This systematic review and meta-analysis was conducted and reported in accordance with the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) 2020 statement [15] and the PRISMA-S extension for literature searches. An expanded PRISMA 2020 checklist, the PRISMA 2020 checklist for abstracts, and the PRISMA-S checklist [16] are provided in Checklist 1. The structure of the Methods section follows the corresponding PRISMA 2020 headings to enhance transparency, reproducibility, and suitability for editorial review. Protocol changes are detailed in Multimedia Appendix 1.

A comprehensive literature search was performed in PubMed (National Library of Medicine), Embase (Elsevier), the Cochrane Library (Wiley), and the Web of Science (Clarivate) from database inception to October 15, 2024, with an updated search conducted on December 10, 2025. The search strategy was independently developed de novo and refined by 2 reviewers to identify RCTs comparing digitally supported Tele-PR or low-technology HBPR with CBPR in adults with COPD. Controlled vocabulary terms (eg, MeSH and Emtree) and free-text keywords were combined using Boolean operators. Searches were limited to articles published in English, but no restrictions were applied based on publication date or publication status (Multimedia Appendix 2). Although specific gray literature databases and clinical trial registries were not searched, we manually screened the reference lists of included studies and relevant systematic reviews to identify additional eligible trials. Furthermore, corresponding authors were contacted to request missing or unpublished data where necessary; however, no additional data were obtained. The reasons for exclusion at the full-text screening stage are reported in Multimedia Appendix 3.

Study eligibility was defined according to the PICOS (population, intervention, comparator, outcomes, and study design) framework. Eligible studies were RCTs enrolling adult participants (aged ≥18 y) with COPD diagnosed according to the Global Initiative for Obstructive Lung Disease criteria (postbronchodilator forced expiratory volume in 1 s [FEV₁]/forced vital capacity of <0.70), regardless of Global Initiative for Obstructive Lung Disease stage (I to IV) or exacerbation history. Tele-PR interventions were eligible if they could be classified a priori into one of 2 conceptually distinct models based on technological infrastructure and supervision intensity. Digitally supported Tele-PR was defined as any comprehensive PR program delivered primarily in the home setting and supported by information and communication technology, enabling real-time or asynchronous interaction with health professionals (eg, synchronous videoconferencing, mobile apps, web-based platforms, or wearable-supported monitoring with structured feedback), while HBPR encompassed lower-technology home programs relying primarily on structured telephone coaching and/or printed training materials, provided that systematic professional guidance or supervision was delivered (ie, not fully unsupervised self-training). To harmonize heterogeneous Tele-PR implementations, we used a prespecified hierarchical taxonomy: delivery models (digitally supported Tele-PR vs low-technology HBPR), supervision modality (synchronous real-time supervised sessions vs asynchronous delayed feedback or periodic contacts), and supervision intensity (high: ≥2 structured contacts per week; medium: ~1 per week; and low or minimal: <1 per week or instruction only). These categories were defined based on trial descriptions and were used for subgroup and exploratory analyses. Meanwhile, the comparator was CBPR, which was defined as any standardized PR program delivered in hospital or community settings with direct, in-person professional supervision. Studies were required to report at least one prespecified outcome (eg, 6-min walk distance [6MWD], COPD Assessment Test [CAT], St. George’s Respiratory Questionnaire [SGRQ], Chronic Respiratory Questionnaire, modified Medical Research Council Dyspnea scale, Hospital Anxiety and Depression Scale, daily step count, self-efficacy, dropout, adverse events, or exacerbations). Outcomes reported by only a small number of studies were synthesized descriptively or exploratorily.

Studies were excluded if they (1) used nonrandomized designs (eg, observational studies); (2) were reviews, case series, editorials, protocols, or conference abstracts without accessible full texts; (3) enrolled participants with contraindications to exercise due to severe cardiovascular, musculoskeletal, or other limiting comorbidities; or (4) evaluated interventions not primarily designed for COPD rehabilitation or involved mixed programs in which the specific effects of Tele-PR or HBPR could not be isolated.

