Two independent reviewers (YW and MY) embarked on a comprehensive literature (the screening of the literature is shown in Figure 1) search across several databases that span both English (PubMed, Web of Science, Embase, the Cochrane Library, Scopus) and Chinese languages (CNKI, WanFang Data, VIP), concluding their search on July 1, 2023. In instances where disparities emerged, a third-party arbitration (XQ) was sought to reach consensus. The search methodology was thoughtfully devised to encapsulate pertinent keywords related to: (1) Atrial Fibrillation, (2) Exercise, (3) Quality of Life, and (4) Randomized Controlled Trials. Further depth was added to our review by meticulously sifting through the reference lists of articles assimilated into this NMA, as well as prior reviews, with a view to discern any additional studies of relevance that may have been overlooked in the initial search.

The present study undertakes an in-depth exploration of the implications of physical exercise on the HRQoL in patients diagnosed with AF. The selection criteria for this investigation were stringently defined to ensure robustness and specificity of the results. These criteria encompassed: (1) Inclusion of patients diagnosed with atrial fibrillation; (2) Restriction to RCTs to ensure high levels of evidence; (3) Examination of physical exercise interventions, unrestricted by intensity, duration, or frequency, offering broad implications for varying exercise regimes; (4) Comparative analysis of exercise interventions against other categories or control groups receiving standard care; (5) Primacy given to the evaluation of HRQoL as the primary outcome measure.

Studies were excluded under the following conditions: (1) Where physical exercise was amalgamated with other multidisciplinary interventions, such as pharmacological treatments or nutritional modifications, as these could potentially dilute the impact of exercise alone; (2) Where the interventions were purely educational in nature, as the study seeks to gauge the impact of physical activity; (3) Where the type or nature of exercise was indeterminate, to ensure the reproducibility and specificity of findings; (4) Where there was inadequate reporting of data, precluding reliable calculation of effect size; (5) Where the source of data was a conference abstract lacking full-text publication, to maintain the comprehensive and robust nature of the data; (6) Where publications were not scribed in either English or Chinese, to ensure consistency and reliability of data interpretation.

Employing a prespecified table, both reviewers (ZY and XQ) undertook an independent extraction of the data, with any ensuing disputes deftly resolved by a senior researcher, thereby ensuring an unbiased and rigorous process. The extracted data encompassed both demographic and outcome parameters, inclusive of the year of publication, country of origin, study design, sample size, age range of the participants, nature of the interventions and controls, as well as the outcome measures (quality of life, physical components, mental components, among others).

In this NMA, the exercise intervention modalities were meticulously classified as follows: •

Aerobic Exercise: Interventions designed to elevate heart rate and energy expenditure through sustained, rhythmic activities, such as cycling, running, and brisk walking.

In all included studies, the HRQoL outcomes were gauged through one or more self-reported questionnaires, with a majority indicating that higher scores denote superior HRQoL. However, in the event of a study utilising an inverse scoring system (higher scores representing poorer HRQoL), the mean value of each group was multiplied by −1 to align with the conventional interpretation. When a composite HRQoL score was not provided by the scale in use (for instance, in the SF-36 questionnaire), we consolidated the individual dimensions into a single composite measure, concurrently calculating its standard mean deviation. Ultimately, the analyses incorporated overall quality of life, physical components, and mental components, thereby providing a comprehensive and multidimensional evaluation of HRQoL.

Our research team is steadfastly dedicated to advancing diversity, equity, and inclusion within the scope of our clinical endeavors, research initiatives, and educational programs. We recognize that systemic disparities and inherent biases have the potential to influence the integrity of the research process. In response to these challenges, we have initiated measures to mitigate these concerns. Accordingly, the procedures of data extraction, processing, and interpretation have been conducted with minimal constraints to ensure that the outcomes are a comprehensive and authentic representation of reality, embracing the broadest possible spectrum of diversity to enhance the accuracy and inclusivity of our findings.

