This systematic review and meta-analysis was conducted in strict accordance with the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) guidelines and the Cochrane Handbook for Systematic Reviews of Interventions. This study’s protocol was prospectively registered in the PROSPERO (International Prospective Register of Systematic Reviews: CRD420261342287).

Several modifications to the registered protocol (PROSPERO CRD420261342287) were made during the conduct of the review and are declared here for transparency. With respect to outcomes, quality of life and BMI were originally listed as primary outcomes but were removed from the meta-analysis—quality of life because the included trials used heterogeneous instruments (36-Item Short Form Health Survey; SF-36, MacNew, EQ-5D, disease-specific questionnaires) at incompatible time points that precluded meaningful pooling, and BMI because only one included trial reported an extractable end-of-intervention measurement. SBP and DBP, originally grouped as a single outcome, were redesignated as 2 separate prespecified secondary outcomes, with VO₂ peak elevated to the sole primary outcome. With respect to information sources, Web of Science was added to the 4 databases originally specified, expanding the search to 5 databases. The supplementary search was confined to backward citation searching of relevant reviews; we did not contact original trial authors, search conference proceedings, dissertation databases, or trial registers as originally planned, and the final synthesis was restricted to English-language publications despite the protocol imposing no language restriction. Risk of bias was assessed using Cochrane RoB 2 (Cochrane Risk of Bias Tool version 2) only, rather than both RoB-1 and RoB-2 as originally stated. The originally planned I²-thresholded model selection has been replaced by a uniform random-effects model with the Hartung-Knapp-Sidik-Jonkman adjustment. Five digital-health-specific moderators (telemedicine modality, guidance type, technology platform, intervention duration, and intervention composition) were prespecified during the analysis phase, operationalizing the broader category of “technology type and intervention duration” mentioned in the protocol. A full GRADE (Grading of Recommendations Assessment, Development, and Evaluation) assessment was added, although the protocol stated certainty would not be assessed, and reporting now follows PRISMA 2020 (expanded) and PRISMA-S (Preferred Reporting Items for Systematic Reviews and Meta-Analyses literature search extension). Safety and feasibility, originally planned to be assessed, are addressed narratively in the Discussion rather than as formal pooled analyses, owing to inconsistent reporting in the primary trials.

This systematic review and meta-analysis analyzed previously published aggregate-level data and did not involve primary data collection, recruitment of participants, or any form of intervention. Ethical approval and informed consent were obtained by the original investigators of each included trial as documented in their respective publications. As a secondary analysis of deidentified summary statistics extracted from the published literature, this present review did not require separate ethics committee approval or informed consent.

A comprehensive literature search was performed in PubMed, the Cochrane Library, Web of Science, Embase, and MEDLINE using a combination of controlled vocabulary terms and free-text terms. The main search terms included “myocardial ischemia,” “coronary artery disease,” and “telemedicine.” No language restrictions were imposed, and all databases were searched from inception to March 20, 2026. To ensure comprehensive study identification, the reference lists of relevant reviews, systematic reviews, and meta-analyses were also manually searched. All retrieved records were imported into EndNote (Clarivate; version 25) for deduplication and record management.

RCTs evaluating the effects of exercise-based telerehabilitation in adults with CVD or related cardiovascular conditions were eligible for inclusion. Studies were included if (1) participants had CAD, ischemic heart disease, myocardial infarction, post–percutaneous coronary intervention status, heart failure, low-risk coronary heart disease, essential hypertension, or had completed CR; (2) the intervention primarily involved exercise-based telerehabilitation, including web-based or mHealth programs, delivered alone or in combination with text messaging, telephone calls, video calls, emails, smartwatches, or related modalities; (3) the control group received usual care without structured exercise training; (4) VO₂ peak was reported as the primary outcome, with SBP and DBP as secondary outcomes; and (5) the study design was an RCT.

Studies were excluded if (1) they involved healthy individuals or patients without CVD or related cardiovascular conditions; (2) the intervention mainly focused on health education, psychoeducation, or lifestyle counseling rather than exercise training; (3) the control group received structured exercise training or telerehabilitation; (4) outcomes related to VO₂ peak or blood pressure were not reported; or (5) the study was a cohort, case-control, cross-sectional, or nonrandomized controlled study, or was published as a conference abstract, review, or meta-analysis.

For the purposes of synthesis, studies were grouped according to telemedicine modality (synchronous vs asynchronous), guidance type (professional-led, patient self-managed, or technology-system-led), technology delivery platform (smartphone or mHealth app-based, wearable sensor-based, or SMS-, web-, or telephone-based), intervention duration (≤12 wk vs >12 wk), and intervention composition (exercise-only interventions that did not incorporate any additional components vs multicomponent interventions that combined exercise with supplementary elements such as health education, dietary counseling, or psychological support) to enable meaningful subgroup comparisons.

