We systematically searched the PubMed, Web of Science, Embase, Cochrane Library, and CINAHL databases from their inception until October 28, 2025. This cutoff date was selected as the latest point prior to the commencement of data extraction and statistical analysis to ensure the inclusion of the most up-to-date evidence available at that time. The search strategy utilized a combination of Medical Subject Headings or MeSH and free-text terms covering three core themes, namely chronic obstructive pulmonary disease, digital health interventions, and randomized controlled trials. Detailed search strategies for each database are provided in Supplementary material 1.

Inclusion and exclusion criteria were established strictly following the PICOS framework. The study population was limited to adult patients aged 18 years or older with a confirmed diagnosis of chronic obstructive pulmonary disease. Interventions focused on digital health interventions. To precisely evaluate the heterogeneity of different technological modalities, interventions were explicitly categorized as App + Sensor involving the combination of applications and physiological sensors for multidimensional management, wearables only relying solely on asynchronous monitoring, and Synchronous Tele-rehabilitation simulating center-based interaction via videoconferencing systems. Comparison groups included any of the aforementioned digital intervention modalities, Center-based PR as the gold standard reference, or Usual Care. Outcome measures were required to include quantitative data for objectively measured daily physical activity or daily steps, or functional exercise capacity measured by the 6-min walk distance. Study design was restricted to peer-reviewed randomized controlled trials. Non-randomized trials, studies with unavailable data, and pure remote monitoring studies lacking a physical activity promotion component were excluded.

To resolve potential conceptual overlaps between technological modalities and to define their roles as distinct pulmonary rehabilitation delivery models, this study established classification decision rules based on Synchronicity of Supervision and Complexity of Feedback Loop as shown in Table 1. First, Synchronous Tele-rehabilitation was defined as a direct remote extension of center-based PR with real-time synchronous supervision as its core feature. Patients trained under the real-time guidance of professionals via videoconferencing systems such as Zoom. This modality fully preserves and migrates core PR elements, including high-intensity exercise, real-time error correction, and interactive education, to the home setting. Second, Composite Digital Interventions or App + Sensor were defined as a form of asynchronously managed structured PR. Unlike tele-rehabilitation, interactions in this modality are primarily asynchronous or automated, focusing on utilizing smartphone applications combined with physiological sensors to provide feedback via preset algorithms or delayed clinician review such as weekly reports. This modality aims to provide structured self-management support and behavior change interventions through digital means. In contrast, Wearables Only was regarded as an activity-focused lightweight alternative. This modality does not include complex educational modules or interactive platforms and relies solely on off-the-shelf wearable devices such as pedometers to provide immediate biofeedback and promote daily physical activity through continuous objective monitoring. Finally, for hybrid interventions involving multiple technologies, we established the Highest Interaction Level Priority principle. For instance, if an intervention utilized an application while also providing regular real-time video guidance, it was prioritized as Synchronous Tele-rehabilitation based on its higher intensity of supervision.

Following literature screening, two researchers independently extracted data using a standardized Excel form. Extracted information included baseline study characteristics such as first author, publication year, sample size, and age, specific intervention details including device types, exercise prescriptions, and supervision methods used to determine node classification, and endpoint data for designated outcomes including daily steps and 6-min walk distance, reported as means, standard deviations, and sample sizes. For daily steps, priority was given to extracting the average daily steps over at least 3 to 7 days to minimize single-day measurement error and ensure consistency in definitions across studies. Given that the standard duration of chronic obstructive pulmonary disease rehabilitation interventions is typically 8 to 12 weeks, and previous pairwise meta-analyses have emphasized significant fluctuations and uncertainty in the efficacy of digital interventions at different follow-up points (20), this study prioritized data extracted immediately post-intervention or at the point closest to 12 weeks. This strategy aimed to maximize the number of included studies to construct an evidence network with maximum connectivity and robust transitivity for accurately assessing the relative ranking of different intervention modalities within the optimal treatment window. In cases of inconsistent data reporting, statistical conversion methods were used to estimate means and standard deviations from medians, interquartile ranges, or confidence intervals. For missing data, efforts were made to contact the corresponding authors. All extracted data were cross-checked, and any discrepancies were resolved through discussion or third-party adjudication to ensure data accuracy and integrity.

