This study was conducted in accordance with the PRISMA-NMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses extension for Network Meta-Analyses), PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses), and PRISMA-S (Preferred Reporting Items for Systematic Reviews and Meta-Analyses – Search extension) 2020 guidelines [18-20]. A completed PRISMA 2020 27-item checklist is provided in Tables S1-S3 in Multimedia Appendix 1. The study protocol was registered with PROSPERO under registration number CRD420251146712.

Two independent reviewers applied the same predefined search strategy across 4 electronic databases, without restrictions on language or publication date: PubMed, the Cochrane Library, Web of Science, and Embase. The search strategy incorporated both Medical Subject Headings (MeSH) and free-text terms, with “telerehabilitation” and “nonspecific low back pain” serving as the key concepts. The search was reported in accordance with PRISMA-S to ensure that each component was fully described and reproducible [20]. Any discrepancies between the reviewers were resolved through discussion and consensus involving a third reviewer. In addition, manual searches of the reference lists of relevant meta-analyses and review articles were conducted to identify additional potentially eligible studies. The initial search was conducted on September 30, 2025, following iterative refinement of the search strategy. An updated search was performed on April 30, 2026, to capture the most recent literature (Table S4 in Multimedia Appendix 1).

The inclusion and exclusion criteria were established according to the Population, Intervention, Comparator, Outcome, and Study Design framework. We included the following definitions: telerehabilitation involves participants performing the prescribed exercises by watching training videos only, without any real-time AI guidance or feedback. TLRH-AI incorporates AI-enabled posture recognition, real-time movement correction, or instant scoring feedback. Specifically, during human–computer interaction, the AI identifies 30 body key points to detect movement accuracy and errors, provide real-time guidance, and automatically evaluate task completion (eg, graded as qualified, good, excellent, or outstanding). IPR: structured, face-to-face rehabilitation involving therapist-supervised exercise sessions with real-time guidance and correction. UC: basic health education and nonsystematic exercise advice without structured exercise programs or continuous professional supervision. In the final network, we treated telerehabilitation and TLRH-AI as 2 separate treatment nodes. We classified an intervention as telerehabilitation (non-AI) when rehabilitation exercises were delivered remotely (eg, video conferencing, apps with prerecorded exercise videos, or web-based programs) without AI-based functions. We classified an intervention as TLRH-AI only when the study explicitly reported AI features, such as motion recognition, automated feedback, or algorithm-based personalization/adaptation of the exercise program.

The follow-up time points at 4, 8, and 12 weeks were not selected in a results-driven manner; rather, they were prespecified based on the following criteria: (1) these time points were the most frequently reported follow-up assessments among the included studies, thereby maximizing comparability; (2) they correspond to commonly used and relatively consistent evaluation windows for rehabilitation effects in chronic low back pain research, namely short-term (≤4 weeks), short-to-midterm (8 weeks), and midterm (12 weeks); (3) to avoid double-counting highly correlated repeated measurements within the same study and artificially inflating its weight, we prespecified that only 1 assessment time point per study would be included within each time window, selecting the follow-up closest to the target window for analysis. A detailed description of the selection process is provided in Table S5 in Multimedia Appendix 1.

Two independent reviewers performed the study selection and data extraction. A third reviewer consolidated the results, and any discrepancies were resolved through discussion until consensus was reached. The following data were extracted into a standardized Microsoft Excel sheet: title, first author, publication year, study design, sample size, telerehabilitation technology used, intervention duration, follow-up period, and all reported clinical outcomes. Outcomes of primary interest were pain intensity, disability, kinesiophobia, and health-related quality of life.

The methodological quality and risk of bias (RoB) of the included RCTs were assessed independently by 2 reviewers using the Cochrane Risk of Bias tool (2.0), as implemented in Review Manager (version 5.4.1) [21]. The evaluation covered the following domains: random sequence generation, allocation concealment, blinding of participants and personnel, blinding of outcome assessment, incomplete outcome data, selective reporting, and other potential sources of bias. Disagreements between reviewers were resolved through discussion or by consultation with a third reviewer when necessary. The overall RoB for each study was categorized as “low,” “some concerns,” or “high”. The RoB assessment was performed at the study level (not per outcome/per time point). We also corrected the figure caption to indicate that the RoB summary/graph refers to RCTs, not outcomes. The results of the assessment are presented graphically in the study.

