This systematic review was preregistered with the PROSPERO (International Prospective Register of Systematic Reviews; registration number: CRD420251250074) and was conducted and reported in accordance with the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) 2020 statement (Checklist 1) [22]. The literature search process was reported in accordance with the PRISMA-S (PRISMA literature search extension) [23] to ensure full transparency (Checklist 2). Evidence synthesis was conducted in accordance with the Synthesis Without Meta-analysis (SWiM) reporting guideline [24] (Checklist 3).

Compared to the preregistered protocol, the following two protocol deviations were introduced during the study implementation: (1) search update: to ensure currency of the evidence base, a supplementary search was conducted after the initial search cutoff date to capture relevant literature published thereafter, and (2) adjustment of evidence synthesis strategy: the synthesis approach was adjusted from the originally planned narrative synthesis to a structured vote-counting method based on direction of effect, with predefined subgroup and sensitivity analyses, to enhance methodological rigor. Neither of these deviations affected the achievement of the study’s primary objectives.

Eligibility criteria were predefined according to the PICOS (Population, Intervention, Comparator, Outcome, and Study design) framework.

Studies meeting all of the following criteria were included in this review: (1) P (population): patients diagnosed with any type of scoliosis, including adolescent idiopathic scoliosis (AIS), juvenile idiopathic scoliosis (JIS), adult degenerative scoliosis (ADS), and patients who have undergone posterior spinal fusion (PSF), regardless of age, sex, or curve severity; (2) I (intervention): any form of digital health intervention, including synchronous telerehabilitation, VR, AR, mobile health apps, wearable sensor systems, and digital health management platforms; (3) C (comparator): usual care, in-person rehabilitation, no active intervention, or other forms of digital intervention; (4) O (outcomes): at least one extractable quantitative outcome measure must be reported, covering radiographic parameters, quality of life, pain, respiratory function, postural parameters, assessment of surface deformity, adherence, or dropout rates; (5) S (study design): randomized controlled trials (RCTs), non-RCTs, cohort studies, and feasibility studies containing quantitative outcome data. Feasibility studies were included because they provide preliminary evidence of intervention feasibility, which is of particular value in a field where evidence remains limited.

Studies meeting any of the following criteria were excluded: (1) publications such as conference abstracts, case reports, narrative reviews, and study protocols that do not contain original quantitative data; (2) studies involving purely surgical interventions without a component of postoperative digital rehabilitation; (3) studies with insufficient raw data to support outcome extraction; and (4) non-English language literature.

To ensure comprehensive coverage of biomedical, rehabilitation, and engineering literature, we systematically searched 4 electronic databases, each searched independently through its official platform (rather than using a unified, multidatabase search platform), including PubMed (MEDLINE), IEEE Xplore, Embase, and Web of Science Core Collection. The search covered both formally published and preprint literature, with no lower date limit, from the inception of each database to November 7, 2025. Additionally, citation searching was performed on the reference lists of all included studies to identify relevant studies that may have been missed in the database search. To systematically include literature published after the initial search cutoff date, an updated search of the aforementioned 4 databases was conducted on May 12, 2026, using the same search strategy and methods.

The search strategy combined MeSH terms with free-text keywords and was adapted to each database’s search guidelines to maximize search sensitivity. Core search terms covered 3 subject clusters: digital health interventions (eg, “digital health,” “telerehabilitation,” “virtual reality,” “mHealth,” “wearable,” and “exergame”), scoliosis (eg, “scoliosis,” “adolescent idiopathic scoliosis,” and “spinal curvature”), and therapeutic interventions (eg, “exercise,” “rehabilitation,” “Schroth,” “PSSE,” and “stabilization”). Searches were restricted to English-language publications. This decision was informed by methodological evidence indicating that language restrictions have a limited effect on the qualitative conclusions of systematic reviews of conventional medical interventions [25,26]; an English-language restriction is also a common practice, applied in approximately one-third of published systematic reviews [27]. No study design filters or published methodological search filters were applied during the search phase. The search strategy was developed by the research team based on the research questions, independently of previous systematic reviews, and was subsequently reviewed by an information specialist with expertise in systematic review searches to ensure its comprehensiveness and consistency. The full search strategies for all databases, including field tags and the number of records retrieved from each database, are reported in accordance with PRISMA-S [23] and provided in Multimedia Appendix 1.