Study selection and data extraction were performed independently by 2 teams using a standardized workflow. Duplicates were removed in EndNote 21 (Clarivate), after which titles or abstracts and full texts were screened in duplicate. Interrater agreement was assessed using Cohen κ [17], indicating substantial agreement for title or abstract screening (κ=0.75) and almost perfect agreement for full-text eligibility assessment (κ=0.92). Disagreements were resolved through discussion with a third reviewer. Backward citation tracking of included studies was conducted as a supplementary search step after full-text screening. Key study characteristics (eg, first author, year, sample size, design, participant characteristics, intervention details, and outcomes) were extracted in duplicate; missing data were requested from corresponding authors when necessary, and discrepancies were resolved through discussion with a third researcher.

We pooled dichotomous outcomes using risk ratios and continuous outcomes using mean differences (MDs), as outcomes were measured on consistent scales. Meta-analyses were conducted in R (version 4.3.1; R Foundation for Statistical Computing) using the meta package (version 7.0‐0). Given the anticipated clinical and methodological heterogeneity across interventions, populations, and delivery models, all meta-analyses were conducted using random effects models, irrespective of the magnitude of statistical heterogeneity. This choice was conceptual rather than data-driven and reflects the assumption that true effects may vary across settings. Pooled effect estimates were calculated using the Hartung-Knapp-Sidik-Jonkman (HKSJ) method, which provides more robust CIs and reduces the risk of false-positive findings, particularly when the number of included studies is small [18]. Between-study variance (τ²) was estimated using the Sidik-Jonkman estimator. When outcomes were reported at multiple time points, separate meta-analyses were conducted for the end-of-intervention and long-term follow-up data (≥6 mo), as these time points address distinct clinical questions regarding short-term efficacy and durability of effects. For the primary meta-analyses, 95% PIs were calculated using the confidence distribution approach proposed by Nagashima et al [19] via the pimeta package. This approach accounts for uncertainty in the estimation of between-study variance and improves coverage probability compared with conventional normal-based methods [20]. While I² values are reported descriptively, interpretation of heterogeneity primarily emphasized PIs, given their greater relevance for real-world applicability. Meta-analyses were not conducted when fewer than 3 studies were available, as estimation of between-study heterogeneity and prediction intervals (PIs) would be statistically unreliable; such outcomes were synthesized narratively. To improve numerical stability for PI estimation, step counts were rescaled (per 100 steps per day) without changing standardized interpretation. Where appropriate (k≥10 studies), we assessed small-study effects using contour-enhanced funnel plots, Egger test (for continuous outcomes), and Peters linear regression test (for binary outcomes), interpreting these as indicators of small-study effects rather than sole proof of publication bias [21,22]. To explore potential sources of heterogeneity, meta-regressions for continuous moderators were visualized using bubble plots, whereas subgroup analyses for categorical moderators were summarized in tabular format to provide detailed group-specific estimates. These analyses are inherently observational and explore potential moderators rather than establish causality [23]. Leave-one-out sensitivity analyses were conducted for all primary outcomes to assess the robustness of pooled estimates and to identify influential studies contributing to between-study heterogeneity. Multiple reports from the same trial were treated as a single study unit, and no study contributed more than one effect size to the same meta-analysis. For multiarm trials, only one eligible comparison per study was included, preventing the duplicate use of shared control groups [24].

Methodological quality was assessed using the Cochrane Risk of Bias 2 tool [25]. Two reviewers independently evaluated 5 domains: randomization, deviations from intended interventions, missing outcome data, outcomes measurement, and selective reporting. Each domain was rated as low risk, some concerns, or high risk. Disagreements were resolved through discussion or consultation with a third reviewer. The certainty of evidence for primary outcomes was graded using the GRADE (Grading of Recommendations, Assessment, Development, and Evaluation) approach [26], considering risk of bias, inconsistency, indirectness, imprecision, and publication bias.

As this study is a systematic review and meta-analysis of previously published data, it did not involve the recruitment of human participants or access to identifiable private patient information. Consequently, institutional review board or research ethics board approval was not required. Ethical approval and informed consent were obtained by the authors of the primary studies included in this review.

