Eligibility criteria were established according to the PICOS framework.

Eight databases were systematically searched: PubMed, Embase, Cochrane Library, Web of Science, Scopus, CNKI, Wanfang, and VIP. Two reviewers independently conducted comprehensive searches combining controlled vocabulary (MeSH terms) and free-text keywords related to spinal cord injury and respiratory dysfunction. The search covered all records from database inception up to July 12, 2025, without language restriction. Reference lists of included studies and relevant reviews were manually screened to ensure completeness. The detailed PubMed search strategy is presented in Supplementary Table 2. Full search strategies for all databases are available upon request.

Two reviewers independently screened the literature and extracted data using EndNote X9 for reference management. Discrepancies were resolved by consensus or consultation with a third reviewer. Standardized data extraction forms were used to collect study characteristics and outcome variables. When data were incomplete, corresponding authors were contacted by email. For inconsistent outcome units, data were standardized using the methods of Luo et al. (10), Wan et al. (11), and Shi et al. (12). Numerical data presented in figures were digitized using GetData Graph Digitizer, and all results were expressed as mean ± standard deviation (SD) for analysis.

Risk of bias was assessed using Review Manager 5.3, based on the Cochrane Risk-of-Bias Tool with seven domains rated as low, unclear, or high risk. Methodological quality of included RCTs was further evaluated using the Physiotherapy Evidence Database (PEDro) scale, consisting of 11 items (10 scored). Studies were classified as high quality (≥7), moderate (5–6), or low (≤4). Detailed results are presented in Supplementary Table 3.

A Bayesian network meta-analysis (NMA) was performed using the gemtc and multinma packages in R (version 4.4.0) and Stata/MP 14.0. Continuous outcomes were expressed as mean difference (MD) with 95% credible intervals (CrI). Four Markov chains were run with 100,000 iterations (20,000 burn-in, 80,000 sampling). Convergence was verified by trace and density plots and confirmed when the potential scale reduction factor (PSRF) was <1.05. Global consistency was assessed by comparing the Deviance Information Criterion (DIC) between consistency and inconsistency models (ΔDIC < 5 indicating good model fit). Node-splitting analysis was used to detect local inconsistency (p < 0.05 = significant). Surface under the cumulative ranking curve (SUCRA) values were used to rank the relative effectiveness of interventions, with higher SUCRA values indicating a greater probability of being the most effective. Publication bias was assessed using comparison-adjusted funnel plots and Egger’s test (α = 0.05). Heterogeneity was evaluated using the I² statistic, with values >50% considered substantial. Sensitivity analyses were conducted by excluding low-quality studies to assess the robustness of results.

In accordance with recommended principles for network meta-analysis, the plausibility of the transitivity assumption was considered a priori. Potential effect modifiers—including neurological level and completeness of spinal cord injury, injury phase (acute/subacute vs. chronic), intervention dose (training intensity or load, frequency, and duration), and concomitant rehabilitation—were identified based on clinical relevance. These characteristics were extracted and summarized at the study level to facilitate qualitative assessment of clinical comparability across intervention nodes. Subgroup analyses or network meta-regression were planned a priori but were to be conducted only if sufficient numbers of studies with consistently reported data were available within each intervention node. As these predefined conditions were not met, subgroup analyses and network meta-regression were not performed.

A total of 2,866 records were initially identified through comprehensive searches of five English databases (PubMed, Embase, Web of Science, Cochrane Library, and Scopus) and three Chinese databases (CNKI, Wanfang, and VIP), supplemented by manual searches of reference lists from high-quality studies. After automatic and manual deduplication, titles and abstracts were screened for relevance, followed by full-text assessment for eligibility. Finally, 40 randomized controlled trials met the inclusion criteria and were included in the network meta-analysis. The detailed study selection process is presented in Figure 1.

A total of 40 randomized controlled trials published between 1992 and 2023 were included, comprising two three-arm studies and 38 two-arm studies. The sample sizes of individual trials ranged from 10 to 136 participants, with a total of 1,878 patients—969 in the intervention groups and 909 in the control groups. All participants had spinal cord injuries at or above the thoracic segment.