Statistical analyses were performed within a Bayesian framework using R software (version 4.4.1) interfaced with JAGS (version 4.3.1). We used the default vague priors in gemtc: treatment effects followed Normal (0, [15 × om.scale]²), and heterogeneity was specified as τ ~ Uniform (0, om.scale) (τ² implied). In gemtc, om.scale is an internally determined scale parameter based on the observed variability and measurement scale of the outcome, and is used to set weakly informative prior ranges that are appropriate for the analyzed data. Markov Chain Monte Carlo (MCMC) sampling was used to jointly estimate direct and indirect treatment comparisons. A random-effects NMA model was constructed using the gemtc package, and the MCMC sampling was executed via the rjags package. The model was run with 4 chains, each with an initial adaptation phase of 50,000 iterations (discarded as burn-in), followed by 200,000 simulation iterations thinned by a factor of 10 to reduce autocorrelation [22]. We used the standardized mean difference (SMD) for pain intensity because the included trials measured pain using different scales, mainly the Visual Analogue Scale (VAS) and the Numerical Rating Scale) NRS, rather than a single uniform instrument. Using the SMD allows us to standardize effects across different pain scales and combine them within the NMA. Effect sizes were expressed as SMDs with 95% credible intervals (CrIs) for pain intensity, and as mean differences with 95% CIs for other continuous outcomes. Statistical heterogeneity was described using the I² statistic for both pairwise and NMAs, with values greater than 50% considered suggestive of substantial heterogeneity. However, I² was not used as the sole basis for interpreting heterogeneity because it does not directly quantify the absolute variation of true treatment effects across different populations or clinical settings [23]. Therefore, 95% prediction intervals (PIs) were additionally estimated for key network estimates where applicable to aid interpretation of the expected distribution of true effects in future comparable settings.

The 95% CrI was interpreted as the uncertainty around the average network treatment effect, whereas the 95% PI was interpreted as the expected range of true effects that may be observed in a future comparable clinical setting. SEs were derived from the reported 95% CrIs using the formula SE = (upper limit − lower limit) / 3.92. Estimated 95% PIs were then calculated as effect estimate ± 1.96 × √(SE² + τ²), where τ represents the assumed between-study SD on the scale of each outcome. Because posterior estimates of τ were not consistently available for all network comparisons, τ values were specified according to the outcome scale, follow-up window, and evidence structure. These PIs were therefore interpreted as estimated predictive ranges rather than formal model-based posterior PIs [24]. When the average effect favored an intervention, but the corresponding PI was wide or crossed the null value, the finding was interpreted cautiously as potentially variable across future clinical settings. For outcomes where lower scores indicate improvement (pain, disability, and kinesiophobia), effect estimates were oriented so that positive values indicate improvement. Consistency between direct and indirect evidence was evaluated using the node-splitting approach; results were visualized using forest plots. Treatment rankings were inferred using the surface under the cumulative ranking curve (SUCRA) values, derived from the posterior probability matrix of each intervention being the most effective. Rank probability plots were generated to illustrate the likelihood of each treatment being ranked at each possible position. Network graphs and funnel plots were produced using Stata (version 16.0; StataCorp LLC). Small-study effects and funnel plot asymmetry were assessed using funnel plots and the Egger test. Funnel plot asymmetry was not interpreted as publication bias alone, because asymmetry may arise from several sources, including between-study heterogeneity, methodological differences, selective outcome reporting, chance, and other small-study effects [25].

The certainty of evidence for all estimated outcomes was assessed using the GRADE (Grading of Recommendations Assessment, Development, and Evaluation) framework for NMA [26]. This approach evaluates the certainty of the evidence for each pairwise comparison within the network separately, considering both direct and indirect evidence, and then rates the certainty of the NMA effect estimates. The assessment of certainty in effect estimates for direct comparison evidence includes the RoB, inconsistency, indirectness, imprecision, and other considerations. The assessment was performed independently by 2 reviewers, with any disagreements resolved by consensus or by consulting a third reviewer. The certainty was judged as high, moderate, low, or very low. The specific meanings are presented in Table S6 in Multimedia Appendix 1.

A total of 2491 articles were initially retrieved. Among them, 2254 duplicate records were excluded, and 46 articles were excluded during the preliminary screening. The remaining 2491 articles underwent full-text review. Finally, 171 studies did not meet the inclusion criteria and were excluded, leaving 20 RCTs [27-46] with a total of 1854 patients. The included interventions were TLRH-AI, telerehabilitation, IPR, and UC. Study numbers and total sample size in each network comparison are provided in Table S7 in Multimedia Appendix 1. We excluded interventions based on phone or text support, nonexercise virtual reality, online self-management or cognitive behavioral programs without specified exercise content, online consultations without guided exercise, and wearable sensors or applications used only for tracking without exercise guidance. The distribution of major clinical and methodological characteristics across intervention nodes was considered broadly comparable, supporting the plausibility of the transitivity assumption (Table S8 in Multimedia Appendix 1). The estimated 95% PIs for the key network are shown in Table S9 in Multimedia Appendix 1. The PRISMA flow diagram of the literature search is presented in Figure 1.