A comprehensive literature search was conducted across 5 electronic databases: PubMed, the Cochrane Library, Web of Science, Embase, and MEDLINE. All databases were searched from inception to March 20, 2026. No language restrictions were imposed. In addition, the reference lists of relevant reviews, systematic reviews, and meta-analyses were manually screened to identify any additional eligible studies. All retrieved records were imported into EndNote (version 25) for deduplication and record management.

The search strategy was developed and reported in accordance with the PRISMA-S extension for systematic review search reporting [30]. A systematic literature search was conducted across 5 electronic databases: PubMed, MEDLINE, Embase, Web of Science, and Cochrane Library. Each database was searched individually on its respective platform; no multidatabase simultaneous searching was performed. The search covered the period from database inception to March 20, 2026. No language, publication date, or study design restrictions were applied at the database search stage. No published search filters were used, and the search strategy was developed de novo without adaptation from previous reviews. Full database-specific search strategies, including all Boolean operators, controlled vocabulary terms (MeSH and Emtree), free-text synonyms, and field modifiers, are provided in Section S1 in Multimedia Appendix 1 [31-43].

To supplement the database search, 2 trial registries were searched: ClinicalTrials.gov and the WHO International Clinical Trials Registry Platform, using equivalent key concepts (coronary heart disease, telerehabilitation or telemedicine, and exercise). In addition, both backward and forward citation searching were performed. Backward citation searching was conducted by manually screening the reference lists of all included studies and relevant systematic reviews. Forward citation searching was performed using Web of Science to identify papers that had cited the included studies.

We did not search gray literature databases, dissertation repositories, or conference proceedings, and we did not contact study authors, experts, or manufacturers to obtain additional or unpublished data. No formal peer review of the search strategy was conducted, and no external information specialists were consulted during search strategy development. No search updates or email alerts were established after the initial search was completed. These limitations of the search methodology are acknowledged in the Discussion section.

All retrieved records were imported into EndNote (version 25) for deduplication, which was performed using the software’s built-in duplicate detection function followed by manual verification.

To quantify the degree of citation overlap between the present review and prior systematic reviews of telerehabilitation-based exercise interventions in CVD, we calculated the corrected covered area (CCA) according to Pieper et al [44], using the formula CCA=(N−r)/(r×c−r), where N is the total number of included primary publications across all reviews, r is the number of unique primary publications, and c is the number of reviews. CCA values were interpreted using the established thresholds of 0%‐5% (slight), 6%‐10% (moderate), 11%‐15% (high), and >15% (very high overlap) [44]. The complete citation matrix is provided in Section S2 in Multimedia Appendix 1.

Study selection was performed independently by 2 reviewers in 2 stages. In the first stage, all retrieved records were imported into EndNote (version 25) for deduplication, after which titles and abstracts were screened against the prespecified eligibility criteria; reasons for exclusion were recorded in detail. In the second stage, the full texts of potentially eligible papers were retrieved and independently assessed for final inclusion. Disagreements at either stage were resolved through discussion with a third reviewer until consensus was reached.

Using a prespecified standardized data extraction form, 2 reviewers independently extracted study characteristics and outcome data from each included report. Extracted data were cross-checked to ensure accuracy and consistency. Disagreements were resolved through discussion with a third reviewer until consensus was reached. No automation tools were used, and no additional data were sought from study investigators.

Adverse event reporting and intervention adherence were extracted as secondary safety and feasibility outcomes. For adverse events, we recorded whether each trial included a dedicated adverse event section, the nature and severity of any reported events, and whether events were adjudicated as intervention-related. For adherence, we recorded the metric used, the numerical value reported, and the method of measurement (eg, device-based monitoring, patient logbook, and platform log-in records). Given the anticipated heterogeneity in how these outcomes were operationalized across trials, no attempt was made to pool them quantitatively; they are presented descriptively alongside study characteristics.

The primary outcome was VO₂ peak. Secondary outcomes included SBP and DBP. For each outcome, the means and SDs of both the intervention and control groups at the end of the intervention period were extracted. Where multiple time points were reported, the end-of-intervention measurement was prioritized.

The following study-level and participant-level variables were also extracted: first author, publication year, sample size, mean participant age, intervention type, intervention duration, and proportion of male participants. Where data were missing or unclear, the original report was reexamined. Studies for which the required information remained unavailable after reexamination were excluded from the analysis.