The following search methods were outside the scope of the predefined search methodology and therefore were not implemented. This review did not search gray literature (including government reports, theses, or unpublished studies); clinical trial registries were not searched; no journal websites, institutional homepages, or conference proceedings were intentionally browsed; during data extraction, corresponding authors were contacted for clarification of unclear data, but field experts and manufacturers were not contacted for unpublished data; conference abstracts were explicitly excluded in the eligibility criteria, and conference proceedings databases were not searched separately; the search strategy was not formally peer-reviewed using the Peer Review of Electronic Search Strategies (PRESS) but was independently reviewed by an information specialist at our institution.

The retrieved literature was imported into EndNote X9 (Clarivate) reference management software for automatic deduplication and manual verification. Two researchers (ZL and XW) independently screened all records based on titles and abstracts. Full-text assessment was conducted for records that passed the initial screening to confirm final eligibility, and reasons for exclusion were recorded for each record. Disagreements during the screening process were resolved through discussion and arbitrated by a third researcher (BZ). Interrater reliability for literature selection was Cohen κ=0.86. The complete literature selection process is presented in the PRISMA 2020 flow diagram.

Two researchers (ZL and QQ) independently extracted data using a predesigned and pilot-tested standardized data extraction form. Disagreements were resolved through discussion, with a third researcher (BZ) serving as the arbitrator. Interrater reliability for data collection was Cohen κ=0.88. For missing or ambiguous data, the corresponding authors were contacted via email to seek clarification; if no response was received, the data were recorded as “not available.”

The data items extracted from each included study were predefined as follows: (1) basic study information: first author, year of publication, country, and study design; (2) participant characteristics: sample size, age, and type and severity of scoliosis; (3) intervention characteristics: type of digital technology, intervention content, supervision model, and intervention dose; (4) comparator group characteristics: comparator type and intervention content; (5) cointerventions: use of braces; (6) outcome data: measurement tools for each outcome measure, assessment time points, means and variability across groups, point estimates for between-group comparisons, P values, and effect sizes; (7) implementation: adherence and dropout and loss-to-follow-up rates. When multiple measurement time points exist for the same outcome measure, data at the end of the intervention were prioritized; when multiple measurement tools are used for the same outcome domain, data from each tool were extracted independently.

Risk of bias in the included studies was assessed independently by 2 researchers (XW and CA). Disagreements were resolved through discussion, with a third researcher (MG) acting as an arbitrator when necessary. Interrater reliability for the risk of bias assessment was Cohen κ=0.88. For RCTs, the Cochrane Risk of Bias tool version 2 (RoB 2) [28] was used to assess 5 domains, including the randomization process, deviation from the intended intervention, missing outcome data, outcome measurement, and selective reporting. For nonrandomized studies, the Risk of Bias in Nonrandomized Studies of Interventions (ROBINS-I) tool [29] was used to assess 7 domains, including confounding, participant selection, classification of interventions, deviation from the intended intervention, missing data, outcome measurement, and selective reporting. Additionally, the Template for Intervention Description and Replication (TIDieR) checklist [30] was used to assess the completeness of intervention reporting across all included studies, focusing on the type of digital technology, intervention content, delivery method, supervision model, intervention frequency, duration, personalization strategies, and adherence monitoring methods.

The results of the risk-of-bias assessment were integrated into the evidence interpretation in 2 ways. First, by explicitly qualifying the conclusions of high-risk-of-bias studies in the outcome narrative and direction-of-effect synthesis; and second, by serving as the stratification basis for the prespecified sensitivity analyses.

This review did not conduct a quantitative meta-analysis for the following reasons: first, the included studies exhibited fundamental clinical differences in terms of technology type and core implementation mechanisms—including technical platforms, interaction modes, and intervention content—and forced pooling could result in unexplained mixed effects estimates. Second, the diversity of measurement tools within the same outcome domain limited direct comparability across studies. Third, for most outcomes, the number of studies eligible for pooling was too small to yield stable and reliable pooled effect estimates. Given these limitations, this review adopted an alternative synthesis strategy in accordance with the SWiM reporting guideline [24].

