Guidance published in Preferred Reporting Items for Systematic reviews and Meta-Analyses [26] and the Cochrane Handbook of Systematic Reviews [27] was adhered to. The a priori protocol for the review is published in the International Prospective Register of Systematic Reviews (PROSPERO): CRD42018089322.

Studies were identified by searching web-based databases with support and consultation provided by an institutional librarian. The search included the Medical Literature Analysis and Retrieval System Online (MEDLINE), Excerpta Medica Database (EMBASE), Cochrane Central Register of Controlled Trials (CENTRAL), Cumulative Index to Nursing and Allied Health Literature (CINAHL), PsycINFO, Physiotherapy Evidence Database (PEDro), and PubMed and gray literature databases (Dissertation Abstracts International WorldCat, The Grey Literature Report, and Open Grey) using both Medical Subject Headings and free-text keywords relating to OA, digital-based structured self-management interventions, and outcomes stated below from inception to May 2018. Search strategies for the first 3 databases above and dates on which searches were conducted are listed in Multimedia Appendix 1. Searches of the reference list of the previous review papers and included studies were also conducted.

Two reviewers (RS and JJ) independently screened abstracts and full text of the search results. Any disagreements were resolved through discussion. Data were extracted by 2 researchers (JJ and EH; mentioned in Acknowledgments) and cross-checked by a third researcher (RS) using a data extraction tool developed a priori based on the Cochrane Handbook recommendations. Data items included participants’ characteristics, guiding theory or rationale, mechanism of effect, the digital medium, method of delivery, who delivered, where it was delivered, adherence, and fidelity. Additional data relating to describing components of the interventions were extracted based on the Template for Intervention Description and Replication (TiDieR) guidance [28] and SMP components [13,14]. The behavior change techniques used were also extracted based on the hierarchical taxonomy by Michie et al [29] to examine if certain techniques were favored or, if there were sufficient data, whether particular techniques yielded bigger effects. The data extraction form was piloted on 2 included papers before full data extraction.

Outcome data were expressed as SMD and were pooled in a pair-wise fixed effects model stratified based on the outcome and type of digital medium. We used the mean (SD) of within-group change from baseline to calculate the SMDs. Per Cohen [30], SMDs <0.2 were classified as small, those between 0.2 and 0.8 were classified as medium, and those >0.8 were classified as large. We converted the SMD to the percentage change in the outcome measure by multiplying the SMD with the SD of the control group of the sufficiently powered study (ie, the trial with the largest sample size).

A sensitivity analysis was conducted for the risk of bias assessment. A subgroup analysis was performed to assess the effect of SMPs on the different digital media used, control conditions, and presence of active exercise component in the intervention. Heterogeneity across pooled studies was examined using the chi-squared test and I² statistics. As there were <10 studies, we did not use a funnel plot or Egger test to assess publication bias. Furthermore, a narrative synthesis of the studies was conducted where there were insufficient data to pool studies in a meta-analysis. We did not conduct additional exploratory analyses to explore potential moderators because of an insufficient number of studies. The common study characteristics tables were supplemented by additional summary tables including the TiDieR process information; risk of bias assessment; effect estimates; Grading of Recommendations, Assessment, Development, and Evaluations (GRADE) analysis; and behavior change taxonomy groupings. We used Review Manager (RevMan, version 5.3; Cochrane Collaboration) for all analyses.

We contacted the authors of included studies for missing data. Aggregated data were provided by authors for 6 studies. We used the RevMan calculator to impute missing SDs from the test statistics reported for 2 studies.

Risk of bias was assessed using the risk of bias tool of the Cochrane Handbook for Systematic Reviews [27]. Quality of evidence for outcomes was assessed according to the 5 GRADE domains, including study limitation (risk of bias), inconsistency, indirectness, imprecision, and publication bias [31,32].

