The effect of self‐management techniques on relevant outcomes in chronic low back pain: A systematic review and meta‐analysis

Among many treatment approaches for chronic low back pain (CLBP), self‐management techniques are becoming increasingly important. The aim of this paper was to (a) provide an overview of existing digital self‐help interventions for CLBP and (b) examine the effect of these interventions in reducing pain intensity, pain catastrophizing and pain disability.


| INTRODUCTION
Low back pain is a global health epidemic, which affects about one-third of the global population (Chen et al., 2022), and it is the leading cause of years lived with disability according to the Global Burden of Disease Study (Wu et al., 2020).Disability due to low back pain vastly affects individuals' quality of life including physical, psychological and social functioning, especially when pain has become chronic (Breivik et al., 2013).Besides the personal impact, low back pain's socioeconomic burden is huge.For example, in western countries, the financial burden of back pain is estimated at 1%-2% of the gross national product (Hansson & Hansson, 2005;Wenig et al., 2009).Eighty to ninety per cent of these costs are indirect costs, caused by loss of productivity and disability (Hartvigsen et al., 2018).Accordingly, there is an urgent need to improve the current therapy for patients with chronic low back pain (CLBP) and raise awareness for this global health issue.
A large number of treatment approaches exist for patients with CLBP, and many treatments have proven to be effective in clinical trials (Qaseem et al., 2017).Common approaches include, for example, lifestyle modifications, diet, education, physical therapy, pharmacologic treatments, psychotherapy and in some cases surgery (Foster et al., 2018).According to the American College of Physicians, patients are encouraged to try non-pharmacologic interventions such as self-management techniques as a first choice to treat their CLBP because these therapies are less harmful than pharmacologic options (Qaseem et al., 2017).Selfmanagement refers to the individual's ability to manage the symptoms, treatment, physical and psychological consequences, and lifestyle changes associated with a chronic condition (Barlow et al., 2002).
According to a systematic review and meta-analysis (Du et al., 2017), self-management is effective for CLBP.Specifically, the authors found that self-management programmes had a moderate effect in relieving pain and a small-to-moderate effect in improving disability.However, factors such as distance, time, cost (Borrelli & Ritterband, 2015) and lack of information impede access to self-management programs for many patients (Gliddon et al., 2017).
Since the beginning of the digital age, new opportunities have arisen to establish new technologies for disease management (Borrelli & Ritterband, 2015).E-health describes a healthcare practice supported by information technologies to prevent disease, treat disease, and promote and maintain health status (Borrelli & Ritterband, 2015).If health information is provided via mobile devices, this is also known as m-health (Gliddon et al., 2017).There is the first evidence that e-health-based self-management programs are effective in improving pain intensity and disability in CLBP patients.In a meta-analytical review, the authors found small effects for the improvement of pain intensity and disability immediately after the treatment as well as for short-term outcomes (Foster et al., 2018).In this systematic review, we aim to address the following research questions: (1) What types of digital self-management interventions for adults with CLBP have been studied thus far? (2) What is the effectiveness of these interventions in reducing pain intensity, pain interference/pain disability, 1 and pain catastrophizing?(3) What characteristics define effective interventions?Answering these questions is crucial in developing digital interventions that can be tailored to individual patients with CLBP.

| METHODS
The systematic review and meta-analysis were reported based on the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines (Page et al., 2021).

| Data sources and search strategy
The search for literature was performed in the databases Cochrane, Embase, Medline, CINAHL & PsychInfo on 23 March 2022.We additionally looked in the International Clinical Trials Registry (ICTRP) from the World Health Organization, to find studies not yet published.The search was limited to the last 10 years.
The search strategy consisted of keywords and Medical Subject Headings (MeSH).In addition, the Cochrane Highly Sensitive Search Strategy for randomized controlled trials (RCTs) was used for our search (Cumpston et al., 2019).As described in the protocol, using Web of Science, citation tracking was performed with the final selection of studies.The search terms are provided in Table 1.