Telerehabilitation technologies primarily include smartphone apps (providing instructional videos or AI-assisted guidance), video conferencing, and exercise tutorials based on computer or web platforms. The intervention period ranged from 4 to 12 weeks. The control group typically received traditional outpatient physiotherapy or paper-based conventional physiotherapy. Pain intensity was measured using the VAS or the NRS; disability or functional status was assessed using the Oswestry Disability Index (ODI) or the Roland-Morris Disability Questionnaire (RMDQ); fear of movement was evaluated using the Tampa Scale for Kinesiophobia (TSK); and health-related quality of life was measured with the 12-item Short-Form Health Survey (SF-12), including the physical component summary (PCS) and the mental component summary (MCS). Table 1 lists the basic characteristics of the included studies.

The overall methodological quality of the included studies is illustrated in Figure 2 and Figure S1 in Multimedia Appendix 1. Most studies described random sequence generation and allocation concealment methods adequately, demonstrating a low risk of selection bias. Regarding blinding of participants and personnel, all studies were at high risk. This was due to the nature of the interventions: telerehabilitation, such as video guidance, AI-assisted training, or app-based guidance, differed significantly in implementation from IPR or UC, making blinding of participants and therapists practically infeasible.

In terms of blinding of outcome assessment, some studies explicitly reported the use of blinding for outcome assessment; for example, data were collected by researchers unaware of group allocation, and these studies were rated as low risk. Studies that did not report assessor blinding were rated as having an unclear risk. For incomplete outcome data and selective reporting, most studies had low dropout rates and reported prespecified outcome measures, thus being rated as low risk.

This study systematically assessed the certainty of the evidence between different comparison groups using the GRADE framework. In addition, the league matrix tables visually highlight the certainty of the effect estimates using different colors. As shown in all league matrix tables, the certainty of the evidence for all outcomes was rated from very low to high according to GRADE.

Available studies suggested possible benefits of TLRH-AI for reducing pain intensity; however, the certainty of the evidence was low, and these findings were insufficient to support reliable comparative conclusions. In contrast, the certainty of evidence for TLRH at 4, 8, and 12 weeks was rated as high or moderate. IPR demonstrated a strong statistical effect in reducing the primary outcome of disability, but the certainty of the evidence was rated as low. To improve the transparency of the methodology, Table 2 provides a detailed explanation of the reasons for evidence downgrading.

We performed node-splitting consistency tests for comparisons in the network that had both direct and indirect evidence. For the 4-week pain intensity outcome, only UC vs IPR could be evaluated, with a direct effect MD of −1.720 (95% CrI −2.239 to −1.204) and an indirect effect MD of −0.983 (95% CrI −2.032 to 0.061), with a difference of approximately 0.74 between the two. The inconsistency test showed no significant difference (P=.22).

For the 4-week ODI outcome, evaluable comparisons included telerehabilitation vs TLRH-AI, IPR vs TLRH-AI, IPR vs TLRH, and UC vs IPR. Node-splitting tests for all comparisons did not reach statistical significance (P=.19, P=.19, P=.17, and P=.05, respectively). Among them, the direct effect for UC vs IPR was −12.63 (95% CrI −15.93 to −9.345) and the indirect effect was −3.715 (95% CrI −12.16 to 4.665), with a difference of approximately 8.92, representing a potentially important inconsistency, even though its statistical test did not reach significance.

For the remaining comparisons, differences between direct and indirect estimates were approximately 2.42 to 2.74. For the TSK outcome, only UC vs IPR could be evaluated, with a direct effect of −2.269 (95% CrI −3.052 to −1.481) and an indirect effect of −0.662 (95% CrI −3.726 to 2.388), with a difference of approximately 1.61. The inconsistency test showed no significant difference (P=.32). For the RMDQ outcome, evaluable comparisons included telerehabilitation vs TLRH-AI, IPR vs TLRH-AI, and IPR vs TLRH, with P=.60, P=.61, and P=.60, respectively, showing no statistically significant inconsistency. Differences between direct and indirect estimates were approximately 0.74, 0.72, and 0.73, respectively, suggesting generally good consistency within this outcome. Overall, node-splitting tests for all evaluable comparisons did not reveal statistically significant inconsistency, suggesting acceptable overall network consistency. However, in some comparisons, particularly UC vs IPR for the 4-week ODI outcome, a potentially important deviation between direct and indirect estimates was observed. Therefore, these network estimates should be interpreted cautiously, especially considering the limited statistical power, the quantity of available evidence, and the width of the CrIs.