The risk of bias of the included studies was independently assessed by 2 reviewers using the revised Cochrane Risk of Bias tool for randomized trials (RoB 2). Disagreements were resolved through consultation with a third reviewer. The following 5 domains were evaluated: bias arising from the randomization process, bias due to deviations from intended interventions, bias due to missing outcome data, bias in measurement of the outcome, and bias in selection of the reported result. Each domain was judged as having a low risk of bias, some concerns, or a high risk of bias, and an overall risk-of-bias judgment was derived for each study accordingly.

For all continuous outcomes (VO₂ peak, SBP, and DBP), mean differences (MDs) with 95% CIs were used as the effect measure.

Eligibility for each synthesis was determined by tabulating the intervention characteristics and outcome domains of each included study and comparing them against the prespecified groupings defined in the eligibility criteria. Studies reporting the same outcome with comparable intervention and control conditions were pooled in the corresponding meta-analysis.

Where SDs were not directly reported, they were calculated in accordance with the Cochrane Handbook for Systematic Reviews of Interventions using the following approaches: when a 95% CI for the difference in means was available, the SD for each group was calculated by dividing the length of the CI by 3.92 and then multiplying by the square root of the sample size; when a t value was available, the SD of the difference in means was calculated by dividing the MD by the t value and then multiplying by the square root of the sample size; and when actual P values obtained from t tests were available, the corresponding t value was first obtained from a table of the t distribution, after which the same formula was applied. No other data conversions were required.

The results of individual studies and pooled syntheses were presented using forest plots. Study-level characteristics were tabulated to provide an overview of the included evidence.

Statistical analyses were performed in Stata (StataCorp; version 18.0) and R (R Foundation; version 4.5.1) with the metafor package (version 4.8.0). Given the anticipated clinical and methodological heterogeneity across included studies—arising from differences in populations, intervention protocols, delivery platforms, and follow-up durations—all meta-analyses were conducted using a random-effects model [45]. Between-study variance (τ²) was estimated using the Sidik-Jonkman method, and CIs for the pooled effect estimates were computed with the Knapp-Hartung adjustment (also known as the Sidik-Jonkman adjustment) to provide more adequate coverage when the number of studies is small [46]. With this adjustment, inferences for the overall effect were based on the Student t distribution. This combination of methods is hereafter referred to as the Hartung-Knapp-Sidik-Jonkman approach [46]. Heterogeneity was quantified using the Cochran Q test and the I² statistic, with significant heterogeneity defined as P<.10 or I²>50%. For meta-analyses including 10 or more studies, 95% prediction intervals were calculated to distinguish the average treatment effect from the range of true effects expected across different clinical settings [47]. All forest plots and the funnel plot were generated in R using the metafor package, with the Knapp-Hartung label explicitly displayed on each forest plot to ensure consistent and transparent presentation across figures. All included studies used nonoverlapping independent subject samples; where a trial included multiple study groups, data from only 1 group were extracted.

For results with significant heterogeneity and sufficient research support (N≥10), we performed meta-regression analysis to explore potential sources of heterogeneity. Prespecified covariates included telemedicine modality (synchronous vs asynchronous), guidance type (professional-led, patient self-managed, or technology-system-led), technology delivery platform (smartphone or mHealth-based, wearable sensor-based, or SMS-, web-, or telephone-based), intervention duration (≤12 wk vs >12 wk), and intervention composition (exercise-only vs multicomponent). Subgroup analyses were conducted along these same dimensions to compare effects across categories. For the guidance type subgroup analysis, professional-led interventions were defined as those in which a qualified health care professional—such as a nurse, coach, or physician—actively directed the intervention through regular structured contact (eg, telephone calls, messaging platforms, or videoconferencing); patient self-managed interventions were defined as those in which participants independently executed the exercise program with asynchronous or automated feedback and without regular proactive contact from a health care provider; and technology-system-led interventions were defined as those in which an intelligent system or algorithm, rather than a human professional, served as the primary agent of real-time guidance, monitoring, or corrective feedback (eg, AI-based motion recognition or automated anomaly detection systems). For the intervention composition subgroup analysis, exercise-only interventions were defined as those in which exercise training was the sole content of the program without any additional components, whereas multicomponent interventions were defined as those that delivered exercise training as the core component alongside supplementary elements such as health education, dietary counseling, or psychological support.

Leave-one-out sensitivity analyses were performed to evaluate the influence of individual studies on the pooled estimates and to assess the robustness of the results.

When at least 10 studies were available for a given outcome, potential small-study effects were assessed statistically using the Egger regression test.