Following the procedures recommended by Boon and Thomson [31] and Chapter 12 of the Cochrane Handbook for Systematic Reviews of Interventions [32], evidence synthesis was performed using a vote-counting method based on the direction of effect. For each outcome, the direction of the point estimate from the between-group comparison served as the basic unit for synthesis, rather than statistical significance or effect sizes; this approach allowed studies with insufficient statistical power to be included in the synthesis based on their effect direction, rather than being excluded simply because they did not reach statistical significance [24,32]. Feasibility studies lacking a parallel control group were not included in the directional vote synthesis but were described qualitatively in a separate section to highlight their unique contributions. Effect direction coding followed 3 predefined categories, including favorable direction (↑), where point estimates indicate that the intervention group’s outcome is superior to that of the control group; unfavorable direction (↓), where point estimates indicate that the control group’s outcome is superior to that of the intervention group; and no clear direction (○), where point estimates are close to zero, suggesting that the outcomes of the 2 groups are substantially equivalent, or where the original study did not provide sufficient data to determine the direction of the between-group effect. When a single outcome domain included multiple measurement indicators, the directional coding was based on the predominant direction of each measure; if the directions of the measures were inconsistent and a predominant direction could not be determined, the coding was classified as no clear direction (○).

Based on the measurement characteristics and clinical significance of the outcomes, all outcome measures were categorized into four predefined domains: (1) clinical outcomes, reflecting the corrective effects of digital interventions on 3D spinal morphology; (2) functional outcomes, reflecting the impact of the intervention on patients’ physiological and motor functions; (3) patient-reported outcomes, reflecting patients’ subjective perceptions of health; and (4) implementation outcomes, reflecting the feasibility and sustainability of the intervention in real-world clinical settings.

For outcomes with at least 3 included studies, we conducted a prespecified 5-dimensional subgroup analysis to explore potential sources of heterogeneity in the direction of effects and their clinical significance: (1) technical pathway; (2) supervision model; (3) comparator type; (4) intervention dose; and (5) bracing cointervention status. The aforementioned subgroup analyses are exploratory in nature and aimed at identifying potential moderators of effect direction and were not designed to statistically test for moderating effects.

A sensitivity analysis was prespecified; the directional vote-counting synthesis was restricted to studies rated as low risk of bias or some concerns, excluding studies rated as high risk of bias by RoB 2 and those rated as serious risk of bias by ROBINS-I, to assess the extent to which the main synthesis conclusions depend on studies of lower methodological quality.

This review did not conduct a meta-analysis and therefore did not produce pooled effect estimates. Statistical tests for publication bias based on pooled effect estimates (such as funnel plots and the Egger test) were not applicable to the synthesis framework of this study and were therefore not performed. The potential impact of not performing these assessments on the reliability of this review’s conclusions is discussed in the “Limitations” section.

The certainty of the evidence was assessed using the Grading of Recommendations Assessment, Development, and Evaluation (GRADE) approach. Given that this review used vote counting based on direction of effect and did not generate pooled effect estimates, certainty grading was conducted according to the method provided by Murad et al [33] for such syntheses, assessing 4 domains, including risk of bias, inconsistency, indirectness, and imprecision. The assessment was performed independently by 2 researchers (CA and GW); disagreements were resolved through discussion and, when necessary, arbitrated by a third researcher (MG).

A search of 4 databases identified a total of 5562 records. After deduplication, 2644 records were screened based on titles and abstracts; 2594 irrelevant records were excluded, leaving 50 records to proceed to the full-text retrieval stage. Of these, 3 studies were excluded due to unavailability of the full text, resulting in 47 studies undergoing full-text assessment. Subsequently, we excluded 34 studies that did not meet the eligibility criteria for study design (n=14), population (n=6), or intervention (n=12), or that had inadequate data reporting (n=2). Citation searching identified 12 additional records; of these, 5 were unavailable, and 7 were excluded after full-text assessment for failing to meet the inclusion criteria. Finally, 13 studies [17,34-45] were included in this review. The characteristics of the studies excluded after full-text assessment (n=41 [database search n=34 and citation searching n=7]) are detailed in Multimedia Appendix 2. The complete literature screening process is shown in Figure 1.