This Bayesian NMA compared the relative effects of TLRH-AI, telerehabilitation, IPR, and UC in patients with CNSLBP. The main findings were that supervised exercise-based interventions were generally associated with better outcomes than UC. Telerehabilitation showed the most favorable effect on pain reduction at 12 weeks and improved the physical component of health-related quality of life, whereas IPR appeared to provide greater benefits for disability and kinesiophobia. Findings related to TLRH-AI were exploratory only and, because of the low certainty of evidence, were insufficient to support reliable comparative conclusions. Overall, these findings support an individualized rehabilitation strategy according to patients’ pain, functional, and psychological recovery needs.

Our findings showed that, compared with UC, all supervised exercise-based interventions were associated with significant improvements in pain and disability [4-6,47]. This result is consistent with previous evidence supporting exercise-based management as a core component of CNSLBP care, not merely a simple alternative or an inferior option. Importantly, telerehabilitation appeared to achieve outcomes comparable to those of conventional outpatient rehabilitation in pain relief and functional improvement, suggesting that remote delivery of structured exercise programs may represent a feasible and effective alternative to IRP in selected populations. Previous studies have similarly reported no significant differences between app-based remote exercise training and traditional outpatient exercise therapy in terms of pain intensity and ODI improvement, further supporting the clinical utility of remotely supervised rehabilitation [48].

The relatively sustained 12-week benefit of telerehabilitation observed in this study may reflect mechanisms that are particularly relevant to chronic pain self-management, including repeated home-based practice, lower access barriers, and better integration of exercise into daily routines. This interpretation is also supported by recent clinical studies of exercise-based telerehabilitation in chronic low back pain, which have reported meaningful improvements in pain and related outcomes after structured remote programs [33]. By contrast, the short-term benefit observed with IPR may reflect the immediate therapeutic advantages of face-to-face supervision. The short-term pain-relieving effect of IPR may partly arise from its ability to reduce patients’ uncertainty and concerns about safety during the early phase of rehabilitation. In IPR treatment settings, therapists can provide immediate reassurance, clarify pain-related concerns, correct maladaptive movement patterns in real time, and build a stronger therapeutic alliance [49]. Therefore, the early benefits of IPR may reflect not only the exercise intervention itself, but also the added value of direct supervision, reassurance, and psychologically supportive care.

TLRH-AI performs in reducing pain intensity and improving health-related quality of life. However, this finding should be considered hypothesis-generating rather than conclusive, and may also reflect sparse-data effects or a winner’s curse. This advantage may be related to the ability of AI-enhanced systems to provide personalized feedback, real-time monitoring, and adaptive exercise progression. From a digital health perspective, these features are particularly relevant because they address some of the major limitations of conventional telerehabilitation, such as insufficient interactivity, delayed feedback, and limited monitoring of exercise quality. In this context, AI may enhance both the precision of exercise execution and patient adherence, thereby contributing to superior short-term symptom control. This interpretation is consistent with the increasing emphasis on personalization in digital rehabilitation, where AI-driven platforms are expected to improve intervention responsiveness through motion recognition, performance tracking, and tailored progression adjustment [49,50].