A total of 2001 titles and abstracts were screened after excluding duplicates, of which 1950 records did not meet the inclusion criteria (Figure 1). Full texts of 51 potential eligible records were read, and 8 studies published between 2008 and 2018 were included. These studies included 2256 (range 113-352) individuals with OA. A total of 8 digital-based structured SMPs were compared with 10 control conditions. Two studies were cluster RCTs and 6 were RCTs. A list of studies excluded at the full text screening stage with reasons for exclusion is presented in Multimedia Appendix 2 [33-75]. A summary of the study characteristics and participants’ demographics are presented in Tables 1 and 2.

A total of 6 studies were conducted in the United States, and 2 were performed in the Netherlands.

The number of participants in the studies ranged from 113 to 855. The median (IQR) age was 63 (3.55) years. Most of the participants in the 6 studies were female (median 74.7%, IQR 6.9%), whereas in the remaining 2 studies, most participants were men (93.0% and 90.7%).

Two studies recorded participants with knee OA only; 5 studies recorded a mixture of knee (median 71.55%, IQR 15.77%), hip (median 16%, IQR 5.75%), or both (median 6%, IQR 10.5%); and 1 study did not record the area affected. No studies recorded OA in the hand or any other type of OA. Five of the studies recorded the time since diagnosis, with a median of 15 (IQR 2) years.

Digital self-management interventions were compared with treatment as usual (n=4), wait list control (n=2), in-person physical therapy (n=2), attention group (health education; n=1), and no intervention (n=1).

The risk of bias was assessed as unclear for random sequence generation (n=3), allocation concealment (n=1), and incomplete outcome data (n=1). There was a high risk of bias for incomplete outcome data in 1 study (Figure 2). The detection bias and performance bias were rated high in all studies because of the nature of the interventions and patient-reported outcome tools.

A summary of findings for the main comparisons of the study with GRADE ratings is presented in Table 7. Detailed meta-analytic forest plots are also presented in Multimedia Appendix 4.

The planned subgroup analysis based on the digital medium used to deliver the intervention revealed that the mobile app (n=1 study, n=197 participants) resulted in the largest effect size in pain reduction (SMD –0.38; 95% CI –0.67 to –0.10), followed by the internet medium (n=4 studies, n=841 participants; SMD –0.33; 95% CI –0.46 to –0.19; I²=0%). The telephone as a medium of intervention delivery resulted in the smallest effect size in pain reduction (n=2 studies, n=576 participants; SMD –0.18; 95% CI –0.35 to –0.02; I²=0%). However, the difference between the 2 groups was not significant (X²₂=2.3; P=.32; I²=13.4%). Subgroup differences for physical function outcome were not significant (X²₂=0.1; P=.97; I²=0%) between telephone, internet, and mobile apps, and resulted in similar small effect sizes (SMD –0.23; 95% CI –0.40 to –0.07, SMD –0.27; 95% CI –0.41 to –0.14, and SMD –0.24; 95% CI –0.51 to 0.03, respectively).

The intervention in most studies had exercise and/or physical activity components, and, in a few studies, it was the main component of the intervention. Therefore, we conducted a subgroup analysis accordingly. The pain reduction in studies with exercise and/or physical activity as the main component of the intervention was greater than the pain reduction in those studies without exercise and/or physical activity as the main component of the intervention (SMD –0.37; 95% CI – 0.54 to –0.20 and SMD –0.23; 95% CI –0.36 to –0.11, respectively). However, the difference was not statistically significant (X²₁=1.7; P=.19; I²=40.6%). The improvement in physical function was comparable in both groups—(SMD –0.27; 95% CI – 0.44 to –0.11) in studies with exercise and/or physical function as the main component and (SMD –0.24; 95% CI –0.38 to –0.11) in studies without exercise as the main component (X²₁=0.1; P=.76; I²=0%).

No subgroup analysis could be conducted for the follow-up time points as all 3 studies reporting the 12-month follow-up data used the internet as the intervention delivery medium. Similarly, no subgroup analysis could be conducted on the studies with high rates of completion because intervention was delivered via the telephone in 2 studies that did not report completion rate, or there was no measure of adherence.

Risk of bias sensitivity analysis, excluding studies with no or unclear random allocation, inadequate treatment allocation concealment, and/or incomplete outcome data, reduced the effect size in both pain reduction and improvement in physical function (SMD –0.19; 95% CI –0.32 to –0.05 and SMD –0.19; 95% CI –0.32 to –0.06, respectively).