The 13 included studies were published between 2015 and 2026, with study designs comprising 7 RCTs [35,37-40,43,44], 3 prospective studies [17,41,45], 1 non-RCT [36], and 2 feasibility studies [34,42]. Studies originated from 5 countries: China (n=4) [34,38,44,45], Turkey (n=4) [35,37,39,43], the United States (n=2) [36,40], Brazil (n=2) [17,41], and South Korea (n=1) [42]. The sample comprised 574 patients with various types of scoliosis: (1) AIS: 11 studies (84.6%) [17,35,36,38-45], involving 514 patients aged 9‐18 years; one study [40] included patients with AIS with concomitant low back pain, and another focused on patients after PSF [43]; (2) ADS: 1 study (7.7%) [34], involving 10 patients aged 53‐65 years; and (3) juvenile idiopathic arthritis (JIA) with scoliosis: 1 study (7.7%) [37], involving 50 patients aged 8‐16 years. Detailed characteristics of the included studies are summarized in Table 1.

This systematic review used 3 complementary assessment tools to comprehensively evaluate the methodological quality of included studies and the completeness of intervention reporting. The RoB 2 assessment (Figure 2) indicated that 3 studies (42.9%) [39,43,44] had a low risk of bias, 3 studies (42.9%) [35,37,40] had some concerns, and 1 study (14.3%) [38] had a high risk of bias. Primary methodological limitations centered on 2 domains: the randomization process (unclear reporting of allocation concealment [35,37,38,40]) and deviation from the intended intervention (difficulty implementing participant and intervention provider blinding [35,38,40]). The ROBINS-I assessment (Figure 3) revealed 2 studies (33.3%) [17,41] were at low risk, 2 studies (33.3%) [34,45] were at moderate risk, and 2 studies (33.3%) [36,42] were at serious risk. Primary sources of bias included inadequate control of confounders [34,45], selection bias [34,36,42,45], and bias in the classification of interventions [36]. The TIDieR assessments (Table S1 in Multimedia Appendix 3) showed reporting completeness ranging from 75% to 100%. All studies adequately reported basic intervention characteristics (name, theoretical basis, materials, procedures, providers, and delivery methods), but reporting on fidelity monitoring and adherence assessment was relatively limited. The synthesis of results from 3 assessment tools indicated that included studies exhibited methodological heterogeneity, with RCTs demonstrating superior overall quality to that of nonrandomized studies. Given the exploratory nature of digital interventions in scoliosis rehabilitation and limited existing evidence, this review retained all included studies to provide a comprehensive evidence base. However, interpretation of findings must fully account for the methodological limitations of each study.

In Figure 2, the upper panel shows the domain-level and overall judgment for each trial, and the lower panel shows the distribution of judgments across trials. Domains are D1, bias arising from the randomization process; D2, bias due to deviations from intended interventions; D3, bias due to missing outcome data; D4, bias in measurement of the outcome; and D5, bias in selection of the reported result. Judgments are low risk (green plus sign), some concerns (yellow minus sign), and high risk (red cross).

In Figure 3, the upper panel shows the domain-level and overall judgment for each study, and the lower panel shows the distribution of judgments across studies. Domains are D1, bias due to confounding; D2, bias due to selection of participants; D3, bias in classification of interventions; D4, bias due to deviations from intended interventions; D5, bias due to missing data; D6, bias in measurement of outcomes; and D7, bias in selection of the reported result. Judgments are low risk (green plus sign), moderate risk (yellow minus sign), and serious risk (red cross).