A total of 8847 records were retrieved in the initial search. After stepwise screening, 17 studies were included. The selection process is presented in a PRISMA flow diagram (Figure 1).

A total of 17 trials [27-43] were identified and included in the quantitative synthesis. To ensure statistical independence, multiple reports originating from the same trial were linked via trial registration numbers, and data were extracted from a single source per outcome to avoid participant double-counting (Multimedia Appendix 4 for a detailed audit of study independence). The included trials were published between 2008 [36] and 2025 [43]. The trials were conducted across multiple regions, including Europe (the United Kingdom [27,34,35,39], Spain [29,37,43], Denmark [30,31], and Greece [40]), Australia [28,33,38,41], Asia (China [32]), North America (Canada [36]), and South America (Brazil [42]). Sample sizes ranged from 54 participants [29] to more than 150 participants [39]. All participants were adults with a confirmed diagnosis of COPD. One trial [33] enrolled patients with chronic respiratory diseases; however, in this meta-analysis, only data from its COPD subgroup were extracted. The most studies included a higher proportion of male participants. Regarding disease severity, participants predominantly had moderate-to-severe airflow limitations. Tele-PR interventions were classified into 3 distinct models: synchronous video supervision, in which direct remote supervision was provided by physiotherapists via real-time videoconferencing [30,31,33]; asynchronous digital platform support, in which structured rehabilitation content was delivered through web-based platforms [27,34,40], mobile apps, or wearable sensors [29,32], with supervision primarily provided as data-driven asynchronous feedback; and low-technology and telephone support [28,35,36,38,39,42,43], which relied primarily on printed materials for self-management guidance and was supplemented by telephone follow-up to enhance adherence and motivation. A minimal supervised model was also reported [37]. All trials compared Tele-PR with CBPR. The intervention duration was typically 8 to 10 weeks [29,31,33,36,38,41], with some studies lasting 7 weeks [27,28,39] or extending to 12 to 13 weeks [42,43]. Several trials reported long-term follow-up data at 6 to 12 months after the intervention [30,32,33,40,41,43]. Detailed characteristics of the included RCTs are summarized in Table 1. Intervention components and supervision models are detailed in Multimedia Appendix 5.

The methodological quality of the included RCTs varied considerably (Figure 2). A total of 4 RCTs were rated as low risk of bias [28,33,36,41], while the remaining studies were classified as high risk or as having some concerns, mainly due to the analysis strategy they used (domain 2) or missing outcomes data (domain 3). For domain 1 (randomization process), risk was generally low, as most studies reported adequate sequence generation and allocation concealment; however, concerns were raised in a study by Cerdán-de-las-Heras et al [29] due to substantial baseline imbalance in a small sample. For domain 2 (deviations from intended interventions), risk depended on the type of analysis because blinding was difficult to implement. Studies using a strict intention-to-treat analysis were rated low risk [28,29,31,33,36,40,41,43], whereas studies relying on per-protocol analyses or completer-only analyses were judged to exhibit high risk [34,35,37-39,42]. Domain 3 (missing outcomes data) was the primary source of bias. High-risk ratings resulted from excessive dropout rates [27,39] or differential dropout between groups [31,37]. Long-term data in a study by Hansen et al [31] and Sacristán-Galisteo et al [43] were affected by substantial missingness, whereas only Cox et al [33] and Vasilopoulou et al [40] showed minimal missing data. For domain 4 (measurement of the outcome), risk was predominantly low, as blinded assessors were used for objective outcomes in most studies; Li et al [32] was an exception, as the absence of assessor blinding may have introduced detection bias. For domain 5 (selection of the reported result), risk was very low, and, in nearly all studies, reported outcomes were consistent with preregistered protocols or trial registry records. The risk of bias assessment of the included RCTs is provided in Multimedia Appendix 6.