The risk of bias for included studies was independently assessed by two researchers using the Cochrane Risk of Bias tool 2.0 (ROB 2.0). This tool evaluates bias across five domains, namely randomization process, deviations from intended interventions, missing outcome data, measurement of the outcome, and selection of the reported result. The overall risk for each study was categorized as low risk, some concerns, or high risk. Furthermore, the Grading of Recommendations Assessment, Development and Evaluation (GRADE) framework was used to grade the certainty of evidence for primary outcomes of the network meta-analysis. By comprehensively considering five dimensions including risk of bias, inconsistency, indirectness, imprecision, and publication bias, the quality of evidence was classified as high, moderate, low, or very low to assess the reliability of the study conclusions.

This study utilized a network meta-analysis method based on the frequentist framework, with all statistical calculations performed using the network package in STATA 16.0 software. Network plots were first constructed to visually present direct comparisons and evidence distribution among interventions. Although outcome measures have uniform units such as steps and meters, standardized mean difference or SMD and its 95% confidence interval were uniformly adopted as effect size indicators to achieve standardized data merging. This decision accounted for the use of measurement devices with different sensitivities such as various brands of pedometers and accelerometers and differences in testing protocols across studies. A random-effects model was prioritized to account for potential clinical heterogeneity in population characteristics and intervention implementation across studies. To assess consistency, a global Wald test based on the design-by-treatment interaction model was used to evaluate overall network consistency, while the node-splitting method was employed to assess local inconsistency between direct and indirect evidence in specific comparisons. Finally, the Surface Under the Cumulative Ranking curve or SUCRA was calculated to probabilistically rank the efficacy of each intervention. A SUCRA value closer to 100% indicates a higher probability of the intervention being the optimal treatment. Furthermore, comparison-adjusted funnel plots were used to identify potential publication bias or small-study effects. For multi-arm randomized controlled trials, intervention arms were strictly screened according to the predefined node definitions in Table 1. If a multi-arm trial included arms that did not meet the definition of digital interventions in this study, such as pure unsupervised training without digital support, those arms were excluded from the network analysis. In all multi-arm trials included in this study, only one eligible intervention arm remained for comparison with the control group after screening. Consequently, these studies were treated as two-arm trials in the network, involving no shared control group issues and thus requiring no additional variance adjustments or sample size splitting.

The initial electronic database search yielded a total of 2,769 records. Duplicates were removed using EndNote and supplemented by manual verification to ensure accuracy, which resulted in the exclusion of 248 records. Subsequently, the titles and abstracts of the remaining 2,521 records were screened, leading to the exclusion of 2,457 irrelevant records. We then retrieved 64 full-text articles for detailed eligibility assessment. Following a rigorous review, 43 studies were excluded for the following reasons, including non-randomized controlled trial design for 9 studies, ineligible study population such as non-COPD patients for 14 studies, inappropriate intervention such as pure remote monitoring lacking a physical activity promotion component for 12 studies, and failure to report outcomes of interest for 8 studies. Ultimately, 21 studies met all inclusion criteria and were included in this systematic review and network meta-analysis. The detailed literature screening process is illustrated in Figure 1.

A total of 22 studies were included in this systematic review (Table 2) (21–41). The sample sizes across these studies varied significantly, ranging from small pilot trials with only 22 participants to large multicenter trials involving 375 subjects (28, 42). Despite the variation in sample sizes, the age characteristics of the participants were relatively consistent, with the mean age in most studies concentrated between 60 and 75 years (35, 41). While all intervention groups received digital health interventions, the control group settings were diverse, including usual care (34), active control groups receiving educational brochures or pedometers (32, 39), and traditional face-to-face pulmonary rehabilitation (37).