Detailed characteristics of the digital technologies included in the studies are presented in Table 2. Based on core technical features, the digital intervention programs included in the studies can be categorized into five primary technical pathways (Table 3): (1) synchronous telerehabilitation technology was applied in five studies [17,35-37,45], with 3 using generic instant messaging platforms (WhatsApp; Meta [17,37]; Tencent Meeting, Tencent [45]), one using a hybrid telerehabilitation model (initial in-person clinic sessions combined with subsequent remote monitoring) [35], and one using professional video conferencing platforms such as Skype (Microsoft Corp) and Zoom (Zoom Video Communications, Inc) [36]; (2) asynchronous digital educational tools were applied in 3 studies [38-40], including 2 prerecorded video studies ([40] using digital versatile discs; DVDs [39], using self-recorded videos) and 1 mobile app video library [38]; (3) immersive VR and AR technologies were applied in 4 studies [34,41-43], with 3 using VR gamified systems (Vicon system; Oxford Metrics Group, HTC Vive; HTC Corp, Kinect Xbox 360; Microsoft [34,41,43]) and 1 developing an AR smart-bar device integrated with a 9-axis inertial sensor and mobile app for real-time guidance [42]; (4) wearable biosensing technology was applied in 2 studies [38,42]. One study [38] used wrist-worn accelerometers to track high-intensity interval training (HIIT) activity levels, while another [42] used a 9-axis inertial sensor for real-time detection of postural deviations; (5) integrated digital health platforms were applied in 1 study [44], in which the authors developed an integrated digital health management platform combining patient-side apps, therapist management interfaces, and cloud-based data analysis to deliver personalized training plans based on the Peking Union Medical College Hospital Scoliosis-Specific Exercise (PUMCH-SSE) classification.

The interventions included in the studies exhibited considerable heterogeneity in frequency, duration, supervision model, and exercise methods. Regarding intervention frequency: low frequency (≤2 sessions/week) was used in 3 studies [17,37,43]; moderate frequency (3‐5 sessions/week) in 2 studies [34,40]; high frequency (6‐7 sessions/week) in 6 studies [35,36,38,39,44,45]; 1 study [41] used an intermittent frequency (≥7 day intervals, totaling 3‐4 sessions); and 1 study [42] did not report frequency. Regarding intervention duration, 6 studies [17,36-38,44,45] delivered long-term interventions (total duration >12 weeks), 6 [34,35,39-41,43] delivered short-term interventions (≤12 weeks), and Nam et al [42] did not report a standardized duration.

Based on the mode and intensity of therapist involvement during the intervention, supervision models can be classified into three categories: (1) real-time synchronous therapist supervision (6 studies [17,34,35,41,43,45]): during each training session, therapists actively observed and corrected movement execution through in-person guidance [34,41,43] or real-time video links [17,35,45]; (2) asynchronous technology-mediated remote monitoring (3 studies [37,38,44]): patients primarily engaged in independent home-based practice, with delayed or on-demand therapist feedback facilitated through instant messaging [37], mobile apps, and wearable device tracking [38], or digital health management platforms [44]; no synchronous supervision occurred during training; (3) minimally supervised self-directed practice (4 studies [36,39,40,42]): following initial instruction, training was primarily self-directed at home, with professional contact limited to occasional remote guidance [36], video review combined with periodic follow-ups [39], or the provision of standardized video instructional materials only [40]; there was no continuous real-time or technology-mediated monitoring.

Among the 11 studies with parallel control groups, the control groups can be classified into three categories based on increasing intervention intensity: (1) inactive comparator (2 studies [36,38]): The control group received usual care or no active intervention; (2) unsupervised home exercise (3 studies [35,43,45]): the control group performed home exercises without real-time supervision by a professional; (3) in-person supervised rehabilitation (6 studies [17,37,39-41,44]): the control group received traditional rehabilitation training under the in-person guidance of a professional therapist. Two feasibility studies [34,42] were not included in the above classification because they lacked parallel control groups.