| Eligibility criteria
Studies meeting the following criteria were eligible for inclusion: 1. Patients reporting nonspecific CLBP, lasting ≥3 months 2. Age ≥18 years 3. Using a digital intervention (accessible via smartphone/ smartwatch/tablet/computer or internet browser) that aims for reducing pain, disability, catastrophizing or pain-related impact on functioning in patients with CLBP 4. Using a self-management technique guided or unguided including psychological and behavioural therapy approaches, patient education, physical activity, physiotherapy, complementary and alternative treatments, interdisciplinary methods and multimodal programs 5. Comparison of intervention with a control group (e.g., waitlist, active control or treatment as usual) 6. Reported outcome at least pain intensity 7. Using a RCT design 8. Study was available as full text in English or German language Two independent reviewers (C.S. and P.S.) screened all search results by title and abstract.For those studies that were not excluded, the full text was obtained and screened.If there was disagreement between the two reviewers after full-text screening, a third independent reviewer (A.O.) was consulted.The same process was performed for the studies found for citation tracking via Web of Science.

| Data extraction and management
Two authors (P.S. and C.S.) extracted the data separately using a structured table.After extracting all the data, both authors compared the extracted information and resolved inconsistencies.Extracted data included: 1. title, DOI, short description of the study 2. outcome and sample type, outcome measure, instrument type 3. the type of intervention, type of control group 4. mean and standard deviation of pain intensity before and after treatment 5. mean and standard deviation of pain catastrophizing before and after treatment 6. mean and standard deviation of pain disability before and after treatment 7. follow-up data of pain intensity, pain catastrophizing and pain disability 8. moderators such as format, country, age, percentage of women, duration of treatment and what kind of device supported the intervention (device support) If the article did not provide data as mean and standard deviation, the given format (e.g., change score and standard error) was extracted and later converted into mean and standard deviation.
In Özden et al. (2022), pain intensity was only described separately for activity and rest periods.Therefore, we chose to use the mean of both values.In the threearmed trial by Irvine et al. (2015), both control groups  T A B L E 1 Search terms for systematic review and meta-analysis.
were combined.Due to the low number of studies that reported follow-up data (only four studies), we decided to refrain from reporting any follow-up data.

| Risk of bias
The RoB was assessed by two authors individually (P.S. and C.S.).In the case of inconsistent ratings, the domain was rated by a third author (C.Y.P.).We used the RoB tool for randomized trials version 2 (RoB2; Sterne et al., 2019) for the assessment.The tool contains five domains, which can be rated separately as high risk, some concerns or low risk.
• Domain 1: RoB arising from the randomization process • Domain 2: RoB due to deviations from the intended interventions (effect of assignment to intervention) • Domain 3: RoB due to missing outcome data • Domain 4: RoB in measurement of the outcome • Domain 5: RoB in selection of the reported result After rating the domains individually, the overall scores were calculated following the RoB2 algorithms.

| Quality of evidence
To assess the quality of evidence of each effect evaluated in this meta-analysis, we used the Grading of Recommendations Assessment, Development and Evaluation (GRADE) method.
With GRADE, quality of evidence is a evaluated according to the criteria defined by the GRADE Working Group for reviews in the healthcare sector, with the goal of better comparability (Schünemann et al., 2013).
Evaluation criteria include the following dimensions: RoB, inconsistency, indirectness, imprecision and publication bias.The extent of the limitations in each dimension can be rated as 'not serious', 'serious' and 'very serious'.Based on another similar study (Schütze et al., 2018), dimensions were rated based on the following criteria: 1. Risk of bias: less than 25% of the study participants are from studies rated as 'high/unclear RoB' by the RoB2 score.2. Inconsistency: significant heterogeneity in pooled effect (I 2 > 50%) 3. Indirectness: the interventions are not directly compared; it will not be possible to generalize from the results 4. Imprecision: <400 study participants 5. Publication bias: the funnel plot shows asymmetry and the funding of a large part of the studies to be evaluated is not presented transparently or comes from profitoriented funding.

