This systematic review and network meta-analysis were registered in the International Prospective Register of Systematic Reviews (PROSPERO; CRD420251169866). The design, conduct, and reporting of this systematic review followed the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) 2020 statement, the PRISMA extension for network meta-analyses (PRISMA-NMA), and the PRISMA literature search extension (PRISMA-S) for reporting literature searches in systematic reviews [21,22]. The complete PRISMA 2020, PRISMA-NMA, and PRISMA-S checklists are provided in Checklist 1.

A systematic search was conducted across 7 electronic databases: PubMed, Embase, Web of Science Core Collection, the Cochrane Central Register of Controlled Trials, CINAHL, MEDLINE, and Scopus. The search covered each database from inception to July 10, 2025. The search strategy combined controlled vocabulary terms and free-text keywords covering CMP, DHIs, mobile apps, virtual reality, wearable devices, telerehabilitation, and randomized controlled trials. Search strategies were adapted to each database’s indexing terms and syntax. The strategy was developed by the research team and validated against preidentified relevant studies to ensure sensitivity. An updated search was conducted on January 29, 2026, using identical databases, strategy, and eligibility criteria to identify newly published records. Full search strategies and results are provided in Multimedia Appendix 1. Reference lists of included studies and relevant systematic reviews were manually screened, and forward citation tracking was performed where appropriate. Gray literature databases and clinical trial registries were not searched separately.

All records were imported into EndNote 21 (Clarivate) for deduplication. Two reviewers (XW and WJ) independently screened titles, abstracts, and full texts. Disagreements were resolved through discussion or adjudication by a third reviewer (Y Liu). Multiple reports of the same trial were merged using the registration number, study center, recruitment period, sample characteristics, intervention details, and baseline data. The most complete report was used as the primary source.

Eligibility criteria were defined according to the PICOS (population, intervention, comparison, outcome, and study design) framework [23].

A standardized extraction form was developed and piloted. Two reviewers independently extracted data. Disagreements were resolved by discussion or third-party adjudication. Extracted data included study characteristics, participant characteristics, intervention and comparator details, outcome measures, and quantitative outcome data. For continuous outcomes, means, SDs, and sample sizes were extracted. Change scores were preferentially used; end point values were used when change scores were unavailable. Missing SDs were calculated from available statistics or imputed using recommended methods [25,26]. If key data could not be obtained from the original articles or supplementary materials or by contacting the authors and could not be reliably converted, the study was excluded from the quantitative synthesis for the corresponding outcome. For studies reporting multiple follow-up time points, data from the final assessment time point were extracted for analysis, and the timing of assessment was considered in assessing transitivity and in network meta-regression.

Outcomes were synthesized using standardized mean differences (SMDs) with Hedges g correction. All outcomes were aligned so that negative values indicated improvement in favor of DHIs. Scales with opposite directionality were reversed before analysis. Functional disability was defined solely by patient-reported functional outcomes. Objective performance measures and psychological constructs were not pooled with disability outcomes. Health-related quality of life was not quantitatively pooled due to heterogeneity and was synthesized narratively.

Risk of bias was assessed using the Cochrane Risk of Bias 2 (RoB 2) tool for randomized trials [27]. Two reviewers independently evaluated each outcome. Domains included randomization, deviations from intended interventions, missing data, outcome measurement, and selective reporting. Each domain was rated as low risk, some concerns, or high risk. Assessments were performed separately by outcome and time point. Disagreements were resolved by consensus or adjudication.

The certainty of evidence was assessed using the Grading of Recommendations Assessment, Development and Evaluation (GRADE) approach adapted for network meta-analysis [33-35]. For each outcome-specific intervention-vs-control network estimate, the standard GRADE domains of risk of bias, inconsistency, indirectness, imprecision, and publication bias were considered. Evidence from randomized controlled trials was initially rated as high certainty and was rated down by 1 or 2 levels for serious or very serious concerns, respectively. Final certainty was categorized as high, moderate, low, or very low. Two reviewers independently performed the assessments, and disagreements were resolved through consensus or consultation with a third reviewer.

The study selection process for this systematic review is presented in the PRISMA flow diagram (Figure 1). A total of 9235 records were identified through systematic database searching, including records retrieved from the initial search and the presubmission update. Of these, 1649 records were retrieved from PubMed, 1390 from Embase, 1415 from Web of Science, 726 from the Cochrane Library, 1852 from CINAHL, 982 from MEDLINE, and 1221 from Scopus. After duplicates were removed in EndNote, 3976 records were excluded, leaving 5259 for title and abstract screening. In the initial screening phase, 4628 records were excluded, including 2594 that did not meet the eligibility criteria for study design and 2034 that did not meet the inclusion criteria for the research topic.