Exercise methods exhibited diverse characteristics. Four studies [37,39,44,45] applied 3D corrective training based on the PSSE method, emphasizing self-corrective posture, rotational breathing, and spinal stability training. One study used a modified Pilates protocol [35], focusing on core muscle strengthening and breathing exercises. Another study used yoga isometric posture training [36], incorporating poses such as side plank and half moon. Four studies [34,41-43] integrated motion-sensing games or VR (AR) systems, combining postural control with movement feedback. One study used HIIT training [38], incorporating 40 high-impact weight-bearing exercises. Additionally, 3 studies combined orthotic treatment; all patients in Andrade et al [17] wore thoracolumbosacral orthoses for 18‐20 hours/day; 50% of participants in Dursun et al [35] used orthoses; and all participants in Chen et al [45] received an individualized computer-aided design and computer-aided manufacturing (CAD-CAM)–fabricated Chêneau brace, worn ≥22 hours/day.

This section presents stratified reports according to the 4 outcome domains defined in the methodology (clinical outcomes, functional outcomes, patient-reported outcomes, and implementation outcomes). For studies with a parallel control group, results were synthesized using a vote-counting method based on the direction of effect; 2 feasibility studies were excluded from the direction-of-effect synthesis due to the absence of a parallel control group, and their outcomes were described qualitatively. The complete extracted outcome data for each included study are presented in Multimedia Appendix 4.

A prespecified sensitivity analysis restricted the direction-of-effect synthesis to the 9 studies not rated as high risk of bias (RoB 2) or serious risk of bias (ROBINS-I), excluding Fishman [36] and Lau et al [38], to assess the extent to which the main synthesis conclusions relied on studies of lower methodological quality (Table S7 in Multimedia Appendix 5). The favorable-direction signal for the Cobb angle was strengthened after exclusion (71.4% → 80.0%), and the study with an unfavorable direction (Andrade et al [17]) was retained, suggesting that this signal does not depend on studies at high or serious risk of bias and is therefore robust. The proportion of studies showing a favorable direction for health-related quality of life decreased (75.0% → 66.7%), but the 2 retained studies [43,45] still showed a favorable direction, and the signal persisted overall. The proportion of studies showing a favorable direction for treatment adherence increased (37.5% → 50.0%), and the unfavorable signal from Zapata et al [40] persisted, confirming robustness. For the remaining 7 outcomes, the direction-of-effect proportions were unchanged in the sensitivity analysis.

Using the GRADE approach, we assessed the certainty of evidence for each of the 10 outcomes included in the direction-of-effect synthesis (Table S8 in Multimedia Appendix 5): only the Cobb angle had moderate certainty; spinal flexibility, respiratory function, postural control and gait, and health-related quality of life had low certainty; and ATR, pain, body image, treatment adherence, and dropout and loss-to-follow-up rates had very low certainty. The certainty of the available evidence was generally low, with downgrades primarily attributed to imprecision and risk of bias, followed by inconsistency in effect directions; indirectness affected only a few outcomes.

Wan et al [34] developed an optical sensor-based (Vicon system) motion-sensing video game system for ADS (n=10), in which patients controlled game characters by adjusting their trunk posture in real time. After a 6-week intervention (3 sessions/week), immediate improvements in participants’ sagittal plane posture were recorded following each training session. After 6 weeks, improvements were observed in shoulder and pelvic tilt angles, and postural awareness was enhanced; however, no significant improvement in muscle symmetry was observed. Nam et al [42] developed an augmented reality smart-bar device integrated with a 9-axis sensor for AIS (n=13), which, in conjunction with a mobile app, provided real-time AR posture framework overlays and audio error feedback. Compared with the no-AR condition, AR guidance reduced movement deviation angles, postural sway was suppressed to some extent, and movement accuracy was improved.

None of the 13 included studies reported any serious adverse events. Mild discomfort was reported in 2 studies. In Moraes et al [41], 2 participants experienced transient dizziness after prolonged VR headset use, which resolved spontaneously after rest; in Wan et al [34], one participant experienced mild muscle soreness after training, which resolved after rest or adjustment of exercise intensity.

To our knowledge, this systematic review represents the first structured synthesis of evidence on the application of digital health technologies in scoliosis rehabilitation using a technology classification framework. A total of 13 studies (574 patients) were included, and the review followed the SWiM reporting guideline [24], synthesizing evidence across 4 outcome domains using a vote-counting method based on the direction of effect [31,32].