Two researchers (MY and YW) independently appraised the risk of bias inherent in the incorporated RCTs, employing the Cochrane Collaboration's Risk of Bias 2 tool (RoB-2) as a gold-standard measure (18). Any divergences in assessment were addressed through consensus-building with a third reviewer (XQ), ensuring an unbiased and rigorous appraisal process. The RoB-2 tool methodically dissects the risk of bias across five distinct domains: (1) Bias emerging from the randomization process; (2) Bias attributed to deviations from intended interventions; (3) Bias originating from missing outcome data; (4) Bias in the measurement of outcomes; and (5) Bias in the selection of the reported result. The cumulative risk of bias was categorized as “low” if every domain of the study was scored as such. If at least one domain raised “some concerns”, the overall bias was labelled correspondingly. The study was classified as having a “high risk of bias” if any single domain was rated “high risk” or if several domains elicited “some concerns” in a manner that could potentially undermine the validity of the findings.

The certainty of the evidence concerning the primary outcomes (namely, efficacy, acceptability, and safety) within the network estimates was scrutinized through the application of the Grading of Recommendations Assessment, Development, and Evaluation (GRADE) system (19). Within the purview of the GRADE framework, the quality of evidence is stratified into four tiers: high, moderate, low, and very low. This is premised upon a comprehensive evaluation of factors such as study limitations, inherent risk of bias, inconsistency, indirectness, imprecision, and the potential for publication bias.

This study's principal analysis was conducted using the netmeta and gemtc packages in R 4.3.1. The Standard Mean Difference (SMD) was employed as the effect measure. Statistical significance was denoted by a P-value less than 0.05 and a 95% credible interval (CI) that excludes zero. The analytical process was divided into seven sections:

First, a network geometry plot was utilized to assess the strength of the existing evidence, where the size of the nodes was proportional to the number of studies included for each intervention, and the width of the lines connecting the nodes was proportional to the number of trials directly comparing two interventions (20).

Second, we evaluated consistency by comparing whether the intervention effects estimated by direct comparisons were in alignment with those estimated by indirect comparisons.

Third, a standard pairwise meta-analysis was performed for each direct comparison using the DerSimonian-Laird random-effects method (21). The standardized mean difference score was calculated using Cohen's d as the effect size statistic. Effect sizes were interpreted as follows: values ≤0.2 were considered low, 0.2–0.5 moderate, 0.5–0.8 strong, and ≥0.8 very strong. Statistical heterogeneity was assessed with the I² statistic, with interpretation as follows: I² = 0% to 40% signifying unimportant, I² = 30%–60% moderate, I² = 50%–90% substantial, and I² = 75%–100% considerable heterogeneity. The corresponding p-values were also considered (17). To understand the size and clinical relevance of heterogeneity, we computed the τ² statistic, with interpretation as follows: τ² ≤0.14 low, 0.14–0.40 moderate, and ≥0.40 high degree of heterogeneity. These findings were depicted in a league table format.

Fourth, we assessed transitivity by testing whether the syntheses of direct comparisons of interventions were carried out in samples with similar baseline clinical characteristics.

Fifth, the relative ranks of the exercise interventions were estimated using the treatment rankings and graphically represented using ranking plots. Furthermore, the Surface Under the Cumulative Ranking curve (SUCRA) for each intervention was calculated. SUCRA assigns a value between 0 and 1 to simplify the ranking of each intervention in the ranking plots. The best intervention receives a SUCRA value near 1, while the worst intervention receives a value close to 0 (20, 22). These data were also represented using a rank-heat plot according to the SUCRA values (23).

Additionally, to evaluate the robustness of the estimates and ascertain whether a particular study accounted for a large portion of the heterogeneity, a sensitivity analysis was carried out, sequentially omitting individual studies. A separate sensitivity analysis also excluded studies with a high risk of bias.

Finally, to investigate the presence of bias due to small-study effects, a network funnel plot was utilized for a visual inspection of symmetry. This portion was analyzed using Stata 17.0 (Stata, College Station, TX, USA).

Upon the evaluation through the RoB 2 tool, as depicted in Figure 2, six studies were identified as having low risk, while five studies exhibited some concerns. A single study was classified as high risk. In distinct domains, eight studies demonstrated some concerns regarding the randomization process. Regarding the deviations from intended interventions, four studies displayed some concerns, and two studies were identified with a high risk of bias. Pertaining to the measurement of outcomes, one study expressed certain reservations. The GRADE evaluations are in Supplementary Table S1.