Subgroup analyses were conducted according to delivery models, supervision intensity, and supervision modality (Multimedia Appendix 7). For 6MWD at both the end of intervention and long-term follow-up (≥6 mo), no statistically significant subgroup differences were observed across delivery models, supervision intensity, or supervision modality (all P>.05), although moderate heterogeneity was noted in some low-technology HBPR strata. For daily steps and Chronic Respiratory Questionnaire–Dyspnea (CRQ-D) domain, no significant subgroup effects by delivery models were identified at either time point (all P>.05). In contrast, for the CAT, delivery models showed significant subgroup differences at both end of intervention and long-term follow-up (both P<.001), with digitally supported Tele-PR demonstrating small, nonsignificant changes, whereas low-technology HBPR showed worsening CAT scores in single-study strata. For the dropout rate, no significant subgroup differences were detected by delivery models or supervision modality (all P>.05), although heterogeneity was substantial in several subgroups. Overall, subgroup effects were generally limited, except for CAT, and several strata were constrained by small numbers of studies.

Leave-one-out sensitivity analyses were conducted for all primary outcomes to assess the robustness of pooled estimates and to identify influential studies contributing to between-study heterogeneity (Multimedia Appendix 8). Overall, pooled estimates for several outcomes showed sensitivity to the exclusion of individual studies, indicating that some findings were influenced by specific trial characteristics.

For 6MWD, exclusion of a single telephone-based study [40] resulted in a shift of the pooled estimate toward CBPR (MD –8.63 m; P=.045), indicating that the overall effect estimate was sensitive to this study. For daily step counts, the marginally significant pooled effect was driven primarily by 1 study [35]; removal of this study rendered the pooled effect nonsignificant (P=.12) and substantially reduced heterogeneity.

For patient-reported outcomes, substantial heterogeneity in health status measured by the CAT was attributable to a single study [43]; exclusion of this study eliminated heterogeneity and attenuated the pooled effect. Similarly, heterogeneity in self-efficacy outcomes was largely driven by 1 study [39], and removal of this study reduced between-study variance and altered the pooled estimate.

In contrast, pooled estimates for modified Medical Research Council Dyspnea score, depressive symptoms, and dropout rate were robust to leave-one-out analyses. Although individual studies [34,41] showed effects in opposite directions for dropout, the overall conclusion of no statistically significant difference between groups remained unchanged across all sensitivity analyses.

Visual inspection of the funnel plot for dropout rate revealed a relatively symmetrical distribution of studies (Figure 24). This was confirmed by Peters test, which showed no evidence of significant publication bias or small-study effects (t(8)=.43; P=.68). However, given the limited number of included studies, the results of publication bias tests should be interpreted cautiously.

The certainty of evidence ranged from moderate to very low across outcomes. Evidence for 6MWD was rated as low certainty and was downgraded by 1 level for risk of bias because blinding was not feasible and attrition occurred in several trials. Evidence for dropout rate was rated as very low certainty and was downgraded for risk of bias, substantial inconsistency, and imprecision, reflected by wide CIs and PIs. Health status outcomes (CAT, SGRQ, and Chronic Respiratory Questionnaire–Dyspnea) were rated as low certainty and were downgraded for inconsistency, likely related to heterogeneous intervention modalities, and for imprecision. These certainty ratings should temper interpretation of pooled estimates and underscore variability in real-world effects. The certainty of evidence assessed using the GRADE approach is summarized in Table 5.

This systematic review and meta-analysis evaluated the relative effectiveness of Tele-PR compared to CBPR in patients with COPD. Overall, no statistically significant differences were observed at the end of the intervention across key outcomes, including functional capacity, dyspnea, and health-related quality of life. These findings indicate that, on average, Tele-PR may achieve short-term clinical effects similar to those of CBPR when implemented under structured and well-defined conditions. However, subsequent subgroup analyses and investigations of heterogeneity suggest that Tele-PR should not be regarded as a single homogeneous intervention [33]. Substantial variation was observed in effect consistency and stability across different remote delivery models, particularly with respect to supervision intensity and interaction modality. Accordingly, interpretations of “equivalence” between Tele-PR and CBPR based solely on pooled average effects should be made with caution, as such summaries may obscure meaningful differences across implementation contexts [19].