The certainty of the body of evidence for each outcome was assessed using the GRADE approach. Evidence was rated as high, moderate, low, or very low based on 5 domains: risk of bias, inconsistency, indirectness, imprecision, and publication bias.

A total of 1301 records were retrieved from the databases (PubMed n=210, MEDLINE n=510, Embase n=205, Web of Science n=199, and Cochrane Library n=177). In addition, 136 records were identified through other methods, including 132 from trial registries (ClinicalTrials.gov n=129 and WHO International Clinical Trials Registry Platform n=3) and 4 from citation searching (backward n=3 and forward n=1).

For the database searches, 436 duplicate records were removed using EndNote (version 25), leaving 865 records for screening. Titles and abstracts were screened against the eligibility criteria, and 808 records were excluded. Subsequently, 57 reports were sought for full-text retrieval, of which 1 could not be retrieved, leaving 56 reports for full-text eligibility assessment. Of these, 25 were excluded for not meeting the inclusion criteria, 20 were excluded because the interventions evaluated were not exercise-based remote rehabilitation, and 1 was excluded because it used a nonrandomized controlled design, resulting in 10 studies included from database searches.

For the 136 records identified through other methods, all underwent screening. Of these, 42 records were duplicates with database search results, 34 records were ongoing trials without published results, and 56 records did not meet the inclusion criteria, leaving 4 reports for full-text eligibility assessment. After full-text assessment, 1 was excluded for not matching the inclusion criteria, resulting in 3 studies included from this pathway.

Finally, a total of 13 studies [31-43] were included in the meta-analysis, comprising 10 studies from database searches and 3 studies from other methods. The study screening process is shown in Figure 1.

Among the 47 papers excluded during the full-text evaluation stage, the main reasons for exclusion were (1) not meeting the inclusion criteria (n=26, including 25 from database searches and 1 from other methods), for example, the outcome indicators of the study by Leemrijse et al [48] did not meet the criteria; (2) the study design was a non-RCT (n=1); (3) although the interventions included remote methods, the core content was not exercise-based remote rehabilitation (n=20).

To quantify the degree of citation overlap between the present review and previous systematic reviews on telerehabilitation-based exercise interventions in CVD, we calculated the CCA following Pieper et al [44], excluding the umbrella review by Shi et al [24], which does not independently search for primary trials. The CCA was 6.6% (N=19, r=13, c=8), falling within the 6%‐10% range conventionally interpreted as moderate overlap [10]. This indicates that, despite addressing a related clinical question, the present review draws on a largely distinct primary-trial evidence base relative to prior reviews, consistent with our exercise-isolated eligibility criteria and broad cardiovascular-disease enrollment. The full citation matrix is provided in Section S2 in Multimedia Appendix 1.

Thirteen studies published in English were included, comprising a total of 958 participants. The studies were published between 2007 and 2026 and were conducted in China [35,39-41,43], Iran [32], Belgium [31,34], the United States [33], New Zealand [36], Israel [37], Brazil [38], and Canada [42]. All studies reported participant age and sex distribution. Mean age ranged from 11.2 to 69.0 years, the proportion of male participants ranged from 47.5% (19/40) to 90.77% (59/65), and sample sizes ranged from 7 to 86.

In the studies reporting VO₂ peak, in terms of telemedicine modality, 5 studies [31,33,34,36,38] used asynchronous delivery, while the remaining 5 studies [37,39-41,43] used synchronous communication. Regarding guidance type, 5 studies [31,33,38,40,43] were classified as professional-led, involving active and structured contact with nurses, coaches, or physicians via telephone, messaging platforms, or video calls; 3 studies [34,36,37] were classified as patient self-managed, in which participants independently executed the exercise program with asynchronous or automated feedback and without regular proactive contact from a health care provider; and 2 studies [39,41] were classified as technology-system-led, in which an intelligent system served as the primary agent of real-time guidance through AI-based motion recognition or automated anomaly detection. With respect to technology delivery platform, 5 studies [33,39-41,43] used smartphone or mHealth app-based platforms, 2 studies [31,34] used wearable sensor-based systems, and 3 studies [36-38] used SMS-, web-, or telephone-based platforms. Concerning intervention duration, 6 studies [31,33,38,40,41,43] had durations of ≤12 weeks, whereas 4 studies [34,36,37,39] exceeded 12 weeks. Regarding intervention composition, 8 studies [31,33,34,36,37,40,41,43] were multicomponent interventions that combined exercise training with supplementary elements such as health education, dietary counseling, or psychological support, while 2 studies [38,39] consisted of exercise-only interventions without any additional components. Detailed study characteristics are presented in Table 1.