The findings for disability require cautious interpretation because they varied according to both the follow-up time point and the instrument used. On the ODI, IPR showed the greatest benefit at 4 and 12 weeks, whereas telerehabilitation ranked highest at 8 weeks, suggesting that the relative functional effects of different rehabilitation modalities may change over time. On the RMDQ, however, IPR showed the most consistent advantage. This pattern may reflect not only genuine differences in treatment effects, but also differences in what the 2 instruments capture, as the ODI covers a broader set of domains, including pain intensity, physical functioning, sleep, and social functioning [51], whereas the RMDQ was designed to capture the impact of back pain on everyday functioning and physical activities [52]. Accordingly, although the overall direction of evidence suggests that IPR may have an advantage for disability improvement, the inconsistency between ODI and RMDQ means that this conclusion should not be overstated. Rather than supporting a single, unequivocal treatment hierarchy for disability, these results suggest a more nuanced interpretation in which the comparative benefit may depend on both the stage of recovery and the outcome measure selected. This cautious interpretation is also supported by prior methodological work showing that the ODI and RMDQ cannot be used interchangeably and may not measure exactly the same construct [53]. Prior studies have likewise suggested that although remote supervision can alleviate pain and improve function, IRP may be more effective in restoring physical function and reducing kinesiophobia in patients with chronic low back pain [10,28,54]. Therefore, while digital rehabilitation may enhance accessibility and convenience, in-person care continues to play an essential role in addressing the behavioral and psychological dimensions of recovery. Interestingly, when disability was assessed using the RMDQ, available studies of TLRH-AI suggested possible benefits for disability improvement. The apparent superiority of TLRH-AI over IPR on RMDQ may partly reflect baseline differences, with TLRH-AI trials enrolling participants with milder disability (mean RMDQ ~9.2) compared to IPR trials (mean ~11.8 for IPR). This differential severity at baseline may explain why different instruments favor different interventions. These findings may indicate that TLRH-AI has the potential to support functional gains related to everyday movement performance, possibly through immediate, individualized guidance delivered in the home setting. Such support may help patients integrate therapeutic exercise more effectively into daily routines and maintain higher-quality practice outside clinical environments. However, because the evidence was limited and of low certainty, these findings should be regarded as exploratory only and not as proof of superior effectiveness.

Another notable finding is that the telerehabilitation group maintained favorable effects at 12 weeks after intervention, suggesting that structured remote exercise programs may have relatively good stability over time. Even without AI support, well-designed home-based rehabilitation protocols may help patients sustain exercise habits and reinforce core muscle function, thereby contributing to ongoing pain relief and functional improvement [55]. This observation has practical implications, as it suggests that standard telerehabilitation may still represent a valuable option when AI-supported systems are unavailable or when health care resources are limited.

The rehabilitation modalities may influence different dimensions of recovery rather than exerting a uniform effect across all outcomes. The apparent advantage of IPR for kinesiophobia may suggest that fear of movement is particularly sensitive to interventions that rebuild movement confidence through repeated therapist-guided exposure and behavioral recalibration [56,57]. In this sense, the benefit of IPR may lie not only in symptom reduction, but also in helping patients reinterpret movement as safe and manageable. By contrast, the clearer benefit of telerehabilitation for the physical component of health-related quality of life may indicate that remote rehabilitation is especially well suited to supporting physical self-management in daily life. Home-based telerehabilitation programs can provide structured exercise plans, flexible scheduling, and ongoing remote guidance, which may facilitate the transfer of exercise behaviors into routine activities and reinforce patients’ perceptions of physical capability outside the clinic [11,58]. This distinction may help explain why IPR appeared more favorable for kinesiophobia, whereas telerehabilitation showed a clearer signal for physical health status. At the same time, the nonsignificant trend toward improvement in the mental component with IPR suggests that face-to-face rehabilitation may still have psychological value, as therapeutic alliance and patients’ experiences of interactional support are regarded as meaningful aspects of low back pain care and have been linked to better outcomes in musculoskeletal rehabilitation [59,60]. Overall, these findings may reflect not a simple hierarchy of treatment superiority, but a domain-specific pattern in which different rehabilitation strategies are more responsive to different aspects of recovery.

Taken together, these findings suggest that telerehabilitation and IPR should not be viewed as competing treatment modalities, but rather as complementary components within a flexible and individualized rehabilitation framework for CNSLBP [61]. Telerehabilitation may be particularly valuable for improving access to structured exercise and maintaining rehabilitation in patients who face barriers to clinic attendance, such as mobility limitations, geographic constraints, or limited health care resources. In contrast, IPR may offer added value when closer supervision, movement retraining, functional restoration, or greater psychological support is required, especially for patients with persistent mobility impairment or fear-avoidant behaviors [3]. Within this framework, the choice between telerehabilitation and IPR should be guided not by a rigid hierarchy, but by the patient’s clinical needs, recovery stage, and practical circumstances [62,63].

Beyond comparative effectiveness, these findings also have implications for the implementation of digital rehabilitation in real-world care. Although AI-enhanced telerehabilitation appears promising, its effectiveness should be interpreted alongside issues of accessibility, digital literacy, patient engagement, and health equity. The implementation of telerehabilitation should also address long-term sustainability, the need for ongoing technical support, and integration with existing electronic health record systems [64]. These factors may influence not only adoption and scalability, but also workflow burden and care coordination in routine clinical practice [65,66].