In this analysis, differences in outcomes between rehabilitation modalities diminished during postintervention follow-up, indicating a time-dependent convergence. This pattern suggests that, regardless of delivery format, remote and center-based PR programs face challenges in sustaining long-term benefits [3]. Several factors may contribute to this attenuation of effects, although these potential explanatory variables were not directly assessed in this meta-analysis. First, adherence to regular exercise training often declines after structured supervision ends, particularly when ongoing support is limited [44,45]. Second, longer follow-up periods may include seasonal variations, acute exacerbations, or intercurrent illnesses, which can disrupt training continuity and negatively impact functional outcomes [46,47]. Third, many existing Tele-PR programs primarily focus on exercise training, while systematic reinforcement of relapse prevention and long-term self-management skills remains relatively underdeveloped [48]. Importantly, this convergence of outcomes does not imply that early rehabilitation effects lack clinical significance. Rather, it underscores the need to conceptualize PR—whether delivered remotely or in person—as an ongoing process that requires sustained behavioral and educational support, rather than as a time-limited intervention. This perspective aligns with the concept of maintenance rehabilitation emphasized in recent clinical guidelines and supports the future integration of educational components, behavioral maintenance strategies, and digital platforms to improve long-term effectiveness [49-51].

The included studies predominantly enrolled patients with moderate-to-severe COPD, a population in whom home-based training has traditionally raised safety concerns [3]. Evidence suggests that, under appropriate screening and standardized implementation, Tele-PR was not associated with increased short-term serious adverse events and demonstrated a safety profile comparable to that of CBPR [31,33,36,39,41]. Functional improvements likely extend beyond aerobic enhancement, as PR improves respiratory mechanics and ventilatory efficiency [14,52-54]. Although physiological markers were not directly synthesized, dyspnea improvements in Tele-PR suggest overlapping mechanisms with conventional PR [55]. Emerging evidence of improved gas exchange in hypercapnic patients undergoing PR further supports integrating refined physiological monitoring in higher-risk subgroups [56].

An important observation from this review is that supervision intensity may contribute to heterogeneity in the effectiveness of remote rehabilitation. Subgroup analyses [31,33] indicated that synchronous, video-supervised programs demonstrated lower statistical heterogeneity in functional capacity outcomes, whereas low-supervision or asynchronous approaches exhibited greater variability. These findings raise the possibility that, beyond training dose alone, real-time interaction may support psychological safety, perceived support, and confidence during exercise, which in turn may influence engagement and adherence [48,57]. Conversely, insufficient supervision in home-based programs may heighten concerns regarding exercise-related risk and limit participation [58,59]. However, these interpretations should be approached cautiously. Given the exploratory nature of subgroup analyses and the limited number of studies within specific supervision categories, the independent contributions of psychological factors, adherence, and training intensity cannot be clearly disentangled. Future trials should prospectively incorporate measures of psychological safety, self-efficacy, and fear of exercise to clarify the pathways through which supervision modalities influence outcomes.

The relative effectiveness of Tele-PR was closely linked to the characteristics of the comparator intervention. When compared to lower-intensity programs or usual care, Tele-PR was more likely to demonstrate a favorable relative effect [60]; however, when compared to high-intensity, face-to-face supervised rehabilitation, these advantages were substantially attenuated [43]. These findings suggest that conclusions regarding the equivalence of Tele-PR should be interpreted as context dependent rather than as intrinsic or universally applicable [61]. Accordingly, Tele-PR is best viewed as complementary to, rather than as a universal substitute for, high-quality center-based rehabilitation, with optimal use determined by patient characteristics, resource availability, and service capacity [62].

A notable gap was observed between improvements in functional capacity and changes in daily physical activity. Although gains in exercise capacity were observed under certain remote rehabilitation models, corresponding increases in daily step counts were not consistently demonstrated and were sensitive to the influence of individual studies [35]. This pattern suggests that functional improvement alone may be insufficient to produce sustained behavioral change [63]. Future Tele-PR strategies may therefore need to incorporate structured behavioral interventions alongside monitoring and feedback mechanisms. Digital literacy and technology access remain implementation barriers [64,65], and simpler, low-technology models with strong human support may achieve higher adherence in certain populations [41]. Platforms should prioritize usability and minimize cognitive burden to avoid undermining effectiveness [66,67].