Adverse event reporting and intervention adherence varied considerably across included trials. Regarding safety, 10 of the 13 trials included an explicit adverse event statement, of which 9 trials declared that no serious exercise-related adverse events occurred during the intervention period [31,32,37-43]. One trial [36] reported 31 serious adverse events across 22 participants, of which only 1 participant, a hospitalization following a cycling accident, was adjudicated as intervention-related. The remaining 3 trials [33-35] did not include a dedicated adverse event section [34]; instead reported rehospitalization rates as a clinical end point rather than as formally adjudicated safety data. A per-study summary of adverse event reporting is provided in Section S3 in Multimedia Appendix 1.

Adherence to the prescribed intervention was quantified using heterogeneous metrics across trials. Exercise sessions per week were reported by 4 trials [31,37-39], daily step counts by 2 trials [33,37], platform login frequency by 1 trial [42], and message or video engagement rates by 1 trial [36]. One trial reported the percentage of prescribed sessions completed [43], and 1 trial applied an attendance threshold as an eligibility criterion rather than reporting adherence as an outcome [32]. One trial described adherence as “good” without supplying numerical data [35], and 3 trials provided no adherence data whatsoever [34,40,41].

The RoB 2 was used to assess the methodological quality of all 13 included studies. The assessment covered 5 dimensions: randomization process, deviation from expected intervention, missing outcome data, outcome measurement, and selectivity in reporting results, and a comprehensive assessment of the overall risk of bias was given. Detailed assessment results for each study are shown in Figure 2 [31-43], and the summarized percentage results are shown in Figure 3.

Regarding the randomization process, 38.5% (5/13) of the studies were assessed as having low risk of bias, 23.1% (3/13) had some concerns, and 38.5% (5/13) were assessed as having high risk of bias. The high risk of bias was mainly due to inadequate description of the random sequence generation methods or failure to report allocation concealment schemes, making it one of the most prominent dimensions in this assessment.

Regarding deviations from intended interventions, 53.8% (7/13) of the studies were assessed as having low risk of bias, 30.8% (4/13) as having some concern, and 15.4% (2/13) as having high risk of bias. Due to the nature of exercise-based telerehabilitation, participant blinding was not feasible in all studies, and some studies had shortcomings in handling intervention adherence and intention-to-treat analyses.

Regarding missing outcome data, 84.6% (11/13) of the studies were assessed as having low risk of bias, 15.4% (2/13) as having some concern, and no studies were assessed as having high risk of bias, suggesting that the overall handling of outcome data completeness was relatively appropriate across studies.

Regarding measurement of the outcome, 84.6% (11/13) of the studies were assessed as having low risk of bias, 15.4% (2/13) as having high risk of bias, and no studies had some concern. The widespread use of objective measurements (eg, VO₂ peak and blood pressure) helps reduce the risk of measurement bias.

Regarding selection of the reported result, 23.1% (3/13) of the studies were rated as low risk of bias, 69.2% (9/13) had some concerns, and 7.7% (1/13) were rated as high risk of bias. This dimension represents the area with the highest concentration of bias concerns, as most studies did not undergo preregistration or predefine the primary outcome, making it difficult to rule out the risk of selective reporting.

Regarding overall bias, based on the combined assessment of the 5 dimensions, 15.4% (2/13) of the studies were classified as having low overall risk of bias, 30.8% (4/13) had some concerns, and 53.8% (7/13) were classified as having high overall risk of bias. The overall bias level is primarily driven by 2 dimensions: inadequate randomization and selective reporting of results. These methodological limitations should be carefully considered when interpreting the results of this meta-analysis.

Summary statistics (means and SDs) and effect estimates (MDs with 95% CIs) for each included study are presented in the forest plots (Figures 4-6) [31-41,43]. Individual study data are displayed alongside the pooled effect estimates to facilitate study-level evaluation and comparison.

For the primary outcome VO₂ peak, a random-effects meta-regression model was used to explore potential sources of heterogeneity, incorporating 5 prespecified covariates: intervention duration, telemedicine mode, guidance type, technology delivery platform, and intervention composition. Detailed results are shown in Table 2.

The results showed that none of the 5 covariates were identified as significant sources of heterogeneity in VO₂ peak. Each moderator was analyzed in a separate univariate random-effects meta-regression with the Hartung-Knapp-Sidik-Jonkman adjustment; for moderators with 3 categories (type of guidance and technology delivery platform), 2 pairwise contrasts were estimated relative to the reference category, and an Omnibus test was used for the joint effect of the moderator. None of the moderators reached statistical significance at the conventional 0.05 threshold: intervention duration (Omnibus P=.30), telemedicine mode (Omnibus P=.56), type of guidance (Omnibus P=.36), technology delivery platform (Omnibus P=.41), and intervention composition (Omnibus P=.44). The 95% CIs for all comparisons crossed 0, and the Omnibus tests showed no statistically significant moderator effect, suggesting that the moderating effect size of each variable cannot be determined based on the current evidence. Detailed coefficients for each comparison are shown in Table 2.