The findings of this review indicate that digital-based structured SMPs compared with the treatment as usual or no intervention control groups resulted in a significant, homogeneous, and medium reduction in pain (SMD –0.28; 95% CI –0.38 to –0.18) and improvement in physical function (SMD –0.26; 95% CI –0.35 to –0.16). The SMDs reduced slightly at longer follow-up time points but remained significant with medium effect sizes. However, the findings should be interpreted with caution as the overall quality of the body of evidence is moderate because of the risk of bias in the included studies. Using SD from the control group of the largest (most adequately powered) study by Lorig et al [82], the SMDs translate to 5.70% reduction in pain and 5.07% improvement in physical function at postintervention time points. In accordance with Tubach et al [84] and Angst et al [85], we determined a minimal clinically important difference (MCID) of 15% in pain and 8% in physical function. Therefore, these effects are unlikely to be clinically significant. The findings on disability and QoL are less clear, as these outcomes were reported in 1 study with very small effect sizes.

Both immediate and longer-term effects of digital-based structured SMPs on pain in our review (SMD –0.28; 95% CI –0.38 to –0.18 and SMD –0.20; 95% CI –0.35 to –0.5, respectively) were slightly greater than those reported in a Cochrane review comparing standard OA SMPs with usual care (SMD –0.26; 95% CI –0.41 to –0.10 and SMD –0.17; 95% CI –0.26 to –0.08, respectively) [19]. However, the immediate and longer-term effect of digital-based structured SMPs on physical function was greater in our review compared with the immediate effect of SMPs reported in the Cochrane review by Kroon et al (SMD 0.01; 95% CI –019 to 0.18 and SMD –0.16; 95% CI –0.25 to –0.01, respectively) [19]. Despite the moderate quality of evidence reported in both reviews and a lack of direct comparison between the digital-based structured SMP and the SMP intervention, it could be postulated that the digital-based structured SMPs may have similar or greater effects than conventional SMPs on pain and physical function. Considering the potential cost-effectiveness of digital interventions, digital-based structured SMPs have the potential to reach populations reluctant or unable to attend face-to-face appointments. This could have a considerable benefit from a public health perspective despite the modest size of the effect.

Previous reviews indicate that eHealth interventions have the potential to reduce treatment costs [21,86]. We could not assess intervention costs because none of the included studies compared intervention costs across groups. In the study by Kloek et al [80], participants in the intervention group (internet-based exercise) had 5 in-person physical therapy sessions, whereas those in the control group visited a physical therapist an average of 12 times. Nevertheless, both groups showed a significant improvement in most health outcomes. Reduced numbers of physical therapy visits would likely result in a reduction in health care costs. Future RCTs of digital-based structured SMPs are encouraged to assess the cost-effectiveness of digital interventions in isolation or combination with face-to-face sessions. It should be noted that smartphones and/or internet devices are now owned by the majority of the population; therefore, these interventions are accessible and can be delivered conveniently and easily to the target audience. However, it should be noted that digital-based structured SMPs may not be suitable for all patients, potentially because of age, preference, comorbidities, and/or severity of illness; therefore, the intervention must be tailored to the needs, preferences, and conditions of patients and include face-to-face, group, and digital modes in the intervention package.

The result of the subgroup analysis indicates that there is no significant difference among the different digital modes of SMP delivery. However, mobile app (1 study) and internet SMPs (4 studies) resulted in medium effects on pain outcome, whereas telephone SMPs (2 studies) showed a small effect. Potential reasons for this observation could be that the majority of participants in the telephone studies were male veterans (>90%) who had more comorbidities and/or severe OA symptoms. In addition, the interventions were low intensity, and telephone call sessions were short (average 16.6 min). However, the small number of studies and the fact that the 2 telephone trials had a low risk of bias cautions against this interpretation; inflated effect size in studies with a high risk of bias is expected. The results of further subgroup analysis indicate that studies with exercise or physical activity as the main component of the intervention resulted in a greater improvement in pain but not in perceived physical function. A possible explanation for the reduced pain in this group of studies could be because of the change in pain tolerance and decreased perception of pain after exercise [87,88]. Four studies (n=2 telephone and n=2 internet) were included in the risk of bias sensitivity analysis. The analysis resulted in a reduction in the effect size compared with the main meta-analysis for both pain and physical function outcomes. Notably, the 2 combined telephone and video studies contributed most to the SMD (weight >65%) in the sensitivity analysis.