Although this review primarily focused on clinical outcomes, economic considerations are critical for the sustainable implementation of remote rehabilitation. Available evidence [32,40,68,69] suggests that Tele-PR is broadly comparable to CBPR in terms of direct program costs and may offer potential long-term economic benefits by improving accessibility and intervention completion rates, consistent with findings from recent cost-utility evaluations of digital therapeutics for PR. However, cost-effectiveness evidence remains heterogeneous [70], underscoring the need for standardized economic evaluations in future trials.

Previous systematic reviews of PR and telerehabilitation in COPD have largely focused on estimating average between-group differences, reporting minimal clinically important difference achievement, or comparing remote and center-based delivery formats [71,72], while foundational Cochrane work established the overall efficacy of PR compared with usual care [73]. Few syntheses have systematically examined supervision-related heterogeneity or incorporated PIs to estimate expected effect ranges across settings. We observed substantial variation in program structure, supervision models, and comparator intensity, complicating the interpretation of pooled averages. By integrating supervision-based stratification with PI-informed interpretation, this review provides a context-sensitive framework for evaluating Tele-PR relative to CBPR. To our knowledge, it is among the first to frame Tele-PR equivalence as context dependent rather than model intrinsic, supporting tailored implementation in real-world health care systems.

Several limitations should be acknowledged. First, most included studies were nonblinded, and implementation bias could not be fully avoided. Second, the number of long-term follow-up studies was limited, resulting in insufficient statistical power for certain long-term outcome analyses. Third, definitions of “usual care” and rehabilitation intensity varied across studies, potentially introducing residual confounding. In addition, several subgroup analyses were based on a small number of studies and should therefore be interpreted as exploratory rather than definitive, particularly where a subgroup was informed by a single trial. The term “noninferiority” as used in this study refers to the overall distribution of effect directions and their 95% CIs, rather than a prespecified noninferiority margin; thus, it should be interpreted as indicating clinical equivalence rather than constituting a strict statistical test of noninferiority. Consistent with the PRISMA-S reporting framework, the omission of trial registry and gray literature searches may have resulted in the underrepresentation of ongoing, unpublished, or null-result trials, potentially contributing to residual publication bias despite comprehensive database coverage. Furthermore, the small number of studies available for several outcomes precluded formal risk of bias–based sensitivity analyses, which should be taken into account when assessing the robustness of the findings. The clinical implications outlined earlier are primarily based on evidence of moderate to low certainty and should be applied cautiously in real-world settings, with due consideration of local resources, patient characteristics, and operational conditions. Subgroup and meta-regression analyses were exploratory and intended to generate hypotheses rather than confirm model superiority.

Distinct from prior reviews that treated Tele-PR as a homogeneous intervention, we stratified remote programs by supervision intensity and delivery models and proposed a structured “supervision gradient” framework to interpret model-dependent consistency of effects, with practical implications for designing Tele-PR as a supervision-structured service incorporating real-time oversight and long-term behavioral maintenance support. Tele-PR is a feasible option for delivering PR to adults with COPD and, on average, yields short-term clinical outcomes comparable to those of CBPR. As effects varied across studies and the certainty of the evidence was low to very low, implementation should be context aware and model specific. Programs incorporating structured real-time professional supervision appear to produce more consistent outcomes, whereas low-technology HBPR may require enhanced behavioral support to achieve stable symptom control and improvements in daily activity levels. Longer-term maintenance remains a challenge, indicating that remote rehabilitation should be designed as an ongoing service rather than a time-limited intervention. Accordingly, Tele-PR may be particularly valuable for expanding access to PR and addressing existing accessibility gaps, while CBPR remains essential for patients who require close in-person supervision or complex multidisciplinary care.