In the clinical outcomes domain, digital interventions generally showed a favorable direction for the Cobb angle and spinal flexibility, whereas the evidence for ATR correction was weaker, with most studies showing no clear direction. This pattern suggests that improvements in overall spinal alignment and flexibility are more reliable, whereas the effect of segmental rotational deformities remains to be confirmed. In the functional outcomes domain, the favorable direction was highly consistent for respiratory function and postural control, reflecting the potential value of digital interventions in the functional dimension. Regarding patient-reported outcomes, health-related quality of life showed a relatively consistent favorable direction, whereas the evidence for pain and body image was inconsistent, indicating that the impact of digital interventions on subjective experiences remains variable. In the implementation outcomes domain, the effect directions were the most dispersed; the proportion of studies showing a favorable direction for treatment adherence was relatively low, and dropout and loss-to-follow-up rates likewise showed no consistent advantage. In summary, the potential benefits of digital interventions for spinal morphology and function are relatively clear, whereas their sustained and standardized use (particularly as represented by adherence) may be a key factor influencing whether clinical benefits can be realized. It should be noted that the included studies exhibit substantial clinical heterogeneity, and the GRADE assessment indicated low certainty of the overall evidence. Therefore, the synthesized findings for each outcome should be interpreted cautiously as preliminary and exploratory evidence.

Subgroup analyses revealed several interpretable patterns. First, the favorable direction for the Cobb angle was most evident when digital interventions were compared with unsupervised home exercise, whereas this advantage diminished when they were compared with in-person supervised rehabilitation. This suggests that their added value lies primarily in serving as a substitute for or supplement to unsupervised training, rather than surpassing standardized in-person rehabilitation. Second, improvements in ATR were more commonly observed in studies with longer intervention durations (>12 weeks), suggesting that an adequate intervention period may be a critical factor in correcting rotational deformities. Third, the favorable direction for health-related quality of life and treatment adherence diminished as supervision intensity decreased, suggesting that the intensity of professional supervision may be a common factor influencing both long-term adherence and subjective benefits. Sensitivity analyses further indicated that these key findings remained generally robust after excluding studies at high or serious risk of bias. Taken together, the directional evidence and subgroup patterns indicate that the principal barrier to the clinical translation of digital interventions may lie less in efficacy itself than in patients’ ability to sustain use, with adherence being a key manifestation. This is consistent with a perspective increasingly emphasized in the digital rehabilitation literature: the real-world effectiveness of a technology depends largely on its implementation pathway and on sustained user engagement [47,48].

It should be noted that the mechanisms discussed in this section are interpretations based on existing theories rather than conclusions directly confirmed by research; since most included studies did not measure neurophysiological or behavioral mediators, the following discussion should be understood as a theoretical explanation rather than an empirical finding.

A key finding of this review is that, in digital scoliosis rehabilitation, whether an intervention can be sustained and produce clinical benefits in the real world may depend primarily on the level of human supervision embedded in its delivery, rather than on the technology category itself. A cross-comparison of implementation outcomes across the 5 technology categories revealed that the proportion of studies showing a favorable direction for treatment adherence and dropout (loss-to-follow-up rates) did not increase with the level of technological sophistication (Table S2 in Multimedia Appendix 5). Notably, even within the same technology category (synchronous telerehabilitation), implementation outcomes varied depending on the supervision model. Together, these observations suggest that implementation success depends less on the technology category than on the degree of professional supervision. This is consistent with recent digital-health evidence that adherence is multifactorial, with support from health care professionals identified as a key determinant [88]. The role of supervision can be understood at 2 levels. First, supervision directly influences treatment efficacy by ensuring the quality of exercise execution [40,89]. Second, supervision sustains long-term patient engagement by providing accountability and feedback—a function particularly important in the conservative management of scoliosis, which relies on sustained training [90,91]. Importantly, this reliance is not absolute: one study reported that a home-based program relying primarily on monthly supervision still achieved outcomes comparable to clinic-supervised training for most measures [39].