Interventions applied in the experimental groups primarily included progressive resistance breathing training (PRT), resistive inspiratory muscle training (RIMT), extracorporeal diaphragmatic pacing (EDP), abdominal compression (AC), singing therapy (ST), normocapnic hyperpnoea (NH), aerobic training (AT), and the traditional Chinese breathing exercise Liuzijue (LZJ), which combines specific diaphragmatic breathing techniques with vocalizations, is designed to improve pulmonary function and respiratory health. By coordinating breathing with controlled vocal sounds, Liuzijue enhances diaphragmatic movement, optimizes lung compliance, and promotes effective airflow, which in turn improves ventilation and respiratory endurance. The control groups typically received routine care, conventional rehabilitation, placebo, sham training, or a different respiratory rehabilitation regimen from the experimental groups.

Detailed characteristics of all included studies are summarized in Table 1, and the definitions and protocols of the various respiratory rehabilitation interventions are provided in Supplementary Table 2.

The methodological quality of the included randomized controlled trials was evaluated using the Physiotherapy Evidence Database (PEDro) scale. Among the 40 included studies, 15 were rated as high quality and 25 as moderate quality, with a mean PEDro score of 6.8 ± 1.24 (range: 6–10). Detailed quality assessment results for each study are presented in Supplementary Table 3.

All 40 included studies reported appropriate random sequence generation and were therefore judged as low risk for selection bias. Only five studies explicitly described allocation concealment, while the remainder did not, and were thus rated as unclear risk.; Regarding blinding, four studies (13–16) stated that participants were not blinded, and were consequently assessed as high risk. Fourteen studies explicitly reported blinding of outcome assessors and were judged as low risk, whereas the remaining studies provided insufficient information and were rated as unclear risk. No studies showed evidence of incomplete outcome data, selective reporting, or other potential sources of bias; therefore, these domains were assessed as low risk across all trials. The overall risk-of-bias summary and graph are presented in Figures 2, 3, respectively.

Compared with the control group, abdominal compression training (AC) (MD = 0.44, 95% CrI 0.09–0.79), Liuzijue (LZJ) (MD = 0.97, 95% CrI 0.57–1.37), progressive resistance breathing training (PRT) (MD = 0.49, 95% CrI 0.19–0.78), and resistive inspiratory muscle training (RIMT) (MD = 0.49, 95% CrI 0.22–0.77) significantly improved FVC.

According to SUCRA rankings, Liuzijue (95.8%) had the highest probability of being the most effective, followed by RIMT (60.6%), PRT (59.8%), normocapnic hyperpnoea (NH, 59.0%), aerobic training (AT, 57.9%), AC (54.2%), extracorporeal diaphragmatic pacing (EDP, 57.4%), singing training (ST, 39.5%), control (13.9%), and comprehensive respiratory rehabilitation (CRR, 6.0%) (Supplementary Table 4A; Figure 6A).

Potential publication bias was examined using funnel plots generated in Stata/MP 14.0 for each outcome. As shown in Figure 7, the data points were largely symmetrically distributed within the funnel boundaries, with only a few studies falling outside the confidence region, suggesting minimal small-study effects.

To further verify this observation, Egger’s test was performed for all four outcomes, yielding p-values of 0.328 (FVC), 0.912 (FEV₁), 0.096 (MIP), and 0.099 (Borg), all of which exceeded 0.05. These results indicate no statistically significant publication bias, confirming the robustness of the pooled estimates.

Previous studies have shown that pulmonary ventilatory dysfunction (PVD) develops in varying degrees depending on the level of spinal cord injury, with higher lesions producing more profound declines in ventilatory capacity (19, 20). PVD impairs cough and secretion clearance, leading to infection and even respiratory failure (21). For this reason, FVC and FEV₁ remain the most representative indices of ventilatory performance in SCI (22, 23). In the current analysis, Liuzijue ranked highest for FVC improvement (95.8%), while AC showed the best efficacy for FEV₁ (91.6%).

Liuzijue is a traditional breathing exercise that combines diaphragmatic breathing with pursed-lip expiration and coordinated limb movement. This pattern improves diaphragmatic excursion and lung compliance, optimizing tidal ventilation and pulmonary mechanics (24, 25).