| Meta-analysis
We conducted separate meta-analyses for the outcomes pain intensity, pain catastrophizing, pain disability and one hierarchical meta-analysis for all outcomes combined to reflect the efficacy of the included interventions across all relevant outcomes.We calculated standardized mean differences with a small-study correction (Hedges' g) for post-intervention comparisons between interventions and control conditions.If only data on pre-post changes were available, we transformed those effect sizes to Hedges' g.For each individual domain, we pooled the effect sizes using random-effects models.The amount of heterogeneity (i.e., τ 2 ) was evaluated using the restricted maximum-likelihood estimator (Viechtbauer, 2005).
In addition to the estimate of τ 2 , the Q-test for heterogeneity (Cochran, 1954) and the I 2 statistic (Higgins & Thompson, 2002) were reported.In case, any amount of heterogeneity was detected (i.e., �  2 > 0, regardless of the results of the Q-test), a prediction interval for the true outcomes was implemented as well (Riley et al., 2011).
Tests and confidence intervals were computed using the Knapp and Hartung method (Knapp & Hartung, 2003).
To examine whether studies may be outliers studentized residuals and Cook's distances are used (Viechtbauer & Cheung, 2010).Studies with a studentized residual larger than the 100 × (1 -0.05/(2 × k)) percentile of a standard normal distribution are considered potential outliers (i.e., using a Bonferroni correction with two-sided α = 0.05 for k studies included in the meta-analysis).Studies with a Cook's distance larger than the median plus six times the interquartile range of the Cook's distances considered to be influential.Additionally, we fitted a three-level meta-analysis, nesting the effect sizes of all three outcomes within the included studies to address effect size dependency.We constructed the confidence intervals with both the conventional three-level model (Assink & Wibbelink, 2016) and robust sampling variance estimation (Hedges et al., 2010).

| Publication bias
The validity of meta-analyses can be compromised by publication bias.Publication bias occurs if negative study results (e.g., RCT without a response of the treatment group) are not published.We use recently suggested methods to control publication bias (Bartoš et al., 2022).We performed precision-effect tests and precision-effect estimates with standard errors (PET-PEESE), selection models and robust Bayesian meta-analysis.In addition, the contour-enhanced funnel plots were inspected.To examine funnel plot asymmetry, the rank correlation test (Begg & Mazumdar, 1994) and the regression test (Sterne & Egger, 2005) were used.

| Subgroup analysis
In the case of sufficient statistical power, we performed subgroup analyses based on the type of control condition, type of intervention and gender.

| Disclosure, preregistration and open science
All sample sizes, data exclusions, manipulations and measurements occurring in the analysis were disclosed.The protocol for this meta-analysis containing the search strategy, eligible criteria following the PICO scheme and other information on data extraction, RoB assessment, data synthesis and subgroup analysis was published before the database search on the PROSPERO platform (CRD42022319992).All data necessary for the reproducibility of the analysis will be made available via the Open Science Framework (https:// osf.io/ vks95/ ?view_ only= 07c1f da6c3 a44c6 d8c4d 9e16d d652a10).The data meet the criteria of the FAIR guidelines for scientific data.

| Search results
The search yielded 1572 results, of which 678 were duplicates.Therefore, titles and abstracts of the remaining 894 records were screened.Of these, 25 were found to be eligible for full-text screening.Eight were assessed as suitable for inclusion in the full-text screening.With these eight studies, we performed citation tracking via Web of Science.Citation tracking identified 10 additional studies, five of which were included after screening the full text.
Seven of 15 full texts were excluded due to inclusion of participants with other diagnoses than CLBP, such as neck pain or fibromyalgia (Baumeister et al., 2021;Darnall et al., 2020;Koppenaal et al., 2022;Malfliet et al., 2018;Sander et al., 2020;Selter et al., 2018;Weymann et al., 2015).Three studies had to be excluded because patients' back pain lasted less than 3 months at the time of recruitment (del Pozo-Cruz et al., 2012;Shebib et al., 2019;Toelle et al., 2019).Two studies did not have an intervention that was regarded as sufficiently digital (face-to-face teaching program taught by a physical therapy course professor, interactive voice programs via telephone; Bianchi et al., 2014;Heapy et al., 2017).One study did not use a randomized controlled study design (Valenzuela-Pascual et al., 2015).Another study had to be excluded because of a measurement scale that was not suitable (Sandal et al., 2021).One study (Riva et al., 2014) would have been suitable for inclusion, because the data provided in the paper was sufficient to calculate the effect sizes.Unfortunately, we were not able to obtain the necessary data from the authors of the study.
The PRISMA flow chart is provided in Figure 1.