Prespecified subgroup analyses were performed on the primary outcome VO₂ peak, and the results are summarized in Table 3. No statistically significant differences were found between groups in any of the subgroup analyses (all P values >.05), indicating that the existing evidence is insufficient to identify the aforementioned categorical variables as effect modifiers.

Regarding intervention duration, numerical differences were found between the effect estimates of the ≤12-week subgroup (n=371, MD 3.33, 95% CI 1.85‐4.81, P=.002, I²=37.13%) and the >12-week subgroup (n=378, MD 1.92, 95% CI −2.11 to 5.96, P=.23, I²=83.46%), with the latter’s CI crossing 0. The difference between groups was not statistically significant (P=.30). It is noteworthy that the >12-week subgroup exhibited extremely high heterogeneity (I²=83.46%). Given the limited number of studies within this subgroup and insufficient statistical power, this result should not be interpreted as a lack of effect with longer intervention durations.

Regarding telemedicine mode, the asynchronous mode subgroup (n=412, MD 3.01, 95% CI 0.88‐5.14, P=.02, I²=63.53%) showed a statistically significant effect, whereas the synchronous mode subgroup (n=337, MD 2.24, 95% CI −0.70 to 5.19, P=.10, I²=65.65%) did not reach statistical significance. Both subgroups exhibited moderate to high heterogeneity, and the effect estimates were of similar direction and magnitude. The between-subgroup difference was not statistically significant (P=.56), and the nonsignificant result in the synchronous subgroup should not be interpreted as evidence of absence of effect, given the relatively small number of contributing trials and the wide CI.

Regarding guidance type, the professional-led subgroup (n=316, MD 3.02, 95% CI 1.14‐4.91, P=.01, I²=31.85%) showed a statistically significant effect with low heterogeneity. The technology-system-led subgroup (n=151, MD 3.75, 95% CI −2.73 to 10.22, P=.09, I²=9.18%) yielded a numerically larger effect estimate but did not reach statistical significance, reflecting the very small number of contributing trials (k=2) and consequent imprecision rather than absence of effect. The patient self-managed subgroup (n=282, MD 1.48, 95% CI −5.87 to 8.83, P=.48) also showed a nonsignificant result with extremely wide CIs and very high within-subgroup heterogeneity (I²=81.90%), further limiting interpretability. The between-subgroup difference was not statistically significant (P=.36).

Regarding technology delivery platform, the smartphone or mHealth app subgroup (n=374, MD 3.51, 95% CI 2.21‐4.81, P=.002, I²=31.41%) showed accurate and statistically significant effect estimates with low heterogeneity within the subgroup. The CIs for the wearable sensor subgroup (n=183, MD 2.26, 95% CI −4.49 to 9.01, P=.29, I²=60.72%) and the SMS-, web-, or telephone-based subgroup (n=192, MD 1.75, 95% CI −28.91 to 32.40, P=.60, I²=73.39%) were extremely wide and crossed 0, reflecting high within-subgroup heterogeneity and insufficient statistical power due to the very small number of contributing trials (k=3 and k=2, respectively), rather than necessarily indicating ineffective intervention. The between-subgroup difference was not statistically significant (P=.41).

Regarding intervention composition, the multicomponent intervention subgroup (n=614, MD 2.34, 95% CI 0.59‐4.10, P=.02, I²=78.08%) showed a statistically significant effect, although with high heterogeneity, suggesting substantial variability in intervention effects within this subgroup. The exercise-only subgroup (n=135, MD 3.51, 95% CI −4.55 to 11.56, P=.11, I²=6.16%) yielded a numerically larger effect estimate but did not reach statistical significance, owing to the very small number of contributing trials (k=2) and the resulting wide CI, rather than to an absence of effect. The between-subgroup difference was not statistically significant (P=.44).

In summary, no statistically significant between-group effect modifications were detected in any subgroup analysis, and the existing evidence is insufficient to draw definitive conclusions regarding any single moderating variable.