In terms of technological modalities, the studies employed various technical solutions, including basic telephones and pedometers (37), more sophisticated smartphone applications (25), customized web platforms (26), Bluetooth peripherals for remote monitoring (27), and videoconferencing systems or virtual reality devices for real-time supervision (35). Exercise types were primarily divided into two categories: one focusing on daily unsupervised physical activity and the other consisting of comprehensive exercise simulation programs including aerobic, resistance, and balance training to simulate professional pulmonary rehabilitation. Exercise prescription designs varied significantly, with intervention durations ranging from short-term intensive 2-week inpatient interventions to long-term home follow-ups lasting up to 24 months (35, 38). Exercise intensity was typically personalized based on individual baseline levels. Progression strategies primarily followed two patterns, namely automatic progression through automated algorithms or manual adjustments by physical therapists or health coaches based on patient feedback and remote data. The interactivity of supervision and feedback also varied, covering high-intensity real-time video supervision, regular telephone coaching, asynchronous messaging, and automated visualizations or motivational prompts generated by applications. By evaluating the baseline characteristics in Table 2, we confirmed the reasonableness of the network transitivity assumption. The results showed that key effect modifiers were balanced across the direct comparison groups. The mean age of participants was between 60 and 75 years, showing high homogeneity, and the intervention period for the vast majority of studies, specifically over 70 percent, was concentrated between 8 and 12 weeks. This consistency in demographic characteristics and experimental design supports the validity of merging data within the network and minimizes the risk of systemic bias caused by effect modifiers.

The methodological quality of the 22 included randomized controlled trials was assessed using the Cochrane RoB 2.0 tool as shown in Figures 2, 3. Overall, the risk of bias varied across the studies, with 14 studies rated as low risk, 4 studies showing some concerns, and 5 studies judged as high risk. Specifically, the randomization process in Domain 1 was generally robust, as most studies provided adequate descriptions of random sequence generation and allocation concealment mechanisms. The risk of bias was primarily concentrated in deviations from intended interventions in Domain 2 and missing outcome data in Domain 3. Six studies were rated as high risk in Domain 2, largely due to the inherent challenges of behavioral exercise interventions. Such interventions often lead to an open label design where blinding of participants and personnel is unfeasible, and some studies lacked appropriate intention to treat analysis to adjust for this bias. The risk in Domain 3 mainly stemmed from high dropout rates in certain studies and a lack of robust data imputation methods. Regarding the measurement of the outcome in Domain 4, the overall risk was low because this study exclusively utilized objective indicators, namely daily steps and 6-min walk distance, which are less susceptible to subjectivity. However, some studies were still rated as having some concerns, primarily due to the failure to ensure blinding of outcome assessors. Despite the objective nature of these measures, a conservative approach was required to exclude potential detection bias.

Sensitivity analysis was conducted for the primary outcome indicators using the leave-one-out method (Figure 6). The results for daily step counts demonstrated high robustness. Upon the sequential exclusion of each study, the recalculated pooled Standardized Mean Difference (SMD) fluctuated between 0.22 and 0.35, consistently maintaining statistical significance. The overall trend of improvement persisted even after excluding Study 5 (Wearable Only vs. Usual Care) (34), which carried the greatest weight in the pooled effect. This indicates that the efficacy of digital health interventions in promoting daily step counts remains highly consistent across different combinations of studies. Conversely, the robustness of the results for 6-Minute Walk Distance (6MWD) was relatively weak, exhibiting high sensitivity to individual large-sample studies. Specifically, after the exclusion of Study 4 (App + Sensor vs. Usual Care) (28, 42), the pooled SMD decreased significantly from the original 0.14 to 0.08. Consequently, the 95% Confidence Interval (CI) crossed the line of no effect (−0.03, 0.19), resulting in a loss of statistical significance. This finding suggests that the currently observed improvement in exercise capacity (6MWD) via digital interventions is heavily driven by specific large-scale trials; therefore, caution is warranted when interpreting these findings and promoting their clinical application.