| Description of studies
Table 2 shows a summary of the 12 included studies.
The studies included in the meta-analyses were published between 2012 and 2022.Of the 12 included studies seven were conducted in the United States (Amorim et al., 2019;Barone Gibbs et al., 2018;Carpenter et al., 2012;Garcia et al., 2021;Irvine et al., 2015;Krein et al., 2013;Zadro et al., 2019) one was conducted in India (Chhabra et al., 2018), one in Turkey (Özden et al., 2022) and three in China (Yang et al., 2019;Zheng, Liu, et al., 2022;Zheng, Zheng, et al., 2022).All studies, except one, followed a two-armed design with one control group.Irvine et al. compared their intervention with both an active (emails with informative content) and an inactive (treatment as usual) control group.Most studies chose pain intensity and/or pain disability as their primary outcome.Three studies (Garcia et al., 2021;Irvine et al., 2015;Özden et al., 2022) did not define a specific primary outcome.Various secondary outcomes were collected in the studies, most frequently physical activity and self-efficacy.One study assessed pain interference (Garcia et al., 2021) Chhabra et al., 2018).Four studies assessed pain catastrophizing as an outcome (Carpenter et al., 2012;Garcia et al., 2021;Irvine et al., 2015;Zheng, Liu, et al., 2022).
Seven of 12 studies in this meta-analysis examined the efficacy of a self-management intervention that was developed based on cognitive behavioural therapy (CBT) in combination with exercise.Three studies designed a merely exercise-based intervention, and two only provided CBT content to patients in the intervention group.Seven studies used inactive control groups; this included wait-list control or treatment as usual.As described above, Irvine et al. (2015) conducted a threearmed study with active and inactive control groups.The remaining four studies established an active control group, each modifying the actual intervention and making it less detailed and patient-centred.The digital formats were limited to web-based, app-based interventions and virtual programs, such as gaming and virtual reality.The average duration of the interventions was 14 weeks (standard deviation = 14.1).The shortest intervention duration (3 weeks) was used in the study by Carpenter et al. (2012), whereas Krein et al. (2013) used the intervention with the longest duration, which was 48 weeks.

| Participants
There were 1545 participants studied in total, with trial samples ranging from 8 to 597 people.Two studies were not included in determining the proportion of women because the gender distribution was not stated in the publication (Chhabra et al., 2018) or no distinction was made between the intervention and control group (Carpenter et al., 2012).Therefore, only 1295 participants could be considered for the calculation of the proportion of women.In addition, three studies disclosed only the sex, but not the gender, of the study participants (Carpenter et al., 2012;Zadro et al., 2019;Zheng, Zheng, et al., 2022).The mean age across all study participants was 48.3, the range was 34.5-68.3.Mean age was reported in all studies except for Carpenter et al. (2012), which did not differentiate the age of intervention and control groups, and Irvine et al. (2015), which did not report age at all.Baseline pain intensity scores were available for 11 studies.The mean baseline score was 3.7 (Standard deviation = 2.1) on a 11-point scale (0 = no pain, 10 = maximum pain).Demographic findings of the study population are summarized in Table 3.  Abbreviations: ac, active control; CBT, cognitive behavioural therapy; NRS, numeric rating scale; sham vr, sham virtual reality; tau, treatment as usual; VAS, visual analogue scale; wlc, wait-list control.

| Risk of bias
The RoB assessment shows that the studies included have mixed 'RoB' with a tendency towards 'high RoB'.The results of the quality assessment for RoB through RoB2 are provided in Figure 2. Four of the studies included show serious RoB.Two studies (Carpenter et al., 2012;Özden et al., 2022) showed bias in domain 2 (Bias due to deviations from intended intervention) and two studies (Zadro et al., 2019;Zheng, Liu, et al., 2022) showed bias in domain 5 (bias in selection of reported results).The poor rating of domain 2 occurred mainly because Carpenter et al. (2012) and Özden et al. (2022) analysis did only include cases that completed the intervention, and the attrition rates of these two studies were greater than 5%.The high-risk scores in domain 5 occurred in both Zadro et al. (2019) and Zheng, Liu, et al. (2022) because there were deviations from the protocol.In both studies outcomes in the final study were modified after creating the protocol.
Most of the studies (Amorim et al., 2019;(Chhabra et al., 2018;Garcia et al., 2021;Krein et al., 2013;Yang et al., 2019) show some concerns in domain 3 (bias due to missing outcome data).Domain 3 was rated with 'some concerns' because one could only suspect that the missing data were depending on the true value.We did not find any RoB for the studies (Barone Gibbs et al., 2018;Irvine et al., 2015;Zheng, Zheng, et al., 2022).

