2 were used as a basis to structure the results for research question 2. Included protocols for future RCTs were used to consider intervention components, characteristics, and outcome measures, but were not included in synthesis of intervention effects.

We identified a total of 7014 citations, including 8 from searching reference lists of included studies. From these, 2316 were excluded as duplicates, and thus a total of 4698 titles were screened, resulting in the screening of 729 abstracts and 89 full-text papers. A total of 11 references concerning 9 different studies that described 5 RCT study protocols and 6 RCT reports met the inclusion criteria [27-37]. The PRISMA flow diagram demonstrating the screening process is illustrated in Figure 1 (adapted from Moher et al [18]).

Four of the 9 separate studies were undertaken in the United States [28,29,32,33], 2 in Germany [35,37], one in the United Kingdom [31], one in Australia [27], and one in Spain [36]. The studies were published between 2010 and 2016. The 6 completed RCT reports included a total of 2706 participants, with a range of 114-1343 participants per study (Table 1).

The characteristics of the study population in each of the studies are described in Table 1. There was considerable variation between studies in the duration of LBP symptoms, content and delivery of the interventions, and the measured outcomes. In 6 studies, LBP was defined by participant self-report [27-29,32,35,37] and in 3 studies by general practitioner evaluations [31,36] or diagnosis codes from medical records [33]. Seven studies included participants with pain for more than 3 months [27-29,32,33,36,37]. Only one study included participants with current LBP at the time of screening (or within the past 2 weeks) [31], while Simon et al [35] included only acute LBP participants, defined as participants who had experienced pain for less than 3 months. The included populations had a mean age ranging from 42.5 to 52.7 years; one study did not report the age of the population, except to say they were 18-65 years [32], a further study also had an upper age limit of 65 years [36]. All the remaining 7 studies did not report any upper limit in their inclusion criteria, yet only one study reported the age range of participants, 18-79 years [29]. In 5 [28,29,32,35,37] of the 6 RCT reports, the majority of the participants were female (58%-83%). The 6th study, which was conducted within the American Department of Veterans Affairs, included only 11% females in the intervention group and 14% in the control group [33]. Included participants were generally Caucasian (74%-87%) and the majority (42%-75%) reported educational levels as time spent at college or higher.

Results for the 7 items of the Cochrane Risk of Bias tool for the 6 completed RCTs are reported in Table 2. Four studies had one or more items rated as unclear risk of bias [28,29,32,35], 4 studies had one item with high risk of bias [28,32,35,37], and only one study had a low risk of bias for all 7 items [33]. One study was assessed to have a high risk of incomplete data [35] because the attrition rate was extremely high. Two studies were rated as high risk of selective outcome reporting as one had no published protocol and reported the outcome with highest effect size as a primary outcome [28], and the other did not report on the primary outcome stated in the published protocol [32]. One study was assessed to have high risk of bias for other potential biases due to differences in educational level between the groups [37].

Key characteristics of the digital interventions are summarized in Table 5.

We have systematically searched and reviewed the literature pertaining to interactive, digital interventions for self-management of LBP. The effectiveness of interventions was mixed, with only 1 study reporting a positive effect on their primary outcome [28]. We found a large degree of heterogeneity regarding the description of intervention content and delivery, theoretical underpinnings, and outcomes reported, making comparison between interventions difficult. A comprehensive description of intervention development and use of theory has been recommended when reporting on RCTs of digital interventions [38]; however, such descriptions were either brief or completely lacking in the included studies. Participants were predominantly female, white, younger, and well educated, which renders the external validity of the identified studies as low.

Despite international recommendations for reporting core outcome domains (physical functioning, pain intensity, and health-related quality of life) in LBP studies [23], we identified 16 different primary outcome measures and a total of 52 outcomes covering a wide range of domains. Better consistency in choice of outcome measures was seen in the 3 RCT protocols of planned RCTs [27,31,36]. We expect that these trials will provide more useful information and data for future meta-analyses. Generally, the included studies were not able to demonstrate significant beneficial effects on either the primary or secondary outcomes and we were unable to identify specific characteristics of interventions to explain these findings. However, it may be that the most important factors related to whether an individual engages with a digital support tool were not taken into account, for example, low mood and additional physical comorbidities. Surprisingly, physical activity, which is considered one of the mainstays of the treatment of LBP [10], was only included as a key component in one study. Consequently, no evidence was presented to support effects on physical activity behavioral changes from digital self-management for LBP. This should be a matter for focus for future RCTs in this area. None of the studies showed any evidence of harm from interactive digital interventions. There was no evidence regarding cost-effectiveness of interactive digital interventions.