Abdominal compression training, implemented through banding, manual pressure, or biofeedback systems (26–29), enhances expiratory strength by increasing intra-abdominal pressure and facilitating diaphragmatic elevation (30, 31). This repetitive pressurization directly stimulates the abdominal musculature, reinforcing expiratory flow and aiding airway secretion clearance. Consequently, FEV₁ improvement may reflect both increased expiratory muscle recruitment and reduced airway resistance.

Approximately two-thirds of SCI patients with dyspnea exhibit inspiratory muscle weakness due to paralysis of the diaphragm or intercostal muscles (32). In this context, PRT emerged as the most effective intervention for improving MIP (87.3%). As a form of inspiratory muscle training (IMT), PRT employs graded pressure thresholds to induce adaptive hypertrophy and endurance in respiratory muscles (8).

Repeated resistance loading enhances the cross-sectional area of muscle fibers, particularly in the diaphragm and external intercostals, which increases contractile velocity and strength (33). This hypertrophy improves respiratory muscle power and contributes to greater force production during inspiratory efforts. Additionally, PRT improves neuromuscular coordination, optimizing diaphragm contraction and increasing inspiratory pressure and endurance. Furthermore, PRT increases oxidative capacity in the respiratory muscles, improving their endurance during prolonged inspiratory efforts. Several included trials also reported increased maximal expiratory pressure and reduced pulmonary infection incidence following PRT, indicating that this intervention supports both inspiratory and expiratory respiratory function through enhanced muscular control and airway clearance (15, 16, 34–36).

Dyspnea is one of the most distressing and life-limiting symptoms of cervicothoracic SCI (37). Using the Borg scale (38), this analysis showed that NH achieved the greatest improvement in perceived dyspnea (SUCRA = 99.8%). The included protocols applied sustained hyperventilation at 30–50% of maximal voluntary ventilation with visual and auditory feedback (39).

Mechanistically, NH trains patients to maintain deep, rhythmical breathing, improving alveolar ventilation and oxygen–carbon dioxide exchange (40). The enhanced gas exchange reduces CO₂ retention and respiratory effort, while promoting a more efficient and economical breathing pattern. Beyond physiological benefits, NH also improves patient confidence and tolerance to physical activity, contributing to higher quality of life and reduced anxiety related to breathlessness (41–43). These findings suggest that NH is an effective and accessible strategy for mitigating dyspnea in both acute and chronic SCI phases. However, perceived dyspnea is influenced not only by ventilatory mechanics but also by psychological and contextual factors, which should be considered when interpreting these results.

This review has several limitations. First, substantial clinical heterogeneity existed across included trials with respect to neurological level and completeness of injury, injury chronicity, intervention dose (training intensity, frequency, and duration), and the presence of concomitant rehabilitation. Although statistical assessments indicated no major inconsistency in closed-loop networks, such variability may act as an effect modifier and influence indirect comparisons. In principle, such effect modification could be explored using subgroup analyses or network meta-regression. Due to inconsistent reporting of key clinical variables and limited numbers of studies within some intervention nodes, subgroup analyses and network meta-regression could not be reliably performed.

Most participants were male, potentially limiting generalizability. Intervention frequency, intensity, and duration varied across studies, introducing methodological heterogeneity. Accordingly, SUCRA values should be interpreted as probabilistic rankings rather than indicators of absolute clinical superiority, particularly when differences between interventions are small or when evidence is derived from a limited number of studies. Additionally, some promising modalities—such as aquatic therapy and combined respiratory–neuromuscular stimulation—were insufficiently studied to be included. Future research should establish standardized intervention protocols, explore dose–response relationships, and include longitudinal follow-up to evaluate the persistence of benefits. Further studies integrating respiratory mechanics, muscle performance, and quality-of-life measures may better clarify optimal rehabilitation sequencing and combination strategies.

In addition, the Borg dyspnea network was informed by a relatively small number of studies and did not form a closed-loop structure, reducing the certainty of indirect comparisons for this outcome.

Due to inconsistent reporting of key clinical variables and limited numbers of studies within some intervention nodes, subgroup analyses and network meta-regression could not be reliably performed.