| Quality of evidence
Because our study did not synthesize results from pharmacological trials, only the quality of evidence assessment and not the strength of recommendation of the GRADE rating was meaningful for this meta-analysis.The quality of evidence of the outcome pain intensity and of the combined outcome was rated as high.The quality of evidence of the outcomes pain disability and pain catastrophizing was rated as moderate because one category in each certainty assessment was rated as serious.In the case of the outcome pain catastrophizing, the inconsistency was rated as serious because I 2 was 78% and thus well exceeded the cut-off >50%.For the outcome pain disability, the RoB category was as serious because 31% (cut-off >25%) of the participants originated from studies that were classified as having high RoB.The GRADE rating results are provided in Table 4.
3.6 | Meta-analysis 3.6.1 | Pain intensity A total of k = 12 studies were included in the analysis.The results of the meta-analysis indicate a significant difference between the intervention and control groups, with a pooled summary effect size of Hedges' g = 0.24, 95% CI [0.09, 0.40], indicating a small effect in favour of the intervention group (p = 0.0052).The observed standardized mean differences ranged from −0.07 to 0.88, with the majority of estimates being positive (83%).Figure 3 shows the forest plot for the studies observing the pain intensity.
According to the Q-test, the heterogeneity of the true pain intensity outcomes was not significant, Q(11) = 15.05,p = 0.1800, ̂ 2 = 0.02, I 2 = 33%.The 95% prediction interval for the true outcomes ranges from −0.11 to 0.59.This means that although the average effect was estimated to be positive, the true outcome may actually be negative in some studies.
An examination of the studentized residuals revealed that none of the studies had a value larger than ±2.87, which means that there was no evidence of outliers in this model.According to Cook's intervals, one study (Carpenter et al., 2012) may be overly influential.

| Pain catastrophizing
The results of the meta-analysis on pain catastrophizing based on four studies found a statistically nonsignificant difference between the intervention and control groups, with a pooled effect size of Hedges' g = 0.38, 95% CI [−0.31, 1.06], p = 0.1790.Observed results ranged from −0.00 to 1.04, with three out of four estimates being positive (75%).Figure 4 shows a forest plot with the observed results and the estimate based on the random-effects model.
According to the Q-test, the true outcomes seem to be heterogeneous [Q(3) = 15.89,p = 0.0012, ̂ 2 = 0.13 , I 2 = 84%].The 95% prediction interval for the true outcomes lies between −1.43 and 2.18.Thus, although the average outcome was estimated to be positive, the true outcome may actually be negative in some studies.
An inspection of the student residuals revealed that none of the studies had a value greater than ±2.5, which means that there was no indication of outliers in this Outcome According to Cook's distances, none of the studies could be considered overly influential.

| Pain disability
The meta-analysis including 11 studies on pain disability found a significant difference between the intervention and control groups, with a pooled summary effect size of Hedges' g = 0.43 95% CI [0.27, 0.59], indicating a smallto-medium effect in favour of the intervention group (p = 0.0001).Observed standardized mean differences ranged from −0.26 to 1.28.Most of the results were positive, with 91% of the estimates having a positive outcome.Figure 5 shows the forest plot for the studies observing the pain disability.
One study (Garcia et al., 2021) had a relatively large weight compared with the other studies (i.e., a weight at least three times as large as an equal weight for all studies).An examination of the student residuals revealed that none of the studies had a value greater than ±2.8376.Thus, there was no evidence of outliers in this model.According to Cook's intervals, none of the studies could be considered overly influential.

| Combined outcomes
The results of a three-level meta-analysis, nesting all 27 effect sizes within the 12 included studies for the combination of all three pain outcomes (Figure 6), found a significant difference between the intervention and control groups, with a pooled effect size of Hedges' g = 0.33, 95% CI [0.17, 0.49], indicating a small-to-moderate effect in favour of the intervention group (p < 0.001).The robust variance estimation yielded similar results with g = 0.31, 95% CI [0.14,0.49].The observed outcomes for all k = 27 effect sizes ranged from −0.26 to 1.28, with the majority of estimates being positive (85%).
According to the Q-test, the true outcomes appear to be heterogeneous [Q(26) = 52.47,p = 0.0016, ̂ 2 between = 0.042 , ̂ 2 within = 0.0455, I 2 = 50%].The 95% prediction interval for the true outcomes ranged from −0.14 to 0.80, indicating that although the average effect was estimated to be positive, that the true outcome may be negative in some studies.
An examination of the studentized residuals revealed that none of the studies had a value larger than ±3.11, which means that there was no evidence of outliers in this model.According to Cook's distances, none of the studies could be overly influential.