Subsequently, 631 full-text articles were assessed for eligibility. Of these, 49 were excluded because the full text could not be retrieved. Other reasons for exclusion included ineligible population (n=174), ineligible intervention (n=131), irrelevant outcomes (n=83), inappropriate study design (n=51), and review articles (n=51). Ultimately, 92 articles were included in the systematic review.

A total of 92 randomized controlled trials were included in this systematic review, involving 12,595 participants [36-127]. The overall mean age of participants across studies was 51.33 years. The mean intervention duration was 9.15 (SD 6.10) weeks, ranging from 1 to 36 weeks. The included studies were conducted across 28 countries. Turkey [37,42,43,59,85,92,93,95-97,107,119] and the United States [46,48,49,56,63,69,80,81,101,103,118,125] contributed the highest number of studies (n=12 each), followed by Germany and Spain (n=9 each) [50,51,58,64,74,78,98,102,106,110,114-116,120-124]; Australia (n=8) [39,41,65,66,82,100,108,109]; South Korea (n=5) [71,72,75,76,99]; China and Iran (n=4 each) [36,47,60,67,79,111,113,127]; and the United Kingdom [52,55,112], Thailand [49,92,117], Japan [61,62,104], and Sweden [70,73,94] (n=3 each). In addition, 2 multicountry trials were included, one conducted in Egypt and Saudi Arabia [87] and another in Norway and Denmark [105]. Detailed study characteristics are presented in Multimedia Appendix 4.

All 92 trials reported pain outcomes, and 77 reported functional disability outcomes [36-51,53-56,58-69,71,72,74-77,80-86,88-90,92-103,105-110,112-114,116,117,119,120,122,123,125-127], which were included in the corresponding analyses according to available data. Among the included trials, 51 reported quality of life outcomes [36,38,40-43,48,50,51,53,57-59,62-66,68,70,71,74,76-79,82-86,91,92,94-96,99,104-106,108,110,113,114,117,119-121,123,126,127], of which 46 (90.2%) indicated improvements in quality of life [36,38,40,42,43,48,50,51,53,58,62-65,68,70,71,74,76-79,82-86,91,92,95,96,99,104-106,108,110,113,114,117,119-121,123,126,127], while 5 (9.8%) showed no clear advantage [41,57,59,66,94]. Detailed quality of life results are presented in Multimedia Appendix 5.

Risk of bias was assessed at the outcome level using the RoB 2 tool. Overall, of the 92 studies, 21 (22.8%) were judged to have a low risk of bias, 67 (72.8%) to have some concerns, and 4 (4.3%) to have a high risk of bias. A high risk of bias was primarily attributed to missing outcome data without appropriate analytical handling, deviations from intended interventions, and bias in outcome measurement. In some studies, multiple domains were rated as high risk. An overall summary is presented in Figure 2, and detailed RoB 2 judgments and supporting information are provided in Multimedia Appendix 6.

The certainty of evidence for the network estimates of pain intensity and functional disability was assessed using the GRADE approach adapted for network meta-analysis. Overall, the certainty of evidence was low or very low, with no comparisons rated as moderate or high certainty. Rating down was primarily attributable to risk of bias and imprecision, with heterogeneity and potential publication bias contributing to some assessments. Indirectness was not considered serious, and no serious incoherence between direct and indirect evidence was identified, although the ability to detect incoherence was limited for sparsely informed comparisons. Network estimates, judgments across the 5 standard GRADE domains, reasons for rating down, and final certainty ratings are presented in Table 1.

This systematic review and Bayesian network meta-analysis included 92 randomized controlled trials involving 12,595 participants with CMP. Overall, DHIs were associated with small to moderate average improvements in pain and functional disability. App-based, virtual reality–based, and telerehabilitation interventions showed more favorable average effects, whereas the estimates for wearable device and multimodal interventions were less precise. Subgroup analyses suggested that interventions incorporating exercise or motor function training tended to produce larger effect estimates than those centered primarily on education or cognitive behavioral therapy. However, several subgroups contained relatively few studies; therefore, these patterns should be considered exploratory rather than evidence of component-level superiority. In the network meta-analysis, motor function–oriented virtual reality and exercise-based telerehabilitation showed potentially favorable effects across pain and disability outcomes. Some attention- or cognition-oriented interventions ranked relatively highly for pain, but their effects on functional disability were less certain.