Attrition at postintervention was modest; median 10% (IQR 7%) in the intervention groups and median 8% (IQR 4%) in the control conditions. Despite common concerns with high dropout rates in digital interventions [89], the reported attrition rates in the included studies appear reasonable compared with behavioral studies. In a recent study, Bennell et al [90] reported that a web-based exercise program improved home exercise adherence and confidence in the ability to undertake exercise compared with a home-based exercise program prescribed by a physical therapist’s usual methods [90]. In our review, treatment adherence in 6 studies reporting treatment completers was 46% in 1 study [79], >80% in 3 studies [78,80,81], and >90% in 2 studies (Table 1) [82,83]. There are a few possible reasons for the relatively low dropout and nonadherence rates. First, interventions in all studies were tailored to participants’ needs and conditions. Second, researchers used features in web-based or mobile apps to develop reminder and monitoring systems and created higher interactivity of the intervention delivery. Third, in some studies, health professionals maintained contact with the study participants during the study through either face-to-face meetings or telephone calls. However, it should be noted that the participants were highly educated in most studies and, in some studies, self-selected (ie, responded to the study participants’ recruitment advertisement). Thus, they may have been enthusiastic about the new intervention. In future research and development, the advanced features of mobile apps and internet interventions could be employed to deliver even more effective and tailored monitoring and/or motivational interventions. Digital interventions can also be used to show patients’ symptom improvement, which is a useful and effective way not only to improve retention and compliance but also to increase their effect through self-efficacy [91,92].

There are a few limitations to the current review. First, only studies published in English were considered. Second, blinding of intervention and outcome assessment is not possible; therefore, all studies suffer from performance and detection biases. Moreover, some of the most powered studies have a high risk of bias, lowering the overall quality of evidence. Third, disability and QoL outcomes were only reported in 1 study each. Fourth, this review has limited generalizability because participants in most studies were highly educated and self-selected; thus, they may have been highly motivated. Finally, interventions in some studies were targeted to exercise and/or physical activity as the main component of the program, and behavior change techniques were used to improve exercise adherence [78-81]; the multiplicity and heterogeneity of techniques prohibited meta-analysis of the effects of particular techniques. However, almost all studies employed theoretically driven interventions incorporating important components of the SMPs (ie, education, goal setting, action planning, problem solving, skills acquisition, self-monitoring, understanding illness, and managing emotions).

In conclusion, digital-based structured SMPs resulted in medium improvements in pain and physical function postintervention. However, the effects are below the MCID and may, therefore, not be clinically significant. The quality of the evidence for pain and function was graded as medium because of the high risk of bias in the studies; therefore, the true effect is likely to be close to our estimate of the effect, but there is a possibility that it is substantially different. The effect of digital-based structured SMP intervention on pain and physical function was slightly reduced at the 12-month follow-up, but remained at the medium threshold. One study reporting results on disability and one study reporting results on health-related QoL indicated small improvements in both outcomes. We rated the quality of the evidence as low and very low for disability and health-related QoL, respectively, indicating that the true effects are likely to be substantially different from our estimate of effect.

This review of digital-based structured SMPs on self-reported outcomes including pain, physical function, disability, and health-related QoL in patients with OA revealed 6 RCTs and 2 cluster RCTs. digital-based structured SMPs resulted in medium improvements in pain and physical function postintervention, but these may not be clinically significant. These effects were slightly reduced at the 12-month follow-up, but remained at the medium threshold. The quality of the evidence for pain and function was graded as medium because of a high risk of bias in some studies. More high-quality studies are needed, and the routine assessment of QoL and disability would be useful.