(a) Sensitivity analysis for Daily Steps. The vertical red dashed line represents the original pooled Standardized Mean Difference (SMD = 0.30). The results demonstrate robustness, as the pooled effect size remained statistically significant even after excluding the outlier study (Study 5). (b) Sensitivity analysis for 6-Minute Walk Distance (6MWD). The vertical red dashed line represents the original pooled SMD (0.14). The results indicate lower robustness, as excluding a large-scale study (Study 4) led to a substantial decrease in the effect size, with the 95% CI crossing the line of null effect. Note: Each point and error bar represents the recalculated pooled SMD and 95% CI after omitting the specific study labeled on the y-axis.

Funnel plots and Egger’s regression tests were utilized to evaluate potential publication bias or small-study effects for the primary outcomes (Figure 7). For daily step counts, the scatter points within the funnel plot were symmetrically distributed around the mean difference, and Egger’s test yielded a p-value of 0.543. Similarly, for the 6-Minute Walk Distance (6MWD), the distribution of studies in the funnel plot appeared relatively uniform, with an Egger’s test p-value of 0.609. Statistical results indicated that the p-values for both primary outcomes were greater than 0.05, suggesting that no statistically significant publication bias was detected within the evidence network of this study.

The certainty of evidence for the primary outcomes was systematically evaluated using the GRADE framework (Table 5). The certainty of evidence for daily step counts was graded as “Low,” primarily attributed to substantial statistical inconsistency (global test p = 0.0001) and local inconsistency confirmed by the node-splitting method. Furthermore, core nodes such as Synchronous Tele-rehab exhibited serious imprecision due to wide confidence intervals resulting from limited participant numbers. Additionally, the prevalence of open-label designs in the included studies increased the risk of performance bias. Similarly, the certainty of evidence for 6MWD was graded as “Low.” Although this outcome demonstrated good network consistency (p = 0.66), the rating was downgraded due to the risk of bias inherent in the inability to blind exercise rehabilitation interventions, the borderline statistical significance of the pooled effect size, and the weak robustness of conclusions revealed by sensitivity analysis, which indicated high sensitivity to specific large-sample studies.

This study utilized a network meta-analysis to evaluate the relative efficacy of various digital health intervention modalities on daily physical activity and exercise capacity in patients with chronic obstructive pulmonary disease. Given the observed statistical inconsistency and the overall low certainty of evidence, these findings should be interpreted with caution in clinical practice. Regarding the improvement in daily steps, Synchronous Tele-rehabilitation ranked highest in the probability ranking, which may be attributed to its inherent real-time guidance and supervision functions. App + Sensor ranked second and demonstrated a statistically significant difference in pairwise comparison with Usual Care, highlighting its effectiveness in promoting physical activity.

Concerning the improvement in exercise capacity measured by the 6-min walk distance, the Wearables Only group achieved the highest probability score and showed marginal statistical significance compared with the Usual Care group, suggesting that wearable-based monitoring has a positive impact on enhancing exercise tolerance. The App + Sensor group also demonstrated an improvement trend for this outcome. Notably, within the evidence network of this study, Center-based PR did not show significant superiority over digital interventions, indicating that digital health interventions may serve as viable alternatives when traditional rehabilitation resources are limited. Overall, the performance of different intervention modalities varied across outcomes. Due to the clinical heterogeneity in the design of included trials and the risk of bias in some studies, the current findings should be regarded as preliminary observations. In clinical application, it is recommended to select the most appropriate intervention modality based on the specific needs and circumstances of the patient. Although this study employed standardized mean differences to eliminate methodological heterogeneity caused by different measurement devices such as accelerometers with varying sensitivities and testing protocols, we back-transformed key effect sizes into absolute units to enhance clinical interpretability. Based on the pooled standard deviations of the control groups in the included studies, we estimated that for daily steps, the effect size of Composite Digital Interventions or App + Sensor relative to usual care (SMD = 0.20) corresponds to an increase of approximately 450 steps per day. For exercise capacity, the effect size of Wearables Only relative to usual care (SMD = 0.33) corresponds to an improvement of approximately 23.1 meters in the 6-min walk distance. These transformed values suggest that although some statistical differences are only at marginal levels, the magnitude of improvement approaches or reaches the reference range for the Minimum Clinically Important Difference in chronic obstructive pulmonary disease rehabilitation, representing substantial potential clinical value.