In the pairwise meta-analyses, the pooled estimates and their 95% CIs represent average effects across the included studies, whereas the 95% prediction intervals indicate the range of true effects that might be expected in a new, broadly comparable study or clinical setting. App-based, virtual reality–based, and telerehabilitation interventions showed favorable average effects on pain and disability, with 95% CIs excluding the null. However, the corresponding prediction intervals crossed the null, indicating that effects may be smaller, absent, or potentially different across populations, intervention protocols, and health care settings. Prediction intervals therefore describe variation in study-level effects across settings rather than variation in individual patient responses [128].

Variation in effects was observed across intervention modalities and outcomes. Accordingly, I² values were interpreted descriptively rather than being used to classify the magnitude of heterogeneity, with greater emphasis placed on τ² and prediction intervals. This variation may reflect differences in musculoskeletal conditions, baseline severity, intervention intensity and duration, professional supervision, comparator content, and assessment timing. The included interventions also varied in their structure and implementation. App-based exercise programs differed in training intensity, frequency, and feedback delivery [49,55,68]; virtual reality interventions differed in immersion, therapeutic content, and interaction modality [60,85,116]; and telerehabilitation programs varied in supervision, real-time interaction, and follow-up frequency [65,82]. Although leave-one-out analyses did not materially alter the direction of the pooled estimates, robustness to the exclusion of individual studies does not rule out variation in the magnitude or applicability of effects across settings.

In the Bayesian network meta-analysis, 95% CrIs describe uncertainty around the average relative effects between intervention nodes. Although several nodes had 95% CrIs excluding the null, some estimates remained imprecise and several comparisons were supported by limited direct evidence. Moreover, the absence of statistically detectable inconsistency between direct and indirect evidence does not eliminate heterogeneity or ensure stable treatment rankings. SUCRA values summarize relative ranking probabilities and should therefore be interpreted together with the corresponding effect estimates, 95% CrIs, network structure, and certainty of evidence rather than as a definitive clinical hierarchy [129]. Consequently, the relatively high rankings of some virtual reality– and exercise-based interventions suggest that they may warrant further evaluation but do not establish their superiority over other digital interventions. These findings should also be interpreted in light of risk of bias and certainty of evidence. Overall, 72.8% of the included studies were judged to have some concerns using RoB 2. Pain and functional disability were predominantly participant reported, and participant blinding was generally infeasible for digital behavioral and rehabilitation interventions. Awareness of treatment allocation may therefore have influenced participant expectations, engagement, or outcome reporting [130]. Together with heterogeneity and imprecision, these considerations contributed to the predominantly low or very low certainty ratings under the GRADE approach adapted for network meta-analysis [33,131]. Accordingly, the favorable average effects should be interpreted as suggesting potential benefits rather than as establishing clinically important improvement or comparative superiority [132]. Network meta-regression suggested a potential study-level association between intervention duration and functional disability, whereas no clear association was observed for pain. This finding may be consistent with functional improvement requiring sustained training, but it should not be interpreted as evidence of a causal mechanism because intervention duration may be associated with treatment intensity, adherence, supervision, or other study characteristics. Overall, by jointly considering digital delivery modality and core therapeutic content, this systematic review provides a more clinically interpretable framework for comparing digital interventions. The findings suggest that therapeutic content and delivery format should be considered together when designing interventions and selecting candidates for future head-to-head trials.

Digital health modalities may exert their effects through multiple pathways, including behavioral regulation, cognitive reinforcement, and motor learning [133]. App-based and telerehabilitation interventions may facilitate self-management by delivering structured exercise and supporting goal setting, feedback, and professional contact [134]. The mechanisms of virtual reality may vary according to its therapeutic purpose: distraction-based virtual reality may modulate pain perception through attentional distraction, whereas motor function–oriented virtual reality may promote graded movement and motor relearning [135]. Wearable devices primarily enable behavioral monitoring and feedback [136], whereas multimodal interventions integrate multiple delivery formats or therapeutic components and vary substantially in composition [137]. Taken together, these differences suggest that the effects of digital interventions may be shaped jointly by the delivery modality and therapeutic content. Previous reviews have addressed complementary but more focused research questions. Some reviews synthesized the overall effects of DHIs, whereas others focused on mobile health, virtual reality, remote rehabilitation, or economic outcomes [11,12,14,16]. An earlier network meta-analysis compared electronic delivery modalities but did not differentiate the therapeutic components delivered within each modality [20]. Thus, previous reviews support the potential value of digital care, but comparative evidence across a broad range of delivery modality–therapeutic component combinations remains limited.