Future telerehabilitation research should adopt more structured and reproducible approaches to reporting ethical and equity-related dimensions and should prioritize stratified analyses to clarify how race, income, age, education, and other intersecting factors influence access to, engagement with, and outcomes of digital rehabilitation interventions. Such efforts are essential to ensure that telerehabilitation is not only clinically effective but also equitable, scalable, and sustainable across diverse patient populations.

This study has several limitations. First, a major limitation is substantial heterogeneity in intervention protocols (duration 4-12 weeks, varying session frequency, and diverse technologies), which may violate the transitivity assumption fundamental to NMA. This means observed effect estimates may reflect specific protocol characteristics as much as intervention modality, limiting clinical exchangeability across studies. Ideally, sensitivity analyses would explore this issue, but the sparse network structure made such analyses infeasible without fragmenting evidence into unstable estimates. Therefore, findings should be interpreted as indicative trends rather than definitive comparative effectiveness rankings.

Second, assessing publication bias in NMA is challenging, particularly due to sparse evidence and stratification by time points. As a result, we did not conduct specific bias analyses for NMA, such as funnel plot correction, and the results of traditional funnel plots and the Egger test should be interpreted with caution. While funnel plot analysis can provide preliminary indications of bias, its statistical power is limited in networks with a small number of studies, and potential bias cannot be completely ruled out.

Furthermore, the SUCRA rankings reflect the relative trends in the effectiveness of different rehabilitation strategies, rather than deterministic rankings of effectiveness. Since there is currently no widely accepted method for constructing CIs for SUCRA values, they may be misleading in some cases, particularly when the evidence base is weak. As such, we present SUCRA as descriptive supplementary information, rather than a conclusive measure, and have emphasized its interpretative limitations in the discussion.

There are also issues regarding the choice of measurement tools. The ODI scale focuses more on physical function limitations, while the RMDQ scale is more sensitive to limitations in daily activities. The discrepancies in the results of these 2 scales suggest that they reflect outcomes in different ways. Regarding potential influencing factors, such as age, pain duration, and baseline severity, we acknowledge their clinical significance. However, due to our inability to obtain individual participant data, we were unable to conduct reliable formal subgroup analyses or NMA based on these factors.

Regarding missing data, we used complete outcome data from the original trial reports for analysis. Since only aggregated data were available and we could not verify the assumptions necessary for effective imputation, multiple imputation methods were not applicable. To ensure the reliability of the analysis, we assessed the convergence of the MCMC chains through visual inspection of the prespecified number of iterations and sampling behavior. This method aligns with standard practices for Bayesian NMA using gemtc and JAGS. However, detailed convergence diagnostic metrics, such as R-hat values, were not formally extracted, and this methodological limitation should be considered.

As for cost-effectiveness analysis, although this study included various rehabilitation interventions, most studies did not systematically report cost or cost-effectiveness data. The existing descriptive information is insufficient to support formal economic evaluations. We call for future research to prioritize conducting cost-effectiveness analysis alongside clinical trials, using a mixed-effects implementation research design to more accurately assess the economic benefits of interventions.

This review adds to the field by, for the first time, using Bayesian NMA to compare telerehabilitation, TLRH-AI, IPR, and UC for CNSLBP. Telerehabilitation may be particularly advantageous for patients with stable symptoms who require sustained exercise guidance, flexible scheduling, improved accessibility, and long-term self-management support. Its potential strengths lie in extending rehabilitation beyond hospital-based services, reducing travel and resource barriers, and providing scalable, low-cost intervention delivery, especially when access to in-person services is limited. In contrast, IPR may be more suitable for patients with greater functional impairment, poor movement control, higher kinesiophobia, lower adherence, or a need for direct physical assessment and real-time correction. Its advantages are mainly reflected in therapist-supervised functional recovery, individualized movement adjustment, safety monitoring, and psychological reassurance.

Therefore, telerehabilitation and IPR should not be viewed simply as competing approaches, but rather as complementary components of individualized rehabilitation pathways. Telerehabilitation may serve as a scalable strategy for longer-term pain relief, physical function maintenance, and continuous care, whereas IPR may remain essential for patients requiring intensive supervision, functional correction, or psychological support. TLRH-AI remains exploratory and should not guide routine clinical decision-making until supported by adequately powered trials. Future research should determine whether AI-assisted feedback, motion recognition, and adaptive exercise prescription can improve adherence, safety, equity, workflow integration, and long-term outcomes in real-world CNSLBP rehabilitation.