The network geometry diagram visualizes the relative volume of available evidence regarding the impact of sports exercise interventions on overall, physical, and mental HRQoL, encompassing 5, 4, and 4 pairwise comparisons respectively. It is essential to note that all interventions have at least one direct comparison with the control group. As shown in Figure 3, the size of the node represents the number of trials included per intervention, with the line width corresponding to the number of studies making direct comparisons between two particular interventions. The risk of bias assessment is depicted via color coding: green signifies low risk, yellow represents moderate risk, and red indicates a high risk of bias. The coloration of the connecting lines signifies the average risk of bias assessment for the studies making direct comparisons between two interventions.

In Table 2, in pairwise comparisons using SMD as the effect measure, no types of exercise interventions demonstrated a significant impact on HRQoL and its related dimensions. However, in the NMA, we found that aerobic exercise and CR had a significant effect on HRQoL, as well as on its physical and mental components. Specifically, aerobic exercise showed the most considerable effect when compared to usual care, with SMD values for HRQoL at 0.60 (95% CI: 0.02, 1.13), for the physical component at 0.47 (95% CI: 0.04, 0.87), and for the mental component at 0.56 (95% CI: 0.00, 1.05).

In Figure 4, regarding total HRQoL, the highest SUCRA is observed for CR (76.7%). In the physical component, CR (77.0%) again presents the highest SUCRA. For the mental component, HIIT (90.7%) achieves the highest SUCRA. The ranking diagrams for these are depicted in Supplementary Figure S1.

Due to the absence of studies for each subgroup, it's impossible to conduct a priori subgroup analysis based on the duration, frequency, volume, and intensity of the physical activity interventions.

In the sensitivity analysis, we excluded studies with high risk. Although the pooled SMD experienced some variation, there was no statistically significant alteration in comparison to the NMA incorporating all studies.

The heterogeneity assessment, as shown in Supplementary Table S1, reveals moderate heterogeneity in the total Health-Related Quality of Life (HRQoL) between the Aerobic Exercise and Usual Care groups (I² = 51.1, τ² = 0.049). This moderate level of heterogeneity suggests variability in how individuals with atrial fibrillation respond to Aerobic Exercise, potentially influenced by differences in baseline health status, exercise tolerance, or lifestyle factors. In contrast, the comparison between Cardiac Rehabilitation (CR) and Usual Care shows substantial heterogeneity (I² = 75.5, τ² = 0.116), implying that the effectiveness of CR varies significantly among different patient populations. This high level of heterogeneity could be attributed to variations in CR program content, intensity, or adherence, as well as differences in patient demographics or comorbid conditions. Regarding the physical component of HRQoL, substantial heterogeneity is also noted between the CR and Usual Care groups (I² = 75.6, τ² = 0.113). This indicates that physical health outcomes from CR may vary widely, likely due to differences in patients' physical capabilities, comorbidity burden, or specific exercise regimens within CR programs. For the mental component, moderate heterogeneity is observed in the Aerobic Exercise vs. Usual Care comparison (I² = 43.5), while substantial heterogeneity is present in the CR vs. Usual Care comparison (I² = 62.6). The substantial heterogeneity in CR's effect on mental health suggests that psychological outcomes may vary based on factors such as baseline mental health, support systems, or program engagement levels.

The transitivity assessment, Table 1 provides a comprehensive summary of the included studies, showing comparable characteristics across study populations and interventions. All studies focus on adult populations with similar baseline characteristics, ensuring that age-related variability is minimized. The exercise intervention durations range from 3 to 6 months, with each session lasting no more than one hour. These consistent parameters across studies suggest that populations and intervention durations are sufficiently similar, supporting the assumption of transitivity in this network meta-analysis.

Finally, the publication bias was identified through the utilization of Egger's test, while comparing the effects of Yoga vs. Usual Care on the mental component (P = 0.039). The graphical representation of the bias can be observed in the funnel plots (Supplementary Figure S3).