Overall, supervision intensity, as a key determinant of implementation effectiveness [92,93], serves as an important complement to the technology classification framework proposed in this review. While the technology classification defines the functional capabilities an intervention can provide, whether these capabilities translate into sustained clinical benefits may depend substantially on the corresponding supervision framework. Previous evidence indicates that adequate supervision in the early stages helps foster long-term home adherence [94] and that remote supervision under professional guidance can achieve outcomes comparable to in-person rehabilitation [56]. Accordingly, this review suggests that a “hybrid model combined with phased supervision intensity” may represent a promising strategy for balancing efficacy, adherence, and generalizability, namely, transitioning from early intensive in-person guidance [35] to high-frequency, digitally supported remote supervision during the maintenance phase [44], and sustaining efficacy through low-frequency reassessment during long-term follow-up [39]. However, due to the limited number of included studies across different technology categories and subgroups, as well as the lack of head-to-head comparisons, the aforementioned strategy remains an exploratory hypothesis at this stage. The optimal supervision intensity and timing of transition require validation through prospective studies.

This review has several limitations. (1) The available evidence is limited in both quantity and quality: research in this field remains scarce, with only 13 studies included; among the 7 RCTs, only 42.9% (n=3) were at low risk of bias, while 33.3% (n=2) of the 6 nonrandomized studies had a serious risk of bias, and sample sizes were generally small; moreover, reporting on intervention fidelity and adherence monitoring was insufficient, and acceptability and feasibility were formally assessed with validated instruments in only one study [17], precluding a quantitative comparison of these implementation outcomes across technologies. (2) Follow-up was predominantly short- to medium-term, with only one study reaching 24 months [45]; given the progressive nature of scoliosis and the need for sustained conservative management until skeletal maturity, whether the observed improvements persist beyond the active treatment period remains uncertain. (3) As no meta-analysis was performed, the review provides no pooled effect estimates, so the magnitude and precision of effects could not be quantified, and statistical tests for publication bias (eg, funnel plots and Egger test) were not applicable; reporting and publication biases therefore cannot be excluded. Correspondingly, the GRADE-rated certainty of the evidence was generally low. (4) Generalizability is constrained: the included studies originated predominantly from middle- and high-income countries, with a lack of data from low-income nations, and the restriction to English-language publications, with 4 of the 13 studies originating from Chinese institutions, may have systematically excluded non-English, particularly East Asian, literature, thereby underrepresenting the full breadth of current evidence.

Future research should conduct large-sample, long-term follow-up multicenter RCTs to validate the long-term efficacy and safety of digital interventions for scoliosis. Additionally, optimal implementation strategies based on patient characteristics should be explored to achieve personalized, precise interventions. Further health economic evaluations are needed to compare the cost-effectiveness of different digital technologies versus traditional rehabilitation. Future reviews should also incorporate multilingual search strategies to achieve more geographically and linguistically representative evidence synthesis. As new technologies emerge, digital interventions in scoliosis management should evolve alongside them, continuously integrating these advances to further enhance efficacy and patient experience.

This review systematically evaluated the application of digital health technologies in scoliosis rehabilitation: across the 13 included studies, a vote-counting synthesis based on the direction of effect indicated a favorable direction for digital interventions in spinal alignment and flexibility, respiratory function, postural control and gait, and health-related quality of life; however, the certainty of the evidence was low. Its innovation lies in being the first to incorporate diverse digital interventions into a technology classification framework comprising 5 technology categories and 3 maturity tiers, integrating cross-technology evidence of effectiveness. Based on this, the review suggests that whether these benefits can be realized in the real world may depend not only on the technology itself but also on the supervision models that sustain patient adherence. It is precisely this cross-technology integration that distinguishes this review from previous similar reviews, which often focused on a single technology or a single function. Consequently, this review not only provides a structured and actionable basis for clinical technology selection but also shifts the focus from “which technology is most effective” to “how to ensure sustained use of the technology in practice.” At the clinical level, this implies that technologies may be selected according to their maturity tier and paired with appropriate supervision strategies (such as a “hybrid model combined with phased supervision intensity”) to ensure adherence. This approach could expand access to high-quality rehabilitation resources in underdeveloped and remote areas without compromising therapeutic efficacy. Adequately powered studies with sufficient follow-up are still needed to confirm the value of digital interventions in precision rehabilitation for scoliosis.