| Funnel plot
Neither the rank correlation nor the regression test indicated asymmetry of any of the funnel plots for the individual pain domains (Figure 7a-d; pain intensity, p = 0.2496 and p = 0.4513; pain catastrophizing, p = 0.3333 and p = 0.0581; pain disability: p = 0.3587 and p = 0.6237; combined outcomes: p = 0.2132 and p = 0.0593).However, visual inspection of the funnel plots and trim-and-fill analyses for the combined outcomes (Figure 7d) might suggest small-study effects, as small studies were present for positive effects, yet not for negative effects, which were imputed with the trimand-fill method.

| Subgroup analyses
All planned subgroup analyses did not reach sufficient power and are therefore reported in the Supporting Information only (Tables S1-S6).Due to the insufficient power of the subgroup analyses, unfortunately, we were not able to answer the third review question: 'What characteristics define effective interventions?'

| Deviations from the protocol
In the protocol, we did use the term 'pain interference' when we were referring to outcomes that reflect the impact of pain on one's functioning in different areas.However, during the literature search, we recognized that most papers used disability' outcomes, which is a highly similar construct or which can even be regarded as the same construct (Wilson, 2014).Therefore, we chose to combine pain disability and pain interference outcomes for the purposes of this analysis.In addition to the analysis described in the protocol, we also conducted a combined analysis including all three outcomes, which may reflect a better estimate of an 'overall' effect of an intervention on the well-being of patients.Because there was not enough follow-up data available, we did not perform a meta-analysis for this time point.In addition, we did not include subgroup analyses in the paper, due to the lack of statistical significance.

| DISCUSSION
We included 12 studies that investigated the efficacy of various digital self-management interventions in 1545 adults with CLBP.The meta-analyses were based on 27 effect sizes for the outcomes pain intensity, pain catastrophizing and pain disability.We found small-tomedium summary effect sizes for reducing pain intensity and pain catastrophizing, as well as improving pain disability, indicating preliminary evidence for the efficacy of digital self-management interventions in adults with CLBP.
When we analysed all effect sizes (k = 27) for all pain domains simultaneously (i.e., pain intensity, pain disability and pain catastrophizing), we found a smallto-medium summary effect size, which was robust to outliers, influential studies and different analytical strategies.These findings indicate that patients with CLBP can benefit from self-management interventions in terms of several clinically relevant outcomes.However, if we included only low RoB studies, the summary effect size decreased and was not statistically significant anymore.This finding is probably a result of the limited number of available studies, and it suggests that more high-quality (i.e., low RoB) studies are needed to obtain a more robust conclusion on the efficacy of digital interventions in CLBP.For pain intensity, we found the smallest effect size of all investigated pain domains, which was somewhat surprising given that this was the primary outcome in most of the studies.This small effect size might have been additionally inflated by one study with a disproportionate influence (Garcia et al., 2021; Baujat Plot visualizing the influence can be found in Figure S1).However, after removing the study in the influence analysis the summary effect size did slightly decrease, but was still statistically distinct from zero, showing that the overall finding did not depend solely on this study.
The analysis for the outcome pain disability resulted in larger effect sizes than for pain intensity.We think this is an important finding and may be of clinical relevance.Digitally mediated self-management interventions for CLBP may help patients to cope with their pain.Chronic pain is known to be particularly difficult to treat and reduce in severity (Nilges & Nagel, 2007).Self-management interventions might convey ways to make life easier despite the pain, which may be reflected in the more pronounced improvement in painrelated functioning (i.e., pain disability) compared with pain intensity.With methods like CBT, coaching, motivation and relaxation 'psychosocial factors' of pain such as the influence of higher pain severity on the emotional state, or the dysfunctional cycle between increased catastrophizing and greater pain severity might be treated specifically, which play an important role in the development and maintenance of chronic pain (Nilges & Nagel, 2007).
Furthermore, the effect for pain catastrophizing was not as large as we expected, even though the absolute point estimate for the effect was also larger than for pain intensity.Digital self-management interventions are expected to reduce pain-related thoughts in patients and manage feelings of helplessness if they are in pain (Fritz & Kongsted, 2023).Also, recent evidence suggests that pain catastrophizing is a key outcome in patients with back pain and that it is worth investigating treatment programs that are specifically tailored to reduce catastrophizing (Darnall et al., 2020).However, because only a minority of the included studies assessed pain catastrophizing as an outcome, meta-analytic evidence is yet too small to support the expectation that self-management reduces catastrophizing.Due to recent evidence, pain catastrophizing will likely attract further attention in the field, which will lead to additional evidence on whether self-management interventions are effective for reducing pain catastrophizing (Petrini & Arendt-Nielsen, 2020).