This systematic review was undertaken by a team with extensive experience in conducting such reviews. We used multiple databases, and a thorough search strategy that was designed iteratively by the research team and an information specialist to account for the 3 different dimensions of the search (back pain, digital interventions, and self-management). The methodological assessment tool used in our systematic review has been specifically developed to assess the risk of bias in RCTs [26], and its constructs are in line with the recommendations of the PRISMA statement [18]. All aspects of data extraction, quality appraisal, and data analysis were carried out independently by two researchers, with a third party available for adjudication in case of disagreements.

The primary limitation of this systematic review is the sparse literature related to our objectives. Due to the sparsity and heterogeneity of the data, a formal meta-analysis was not possible. Additionally, our search was limited to studies published in English, Danish, or Norwegian, which could be construed as a limitation, although there is increasing evidence that this is not a particular problem [39]; six papers were excluded at the abstract screening stage of this review based on language. Finally, gray literature was not included; however, given the nature of this review and that there is no suggestion of publication bias, it is unlikely that this will have any impact on the results.

To the best of our knowledge this is the first systematic review of RCTs of interactive digital interventions for self-management of LBP. However, systematic reviews of Web-based interventions for LBP (not specifically self-management) [16], nondigital self-management for LBP [13], and chronic musculoskeletal pain [40] have been published. The first review suggests that CBT-based approaches and interventions that offer Web-based support may have some effect on reducing pain-related catastrophizing and improving patient attitudes; however, study quality was relatively low and further studies were recommended [16]. Reviews of interventions targeted specifically at self-management have suggested that there is only moderate-quality evidence that self-management has small effects on pain and disability in people with LBP [13,40]. These reviews have not dismissed self-management as a treatment option for LBP, but rather suggested that further research is needed to understand the limitations of self-management and whether or how effectiveness can be increased. In addition, these reviews have suggested that future studies should extend the outcomes of interest to include aspects of self-efficacy, and also consider the impact of the duration of the intervention [13,40], increase the length of follow-up [16], and also consider the impact of such interventions on health care utilization [16]. Similar conclusions have been made in systematic reviews of digital self-management interventions in conditions like asthma [22], hypertension [19], and problematic cannabis use [41]. Tailoring digital interventions to individual patient needs has been advocated to enhance engagement [42]. Our review highlights that although 5 out of 6 of the RCT reports included some form of tailoring, there was a lack of detail on exactly what this involved and the role it played in the outcome of the RCT or in user engagement. Finally, small and very similar effects across types of interventions such as different types of exercises, manual treatment, or acupuncture for people with LBP are well recognized [43]; however, because of the enormous societal impact of LBP and LBP-related disability, these interventions may still have worthwhile effects both at the patient and population level [43,44]. In this context, digital interventions aiming to promote self-management are particularly attractive because they are easy to deliver, inexpensive, and safe.

The populations within the identified studies were predominantly female, white, well-educated, and middle-aged, and thus the wider applicability of digital self-management interventions remains uncertain and therefore further investigation including a broader range of participants is merited. Seven of the 9 included studies specifically aimed to address the self-management of chronic LBP, and thus the usefulness of supporting self-management for acute LBP using digital tools remains underinvestigated; any such interventions for acute LBP would possibly require different advice and support to that offered for chronic LBP, as directed in clinical guidelines [10,11]. In addition, the absence of any health economics data was surprising and certainly needs to be addressed in future studies. There were a number of areas of reporting that were identified as deficient in the majority of studies in this systematic review. This suggests that going forward greater adherence to published guidelines that have recommended increasing focus on reporting of the technical aspects of the digital intervention as well as reporting the content of the intervention and its theoretical underpinnings [38,45] would be valuable. Finally, there is growing evidence that tailoring of digital interventions may be an important ingredient for success [42], and this will be an important issue to address in future RCTs of digital interventions aimed at promoting self-management of LBP. We are aware of at least one such study currently underway [46,47].

Our review has highlighted that the published literature is extremely heterogeneous and that digital intervention studies for LBP are generally poorly described. The literature provides insufficient detail regarding target and participating populations, and intervention components, theoretical underpinnings, and the rationale for the wide variety of outcome measures used. This makes it difficult to gain a clear impression of what might work best, for whom and in what circumstances. It is clear that the existing evidence has not yet proven the wider utility of digital interventions for self-management of LBP for the population at large, a knowledge gap that future research should address by better characterizing participants and interventions in a way that would allow replication and by providing clear rationales for intervention components and outcome measure selection.