This systematic review extends the existing evidence by synthesizing 92 randomized controlled trials within a 2-level framework integrating digital delivery modality and core therapeutic content. Pairwise meta-analyses estimated effects at the broader modality level, whereas the Bayesian network meta-analysis compared 12 active intervention nodes defined by specific delivery modality–therapeutic component combinations. This approach distinguishes interventions delivered through the same technology but incorporating different therapeutic components while enabling comparisons of similar components delivered through different platforms. Rather than isolating the effects of individual components, the framework compares clinically meaningful delivery modality–therapeutic component combinations as implemented in the included trials. It therefore provides a more clinically interpretable classification, may improve the interpretation of within-modality heterogeneity, and identifies combinations that warrant further evaluation in head-to-head trials.

This systematic review suggests that DHIs may have clinical value in the management of CMP, although their effects may vary across populations, intervention protocols, and health care settings. Clinical decisions should therefore consider the therapeutic content delivered as well as the digital platform, rather than relying on effect estimates or treatment rankings alone. App-based and telerehabilitation interventions may offer flexible delivery, but their implementation depends on sustained engagement, digital literacy, access to suitable devices and internet services, professional support, and integration into clinical workflows [138,139]. Virtual reality may be useful for selected patients but may require additional equipment, technical support, training, and implementation resources [140]. Because adherence, costs, and implementation outcomes were not consistently reported across the included trials, the relative feasibility and cost-effectiveness of these approaches remain uncertain. In practice, intervention selection should therefore be guided by patient preferences, accessibility, supervision needs, and locally available resources. Future head-to-head trials should clearly report the digital platform, core therapeutic components, intervention dose, professional involvement, adherence, adverse events, resource use, and implementation outcomes [141]. Standardized outcome measures and prespecified follow-up periods are also needed to improve comparability and determine whether effects are sustained over time [142].

This systematic review has several strengths, including the joint classification of DHIs by delivery modality and therapeutic content, the complementary use of pairwise and Bayesian network meta-analyses, and the assessment of risk of bias and certainty of evidence using RoB 2 and the GRADE approach adapted for network meta-analysis, respectively.

Several limitations should nevertheless be acknowledged. First, clinical and methodological variation across studies, including differences in musculoskeletal conditions, intervention protocols, comparators, and assessment timing, may have contributed to heterogeneity and influenced some indirect comparisons [46,77]. Second, pain and disability were predominantly participant reported, and participant blinding was generally infeasible. Potential expectation or reporting effects, together with heterogeneity and imprecision, contributed to the generally low or very low certainty of evidence. Third, the final assessment was selected to evaluate the longest available effect while avoiding correlated estimates; although follow-up duration varied, its potential influence was explored through network meta-regression. Fourth, some nodes and active treatment comparisons were supported by relatively few studies, resulting in imprecise estimates and uncertainty in SUCRA rankings [132]. Rankings should therefore be interpreted alongside the effect estimates, uncertainty intervals, network structure, and GRADE certainty ratings [143]. Fifth, the use of SMDs facilitated the synthesis of different outcome measures but limited direct interpretation against a common threshold for clinical importance. Finally, gray literature and trial registries were not systematically searched; therefore, publication bias cannot be excluded, particularly for comparisons supported by few studies.

This systematic review and Bayesian network meta-analysis of 92 randomized controlled trials suggests that DHIs may provide small to moderate average improvements in pain and functional disability among people with CMP. Interventions incorporating exercise or motor function training through apps, virtual reality, or telerehabilitation showed potentially favorable patterns across outcomes. However, between-study variability in effect estimates, some risk-of-bias concerns, prediction intervals crossing the null for many comparisons, and predominantly low or very low certainty of evidence limit confidence in the comparative estimates. The SUCRA rankings should therefore be regarded as exploratory and should not be interpreted as establishing the superiority of any intervention. The principal innovation of this systematic review is the joint classification of interventions according to digital delivery modality and core therapeutic content. Unlike previous reviews that primarily evaluated overall effects, individual platforms, or broad delivery modalities, this framework compares clinically meaningful modality-component combinations within a common analytical structure. It provides the field with a more clinically interpretable classification and comparative evidence base for intervention design and future head-to-head research. In real-world practice, the framework may help clinicians and health care planners consider therapeutic content alongside accessibility, digital literacy, adherence, supervision requirements, costs, and available resources when selecting digital interventions. Adequately powered head-to-head trials with standardized outcomes and implementation and economic evaluations are needed to confirm comparative effectiveness and determine whether the observed benefits are sustained across health care settings.