| Comparison with previous research
Previous meta-analytical research on self-management methods for low back pain has provided similar results like in this study.In summary, regardless of whether self-management programs were provided through digital devices or not, and regardless of whether back pain in participants had already become chronic, summary effect sizes were in the lower to medium range (Du et al., 2011(Du et al., , 2017(Du et al., , 2020;;Oliveira et al., 2012).For example, one more recent study investigated the efficacy of e-health in the self-management of CLBP and found similar results for pain intensity and pain disability immediately after the intervention as well as at follow-up time points (Du et al., 2020).In the present paper, however, the amount of follow-up data was insufficient for analysis.The results from the recent meta-analysis (Du et al., 2020) were well suited for a plausibility check of our search results.We found an overlapping amount of five papers that were included in both analyses.Two papers Du et al. included were not included in this paper due to our 10-year search limitation.Another study was not included in this paper because some patients included did not suffer from pain for over 3 months.
Our preliminary findings did not detect major differences between types of interventions.According to a recent meta-analytical review (Schütze et al., 2018), for example, the combination of CBT and exercise works best to improve pain catastrophizing in patients with chronic noncancer pain.The authors conclude, however, that treatment should be individualized to the patient.As it is too soon to tell which patients benefit most from which interventions, this identified gap in evidence should be considered a crucial next step in developing effective treatment options.Digital interventions offer a promising opportunity to tailor treatments to individual patients, for example, through Just-In-Time Adaptive Interventions (JITAIs; Nahum-Shani et al., 2018).JITAIs aim to provide suitable microinterventions at the best time point and can take other factors into account such as the current location, social context or current symptom severity.The decision on whether a micro-intervention is provided or not can, for example, be based on continuously measured data (e.g., tracking data from a mobile phone) or based on selfreported information that may be collected randomly or at predefined times.Recent evidence points at the effectiveness of JITAIs for a broad range of conditions (Wang & Miller, 2020).We are not aware, however, of such an intervention for patients with pain conditions.
The most recent, but small scoping review on mhealth interventions for low back pain patients (Rintala et al., 2022), supports the need for more targeted interventions.However, due to the limited evidence in this field, far more research is needed to investigate whether self-management interventions may be better in some subgroups of patients than in others, especially if they are provided at the right time and dose and in the right situation.

| Implications for clinical care
As part of the management of patients with low back pain, self-management has great potential.Due to the very large number of patients affected by (chronic) back pain all over the world, it is challenging-or in some parts of the world even not possible-to provide regular face-to-face visits with a healthcare specialist.Thus, easily accessible services are warranted.Digital self-management programs are well suited for this purpose as these programs are usually designed to be used in many places, as long as patients have access to their smartphone or other mobile device.In fact, most (7 of 12 studies) of the self-management programs that we investigated in this study were based on smartphone apps that allow patients to access the selfmanagement information regardless of where they are.Thus, due to favourable properties including accessibility and cost-effectiveness, digital self-management programs are expected to reduce the burden on the healthcare system (Kheirinejad et al., 2023;Stec et al., 2019).Furthermore, digital self-management programs open up opportunities to monitor progress and share data with physicians, contributing to more effective patient care (Didyk et al., 2022).
Given that CLBP is a chronic disease and many patients are affected for life, it is essential to learn how to actively manage their disease (Barlow et al., 2002).Selfmanagement approaches can help patients acquire skills to improve the ability to better manage their disease.A common model suggests that these skills include problemsolving, decision-making, resource utilization, forming of a patient/healthcare provider partnership and taking action (Du et al., 2017;Lorig & Holman, 2003).These and other skills can enhance patients' self-efficacy, a key mechanism that is believed to make self-management effective (Lorig & Holman, 2003).This hypothesis is supported by a large body of evidence suggesting that self-efficacy and changes in self-efficacy are associated with changes in health behaviour and health status (Bandura, 1997).