This network meta-analysis confirmed a positive trend for digital health interventions in improving objective exercise outcomes in patients with chronic obstructive pulmonary disease, aligning with the general conclusions of existing systematic reviews that tele-rehabilitation can produce benefits comparable to traditional modalities (43, 44). However, the uniqueness of this study lies in its stratified analysis of technological architectures, revealing significant differences in the efficacy of various intervention modalities across different outcomes. These findings contradict some previous reviews (45, 46). Past studies tended to aggregate all digital technologies for analysis, often leading to a general conclusion that composite App models were superior. In contrast, this network meta-analysis found that for the 6-min walk distance, the simple Wearables Only intervention was actually superior to the more comprehensive App + Sensor modality. This discrepancy may be attributed to differences in technological burden and the specificity of intervention content. For instance, Gloeckl et al. [2025] noted in a multicenter randomized controlled trial exploring smartphone pulmonary rehabilitation that while fully automated interactive functions can provide multidimensional support, complex interface operations may distract patients from the exercise task itself to some extent (47). Conversely, feedback provided by Wearables Only is more intuitive and focused on daily activity itself (48), thus better targeting the 6-min walk distance test, which is purely an assessment of walking endurance. This difference further supports the conceptual distinction between outcome measures: complex tele-rehabilitation systems primarily drive behavioral changes or daily steps through enhanced supervision, while simple wearable devices optimize walking strategies and exercise efficiency or 6-min walk distance through direct biofeedback.

Regarding the improvement of daily steps, Synchronous Tele-rehabilitation and App + Sensor protocols ranked among the top. This phenomenon reflects the core role of supervision and feedback in driving sustained behavioral change (49). For example, the randomized controlled trial by Stenlund et al. [2024] showed that web-based self-management support effectively increases physical activity levels in patients with chronic obstructive pulmonary disease (50). Farmer et al. [2017] noted that feedback obtained through digital systems helps patients manage their disease in a home environment (51). Spielmanns et al. [2023] further confirmed that monitoring via mobile applications after the completion of pulmonary rehabilitation helps patients maintain achieved activity levels (52). Notably, the significant inconsistency detected in the daily steps analysis (p = 0.0001) reflects real heterogeneity in clinical practice. This inconsistency primarily stems from the lack of a unified definition and implementation standard for tele-rehabilitation; for example, Tsai et al. [2017] (28) employed high-frequency real-time video supervision, while other studies may only involve low-frequency telephone follow-ups. Differences in intervention intensity directly lead to a dispersion of effect sizes. Furthermore, sensitivity analysis showed that the improvement trend remains robust even after excluding studies with extremely large effect sizes. This indicates that the behavior-promoting effect of digital interventions does not depend on the extreme performance of a single study but possesses strong robustness.

In the dimension of enhancing exercise capacity or 6-min walk distance, Wearables Only ranked first in the probability ranking at 90.3%. The study by Mendoza et al. [2015] (42) validated the specificity between this intervention and the outcome measure, indicating that simple pedometer feedback can directly guide patients to adjust walking efficiency. Wootton et al. [2019] also observed the positive impact of continuous feedback on enhancing walking endurance (32). However, the robustness of the 6-min walk distance conclusion is relatively weak. Sensitivity analysis showed that the comparison of App + Sensor versus Usual Care has a decisive influence. After excluding this comparison, the conclusion loses statistical significance because studies such as Demeyer [2017] or Vorrink [2016] are large-sample studies with extremely high weights (25, 27). Such large-scale randomized controlled trials significantly narrow the confidence intervals of network effects due to sufficient sample sizes; once this comparison is excluded, the remaining small pilot studies struggle to maintain statistical power due to high variability. This suggests that the current effectiveness of digital interventions in enhancing physical capacity is highly dependent on the support of specific high-quality, large-scale trials, and clinical interpretation should remain cautious.

This study has several limitations concerning statistical robustness and evidence quality. First, significant global inconsistency was detected in the network meta-analysis of daily steps (p = 0.0001), along with local evidence conflicts. This inconsistency limits the interpretability of pooled effect sizes and likely stems from differences in intervention protocols and heterogeneity in baseline patient characteristics across studies. Furthermore, the sample sizes for specific intervention nodes, such as Synchronous Tele-rehabilitation and Center-based PR, were small, with n = 49 and n = 29 respectively, leading to wide confidence intervals for effect size estimates and statistical imprecision. Consequently, the certainty of evidence for both daily steps and 6-min walk distance was rated as low.

Second, there are limitations regarding study design and generalizability. The included randomized controlled trials primarily utilized open-label designs, and the inherent difficulty of blinding in exercise interventions increases the risk of performance bias. The analysis of 6MWD exhibited high sensitivity to individual large-sample studies; the loss of statistical significance after excluding such studies indicates that the robustness of these conclusions needs strengthening. Moreover, since the study excluded patients unable to use smart devices and most follow-up periods were concentrated between 12 weeks and 6 months, the usability and adaptability of digital intervention tools represent key bottlenecks for long-term adherence among older populations. Complex user interfaces may exceed the cognitive or operational capacities of older adult patients, leading to digital exclusion. As highlighted in recent research, the lack of adaptive designs tailored to the physiological and psychological characteristics of older adults is a primary barrier to the widespread application of digital health technologies (53).

Furthermore, the external validity of this study may be limited in resource-scarce regions, particularly in low- and middle-income countries. The included studies primarily originated from high-income countries where participants generally possess better digital literacy. In low- and middle-income country settings, promoting complex synchronous tele-rehabilitation or composite App interventions faces dual challenges involving not only network infrastructure coverage but also a lack of sufficient medical human resources to monitor and provide feedback on massive amounts of patient data in real time. In contrast, this study found that Wearables Only has significant potential in improving exercise capacity, which suggests that this low-cost, low-tech, and lightweight intervention model, which does not rely on continuous professional supervision, may be a more equitable and scalable rehabilitation alternative for resource-constrained environments. Finally, although we established decision rules, conceptual overlap in hybrid interventions may still introduce misclassification bias and serve as a potential source of network heterogeneity. To address this challenge, we adopted the Highest Interaction Level Priority principle to achieve classification standardization. Additionally, sensitivity analysis confirmed that the primary conclusions remained robust after excluding ambiguous studies, suggesting that the impact of this bias on overall estimates is limited.

According to the current analysis, different digital intervention modalities may exhibit varying improvement trends in chronic obstructive pulmonary disease rehabilitation, suggesting that clinical selection should be customized to specific rehabilitation goals. Preliminary data show that Synchronous Tele-rehabilitation ranks high in the probability of increasing daily steps, which may imply that real-time guidance plays a positive role in promoting behavioral changes. For patients prioritizing the enhancement of functional exercise tolerance or 6MWD, wearable devices based solely on probability rankings show potential utility, possibly due to the direct regulation of walking efficiency through real-time biofeedback. However, given that 6MWD findings are highly sensitive to specific studies and have low certainty of evidence, these findings should be regarded as exploratory references rather than definitive conclusions.

Furthermore, the improvement trends observed for App + Sensor across both indicators suggest its potential value as a comprehensive intervention program. Rather than viewing digital health interventions as a competitive substitute for center-based pulmonary rehabilitation, they should be considered a Capacity Extender. For patients unable to participate in offline rehabilitation due to geographic and transportation constraints, time conflicts, or shortages of medical resources, digital modalities provide a vital bridge for accessibility. Future clinical decision-making should favor a stratified care model where center-based rehabilitation is prioritized for patients with complex conditions or those undergoing initial rehabilitation, while Wearables Only or tele-rehabilitation can serve as effective supplementary maintenance tools for patients in the maintenance phase or those in resource-limited areas.