According to the Cochrane guidelines, the Egger linear regression test was used to assess small sample effects for outcome measures with ≥10 included studies. The Egger test result for the VO₂ peak was statistically significant (P=.03). As shown in Figure 7, the precision-standardized effect size scatter plot exhibited some asymmetry, and the lower bound of the 95% CI of the regression line intercept was close to but did not completely exclude 0, suggesting a possible small sample effect. The analyses of SBP and DBP each included 5 studies, which did not meet the minimum sample size requirement (≥10 studies) for publication bias testing. Therefore, no statistical tests were performed on these outcomes. However, the potential impact of publication bias should still be considered when the number of included studies is limited.

Sensitivity analysis was performed on all outcome measures included in the meta-analysis, with the results shown in Figure 8 [31-41,43].

Regarding VO₂ peak (Figure 8A), after sequentially removing individual studies, the estimated pooled effect ranged from 2.32 to 3.19 mL/kg/min. The effect direction remained consistent across all removal scenarios, and the CIs did not cross 0, indicating that the pooled VO₂ peak results were not excessively influenced by any individual study and exhibited good robustness.

Regarding SBP (Figure 8B), after sequentially removing individual studies, the estimated pooled effect ranged from −3.55 to −0.75 mm Hg. The CIs crossed 0 across all removal scenarios, consistent with the overall pooled result, indicating that the pooled SBP results remained stable under sequential removal testing.

Regarding DBP (Figure 8C), after successively eliminating individual studies, the estimated pooled effect ranged from −2.52 to −1.23 mm Hg. The CIs for all elimination scenarios crossed 0, consistent with the overall pooled results, suggesting that the pooled DBP results are also relatively robust.

In summary, the sensitivity analyses of each of the 3 outcomes after elimination did not find any individual study that had a decisive impact on the pooled results, and the estimated pooled effect remained relatively consistent across all elimination scenarios.

The GRADE framework was used to assess the certainty of evidence for the 3 outcomes (Table 4). For VO₂ peak, the evidence was rated as very low certainty due to serious concerns regarding inconsistency, indirectness, and imprecision. For SBP, the evidence was also rated as very low certainty owing to serious inconsistency, indirectness, and imprecision. For DBP, the evidence was rated as very low certainty, primarily driven by very serious indirectness and serious imprecision. Given the very low certainty of evidence across all 3 outcomes, the pooled effect estimates should be interpreted with caution. Detailed GRADE assessment procedures and justifications for each downgrade are provided in Section S4 in Multimedia Appendix 1.

This systematic review and meta-analysis of 13 RCTs examined the effects of exercise-based telerehabilitation on cardiometabolic outcomes in patients with CVD. Compared with usual care, exercise-based telerehabilitation produced a statistically significant improvement in VO₂ peak, but did not produce statistically significant changes in either SBP or DBP. Crucially, although the average effect on VO₂ peak was favorable, the 95% prediction interval crossed 0, indicating that in some future clinical settings the true effect could plausibly be null or even slightly negative. The pooled estimate should therefore be interpreted as the average effect across the included trial contexts rather than a universal effect that will be reproduced in every implementation [49,50]. Heterogeneity for VO₂ peak was substantial, and GRADE certainty was rated low to very low for VO₂ peak and very low for both blood pressure outcomes, reflecting concerns related to inconsistency, indirectness, imprecision, and possible publication bias [51]. Within this evidentiary frame, our findings support a probable, but not certain, beneficial effect of exercise-based telerehabilitation on cardiorespiratory fitness (CRF), and provide no convincing evidence of an antihypertensive effect at the pooled level.

To explore the substantial heterogeneity observed for VO₂ peak, we conducted both meta-regression and prespecified subgroup analyses across 5 candidate moderators (intervention duration, telemedicine mode, type of guidance, technology delivery platform, and intervention composition). None of these covariates explained a statistically significant proportion of the between-study variance, and no statistically significant between-group differences were detected in any subgroup contrast. The residual heterogeneity is therefore likely attributable to factors that we were unable to model directly, including intertrial variation in baseline CRF, exercise prescription parameters (intensity, weekly frequency, session duration, and progression), supervision intensity within nominally similar delivery modes, comorbidity burden, concurrent pharmacotherapy, and outcome assessment protocols (cardiopulmonary exercise testing vs estimated VO₂ peak from submaximal testing) [52,53]. Patient-level moderators such as age, sex, baseline functional capacity, and engagement with the digital platform are also plausible contributors [54]. We were unable to perform additional moderator analyses on these characteristics because individual participant data were not available, and aggregate-level reporting of exercise dose and adherence was inconsistent and frequently incomplete across the included studies. This limitation is acknowledged below and represents an important target for future trials.

VO₂ peak is the gold-standard index of CRF and a robust independent predictor of cardiovascular and all-cause mortality, with the American Heart Association recommending its use as a clinical vital sign [55,56]. A pooled improvement in the order of 2‐3 mL/kg/min, even at a population-average level, is therefore clinically meaningful, since each 1 metabolic equivalent (≈3.5 mL/kg/min) increment in CRF is associated with an approximately 10%-25% reduction in mortality risk in CVD populations [55]. Our pooled estimate is broadly consistent with that of Zhang and Lin [57], who reported a long-term VO₂ peak benefit of cardiac telerehabilitation vs center-based CR in patients with CAD, and aligns with the direction of effect reported in the network meta-analysis by Li et al [19] and the heart-failure-specific synthesis of Gao et al [21]. Mechanistically, exercise-induced gains in VO₂ peak are supported by well-established physiological adaptations including expanded blood volume, increased hemoglobin mass and capillary density, mitochondrial biogenesis, and improved cardiac output and peripheral oxygen extraction [58-60]. The convergence of our findings with these prior reviews suggests that a remotely delivered exercise stimulus, when adequately prescribed, can engage the same physiological pathways that underpin the benefits of center-based CR [61].

Our null findings for both SBP and DBP stand in apparent contrast to the well-documented antihypertensive effect of structured exercise training, where moderate-intensity aerobic, dynamic resistance, and isometric exercise have all been shown to lower resting blood pressure by 3‐7 mm Hg in hypertensive populations through improvements in endothelial function, reductions in sympathetic outflow, enhanced parasympathetic tone, and reduced peripheral vascular resistance [62,63]. However, our null pooled effect is in close agreement with previous telerehabilitation-focused syntheses: Cruz-Cobo et al [20] reported nonsignificant effects on both SBP (P=.99) and DBP (P=.36) after mHealth-delivered secondary prevention; Yu et al [27] reported a nonsignificant effect on SBP only; and Zhong et al [15] likewise reported no significant impact on cardiovascular risk factors including blood pressure. Several explanations are plausible. First, the included trials were not designed primarily to lower blood pressure: most enrolled normotensive or pharmacologically controlled participants in whom further reductions are mechanistically constrained [64,65], and only 1 trial explicitly enrolled patients with essential hypertension [41]. Second, the exercise dose actually delivered through remote channels is often below the prescribed dose, and adherence in unsupervised settings tends to decline over time, weakening the hemodynamic stimulus [66,67]. Third, the small number of contributing studies (n=5 for each blood pressure outcome) and a pooled sample of 295 participants leave the analyses underpowered to detect modest effects. The nurse-led telerehabilitation meta-analysis by Lee et al [68] demonstrated that statistically significant SBP reductions are achievable in disease-targeted populations (MD 10.48 mm Hg in hypertension and diabetes subgroups), suggesting that null pooled effects in mixed CVD populations are likely population- and dose-specific rather than evidence of biological inefficacy. The current evidence therefore neither establishes nor refutes an antihypertensive effect of exercise-based telerehabilitation, and dedicated trials in hypertensive cohorts are needed.

This systematic review and meta-analysis demonstrates that exercise-based telerehabilitation, compared with usual care, produces a probable improvement in CRF in adults with CVD, while providing no convincing evidence of an antihypertensive effect at the pooled level. The strength of these conclusions is tempered by substantial between-study heterogeneity, low to very low GRADE certainty, and a wide prediction interval for VO₂ peak that admits the possibility of null or slightly negative effects in some implementation contexts. None of the 5 prespecified digital health–specific moderators—intervention duration, telemedicine mode, type of guidance, technology delivery platform, or intervention composition—statistically explained the observed heterogeneity, although exploratory subgroup patterns suggest that smartphone- or mHealth-app-based delivery and professional- or technology-system-led guidance may be associated with more favorable and more consistent gains in CRF; these patterns should be regarded as hypothesis-generating rather than definitive.

Taken together, our findings position exercise-based telerehabilitation as a clinically promising but evidence immaturity–related alternative or adjunct to center-based CR. The most consequential gaps in the existing literature are the bundling of exercise with multiple nonexercise components, the underreporting of digital health specific implementation metrics (adherence, engagement, usability, or retention curves), the absence of long-term follow-up, and the lack of dedicated trials in hypertensive cohorts. Resolving these gaps will require component-isolated trials, standardized digital-health reporting, embedded health-economic and patient-experience evaluations, and equity-conscious implementation strategies. Until such evidence is available, exercise-based telerehabilitation should be offered as a patient-centered alternative pathway for individuals who cannot access center-based CR, rather than as a uniformly equivalent substitute, and clinicians, program designers, and policymakers should treat the cardiometabolic effects reported here as an average signal across heterogeneous implementations rather than as a guaranteed effect for every patient and every platform.