| Strengths and limitations
Strengths of this study are that we used a comprehensive search strategy, combining both large databases (MEDLINE, EMBASE and Cochrane), one database for unpublished studies (ICTRP) and subject-specific databases (PsychInfo and CINAHL).We arrived at 12 included studies through our well-defined and strict eligibility criteria.We restricted our search to the last 10 years, as we A limitation of this study may be that 10 of 22 studies that were eligible for full-text screening were obtained through citation tracking.This unexpected ratio could have been resulted from an insufficient search strategy.However, as described above, our complex search strategy has been well considered which is reflected, for example, by using MeSH terms, a wide variety of search terms and the Cochrane Filter for Randomised Controlled Trials.Furthermore, a recent study suggested that the scientific community should reconsider introducing citation tracking as a stand-alone search strategy, as it improves the quality of the search (Hirt et al., 2023).
We applied several methods to control for publication bias to critically assess the validity of our findings.Some methods (funnel plot symmetry, selection model, trimand-fill method, rank correlation and Egger's regression test) did not detect publication bias while other methods (PET-PEESE and robust Bayesian model) detected publication bias to some extent.Fortunately, in most instances where we adjusted the effect sizes for the presence of small-study effects or publication bias, the adjusted effect sizes remained statistically significant.This overall picture of publication bias assessment does not clearly suggest the inflation of our reported meta-analytical findings-but still is cause for some concern.However, as we included trial registries (such as the Registry Database ICTRP) in our search, it can be assumed we found a substantial portion of existing research-mitigating the risks of publication bias as much as possible.
Overall, the quality of included primary studies was suboptimal.Only three of the 12 studies examined in our RoB2 analysis had a low RoB.However, bias was not found in the most critical domains (randomization process and measurement).When we included only low RoB studies in our meta-analyses, the summary effect sizes were at times substantially decreased, indicating an inflation of effect size estimates due to low-quality studies.This inflation was especially pronounced for pain intensity and pain catastrophizing.Therefore, all results should be carefully considered as preliminary but not as conclusive evidence.This interpretation still holds even when considering the GRADE rating for the quality assessment.While GRADE may suggest a higher quality of evidence than RoB2 based on the percentage of participants from low-risk studies, it may overestimate quality as the majority of included studies in our meta-analysis had a suboptimal RoB rating.
Lastly, the included studies were very heterogeneous and differed greatly in context, digital delivery mode, duration, guidance and settings.This makes it more challenging to draw consistent conclusions for clinical practice.Unfortunately, as the number of available studies was expected to be limited, we did not restrict our search criteria to acquire a more homogeneous sample.

| Conclusion
Our results point towards a small but relatively robust pooled effect of digital self-management interventions for patients with CLBP.The findings provide preliminary evidence for the efficacy of self-management interventions in improving several key CLBP outcomes, including pain intensity, disability and catastrophizing.The quality of evidence was moderate or high.However, the total number of included studies was limited.Further high-quality studies are needed to replicate and expand our preliminary results, ensure the generalizability in other populations and investigate which patients benefit most from which interventions.
interventions can help patients reduce their pain intensity and disability.A minority of studies point towards the possibility that digital interventions can reduce pain catastrophizing.Future research should further explore which patients benefit most from these kinds of interventions.

F
I G U R E 1 Flowchart of the literature search and study selection.CINAHL, Cumulative Index to Nursing and Allied Health Literature; ICTRP, International Clinical Trial Registry Platform; n, number of studies; PROM, patient-reported outcome measures.Source: From Page et al. (2021).) 68 (I: 34; C: 34) 50% (I: 44%; C: 56%) I: 59.5; C: 57.1 24 CBT + Exercise Physical activity booklet and activity tracker, face-to-face and telephonebased health coaching sessions,

F
Forest plot showing the observed outcomes and the estimate of the random-effects model for pain intensity.I G U R E 4 Forest plot showing the observed outcomes and the estimate of the random-effects model for pain catastrophizing.F I G U R E 5 Forest plot showing the observed outcomes and the estimate of the random-effects model for pain interference.F I G U R E Forest plot showing the observed outcomes and the estimate of the three-level model for all pain domains (pain intensity, pain catastrophizing and pain interference).

F
G U R E 7 (a-d) Funnel plots for pain intensity, pain catastrophizing, pain interference and combined outcomes.
were mainly interested in interventions developed during a time when internet-based and smartphone-based interventions were emerging due to the digital transformation in health care.Interventions that were previously developed (such as email-based interventions or bibliotherapy) were not in the focus of our investigation.The trend towards digital technology was evident to us as both the interest in and publication of digital interventions for CLBP increased substantially within the last decade.
The MeSH terms varied in the different databases.All the search terms marked with MeSH could be found as such in at least one database.If the search term could not be found as a MeSH term, it was only searches in the categories title, abstract and keyword ti,ab,kw.= searched in the categories title (ti), abstract (ab) and keyword (kw). Note: