The effect of chronic low back pain on postural control during quiet standing: A meta-analysis

Low back pain (LBP) has been associated with altered body sway during quiet standing, but the pattern of results is inconsistent. The purpose of this meta-analysis is to examine the effects of vision (eyes open, eyes closed) and changing the support surface (foam surface, firm surface) on postural sway during quiet standing in individuals with chronic LBP (cLBP). Five electronic databases were searched on March 27th, 2022. Of 2,856, 16 studies (n = 663) were included. Across all conditions, we found a positive and medium effect size (g = 0.77 [0.50, 1.04]) that represented greater body sway in individuals with cLBP. Subgroup analyses revealed medium effects during eyes open conditions (firm surface: g = 0.60 [0.33, 0.87]; foam surface: g = 0.68 [0.38, 0.97]), and large effects during eyes closed conditions (firm surface: g = 0.97 [0.60, 1.35]; foam surface: g = 0.89 [0.28, 1.51]). We quantified effects of self-reported pain and found a moderate effect during eyes closed plus firm surface conditions (Q = 3.28; p = 0.070). We conclude that cLBP is associated with increased postural sway, with largest effect sizes evident when vision is removed and when self-reported pain intensity is higher.

Study selection. The procedure to select studies was conducted using the following steps. First, duplicate studies were removed using EndNote X9 (Clarivate) 26 . According to the inclusion and exclusion criteria, a researcher (JP) screened titles and abstracts of all studies, and then two researchers (JP and VN) independently rescreened the studies through the full-text versions. In case of disagreement, the last researcher (SAC) was consulted for the final decision. The final studies were selected in our meta-analysis. We included studies that: (1) recruited cLBP subjects with pain duration of greater than or equal to 3 months, (2) recruited adult population ranged from 18 to 65 years, (3) compared a cLBP group and healthy control group (HC), (4) used cross-sectional design, and (5) used CoP parameters during bipedal quiet standing. Exclusion criteria were studies that: (1) recruited athletes with cLBP or cLBP subjects who have had surgery for lumbar disc herniation or other diseases, (2) recruited recurrent or episodic LBP subjects who reported semicontinuous LBP or at least one episode of LBP during the previous 3 months, (3) did not include a control group, (4) did not measure CoP variables (e.g., the center of mass, the center of gravity, and the joint angles) to compare postural control between groups, (5) did not investigate postural control during a bipedal quiet standing task (i.e., perturbation tasks, balancedexterity tasks, anticipation postural control task, one leg standing task, and dual tasks are exclusion criteria), and (6) did not report statistical results for calculating the effect size.

Data extraction.
To calculate the pooled effect size, all the CoP outcomes were extracted from included studies. CoP data was classified according to body sway and body sway velocity to reduce the heterogeneity. Body sway included the sway amplitude, displacement, area, and dispersion, while body sway velocity variables represent the velocity of the CoP 18 . Both parameters were collected as means with their standard deviation and sample size. If a study reported a median and interquartile range, those values were converted into a mean and standard deviation based on the Quantile estimation method 27 to calculate an effect size. In the case that a study did not report group level descriptive data, effect sizes were calculated using a combination of sample size, mean differences, p-values, and t-values. If a study reported a standard error, the standard error was converted into a standard deviation based on the corresponding sample size. Furthermore, we extracted demographic information (i.e., sex, age, height, and weight) and pain related characteristics when available (i.e., pain intensity, pain duration, and disability). Two authors, JP and VN, extracted and checked data. Effect size. Hedges' g was used to calculate effect sizes of each studies' outcome. The rationale for using Hedges' g is that it is less affected by the impact of sample size compared to Cohens' d 28 . The Hedges' g is interpreted based on 3 stages: (1) values from 0.2 to 0.49 represent a small effect size; (2) values from 0.5 to 0.79 represent a medium effect size; (3) values greater than or equal to 0.8 represent a large effect size.

Meta-analyses.
We first calculated overall effect sizes for group differences in body sway and body sway velocity during quiet standing across all experimental conditions (i.e., vision and proprioception). Next, we calculated effect sizes for group differences in body sway and body sway velocity across different experimental conditions: (1) on a firm surface with eyes open (Firm with EO), (2) a firm surface with eyes closed (Firm with EC), (3) on a foam surface with eyes open (Foam with EO) and (4) on a foam surface with eyes closed (Foam with EC). In each experimental condition we also assessed moderator variables based on direction of body sway and body sway velocity (anterior-posterior [AP], the medial-lateral [ML], and both [AP-ML] directions) and pain (low and high). For the levels of pain intensity, the results from a visual analogue scale (VAS) or numerical rating scale (NRS) were extracted. In the case where an included study reported pain scores based on the range from 0 to 100, their scores were converted to a value from 0 to 10. Pain intensity was then classified into two levels using the median value across all studies (4.73): a higher level of LBP is equal to and more than the median, which we labeled 'High' , whereas a lower level of LBP is less than the median and is labeled as 'Low' .  29 was used to evaluate the quality of each study included in the meta-analysis. The QAT contains fourteen questions to access internal validity. The details on the questions are described in the supplementary material (Table S3). All the questions were rated as being positive ("Yes"), negative ("No"), cannot determine ("CD"), not applicable ("AP"), or not reported ("NR"). Two authors, named as JP and VN, rated the quality of studies. The disagreements were discussed until consensus was reached. In cases where JP and VN could not reach an agreement, they discussed it with SAC. The quality of every study was evaluated by "Good", "Fair", and "Poor. " Data analysis. The Comprehensive Meta-Analysis software package version 3 (CMA; BioStat, Englewood, New Jersey) was used to calculate effect sizes and 95% confidence intervals using a random effect model for each analysis. If a study included in the meta-analysis had multiple CoP outcomes, we calculated multiple effect sizes corresponding to CoP outcomes and then averaged those effect sizes into a single effect size per study according to body sway or body sway velocity and standing conditions. For instance, Caffaro et al. 13 reported 12 different CoP outcomes of body sway that were different according to experimental conditions and directions (See the Supplementary material 2). When the meta-analysis for body sway was conducted, the CMA program calculated 12 effect sizes from 12 CoP results and then averaged all effect sizes into a single effect, which was 0.77 (see the first row in Fig. 2a). An I 2 statistic and prediction interval were used to assess the heterogeneity for each analysis. I 2 values higher than 75% represent high heterogeneity, and the prediction interval (PI) represents the range of true effect sizes for 95% of all comparable studies 30 . To assess publication biases for each analysis, Egger's regression test was first used 31 . The significant level for the bias was set to 0.05. If publication bias was found through Egger's regression test, Duval and Tweedie's Trim and Fill 32 was conducted to reassess publication bias using a funnel plot which can test the symmetry of effect sizes of included studies. Asymmetric funnel plots suggest that publication bias may exist. Using the trim and fill approach, some data was imputed to make the figure look symmetric, and then the changed overall effect size including the imputed data was compared to the previous overall effect size without imputed data. Unless 95% confidence intervals of both overall effect sizes were overlapped, there is an effect of publication bias. For subgroup analyses, Q-statistics were used to assess the moderator effects of movement direction (AP, ML, and AP-ML) and pain intensity (Low and High). The significant level for Q-statistics was set to 0.10. Effect sizes for sublevels of variables (e.g., AP, ML, AP-ML, Low, and High) were statistically analyzed. Significant p-values were set at a level of 0.05, corrected using the false discovery rate (FDR). Meta-or subgroup analyses with only one study were excluded 33 .
GRADE assessment. Grading of recommendations, assessment, development, and evaluations (GRADE) was used to evaluate the evidence certainty based on five subcategories: 'Risk of bias' , 'Inconsistency' , 'Indirectness' , 'Imprecision' , and 'Publication bias' . The ratings for those categories were 'No serious' , 'Serious' , and 'Very serious' . The level of evidence certainty was downgraded if the meta-analysis (i) included 25% of studies with poor quality assessed by QAT (1 level of Risk of bias) or all studies with poor quality (2 level of Risk of bias), (ii) showed higher than 50% of I 2 (1 level of Inconsistency)-I 2 values have been used previously to evaluate inconsistency in CoP outcomes 34,35 -(iii) did not include direct evidence related to the main question (1 level of Indirectness), (vi) included less than the optimal sample size (n = 244) calculated based on 5% of margin of error, 95% confidence intervals, and total sample size (n = 663; 1 level of Imprecision), and (v) showed publication bias resulting from Egger's regression and trim and fill test (1 level of Publication bias). Combining all the ratings, the certainty of the evidence in meta-analyses were interpreted based on four levels: 'High' , 'Moderate' , 'Low' , and 'Very low certainty' . The high certainty of evidence represents high confidence that the true effect is close to the effect of the estimate, while very low certainty represents very little confidence that the true effect is close to the effect of the estimate. Initial ratings began at a low level since all studies included in the meta-analyses were observational studies. However, the level can be graded up (i) if the pooled Hedges' g was higher than or equal to 0.8, (ii) evidence of a dose-response relation was found, and (iii) if all the bias and/or confounds were plausible 36,37 .

Results
Fig. S1 in supplementary material 1 shows the PRISMA flow chart for the first search. Figure 1 shows the PRISMA flow chart for the second and final search. The search strategies initially yielded 2856 relevant articles, and one additional article 38 extracted from a systematic review 17 . Among them, 842 duplicates were removed, leaving a total of 2014 articles. After investigating titles and abstracts, 1922 articles were excluded. From the remaining 92 articles, we excluded an additional 76 articles based on the criteria outlined in Fig. 1. Additional information on why studies were excluded from the meta-analyses is shown in Table S2. A total of 16 articles were included in this meta-analysis.
Demographic data. Table 1 represents demographic information for cLBP and HC groups for each article.
The bottom three rows in the Total section of Table 1 indicate the summed and averaged demographic information on the overall population of cLBP, HC, and total individuals respectively. The first row in the Total section of Study characteristics. Table 2 shows information extracted from the 16 included studies, including which questionnaires/scales were implemented, the methods used to collect CoP data, the types of quiet standing tasks, and whether significant differences were found between cLBP and HC groups. Across all studies, pain intensity was most often assessed using a VAS and disability was most often assessed using the Oswestry disability index (ODI). Three studies did not report pain intensity 22,23,40 . Eight studies reported disability for pain using ODI 15,[21][22][23]39,[41][42][43] , and four studies used the Roland Morris disability questionnaire (RM) 13,14,20,24 . Four studies did not assess disability 6,38,40,44 . Anxiety, cognitive function, and health condition questionnaires are also included in Table 2 if included in the corresponding study. For the measurement of CoP variables, the sampling rate ranged from 40 to 1500 Hz. Eight studies used 100 Hz 13,14,[20][21][22]24,38,42 , and two studies did not report the sampling rates 23,44 . A Butterworth filter was used for all studies with cut-off frequencies ranging from 2.5 to 35 Hz and eight studies not reporting cut-off frequencies used 14,15,23,[38][39][40][41]44 . The duration of each trial between studies varied from 10 to 120 s. Six articles used 30 s 6,14,20,38,39,41 , and one study did not report trial duration 44 .
Most studies (5 studies) used three trials of quiet standing 13,20,22,23,39 , and three articles did not report the number of trials 6,15,42 . A total of seven studies gave subjects instructions for quiet standing tasks 15,21-23,41-43 . Among them, five studies instructed subjects to stand as still as possible 15,21,23,41,43 . Two studies instructed subjects to relax 22 and to be comfortable 42 during quiet standing. The remaining nine studies did not report their instruction for subjects 6 Table 2 shows the significant differences between cLBP and HC groups. Twelve studies found CoP movement was greater or faster in cLBP group as compared to HC group 6,[13][14][15][20][21][22][38][39][40][41]44 , while one study found that CoP movement was smaller or slower in cLBP group than HC group 21 . Four studies did not find significant between group differences 23,24,42,43 . Data handling. From 16 studies, we extracted 111 combinations of means, standard deviation, standard errors, sample sizes, p-values, and/or t-values. A total of 58 effect sizes were calculated from 11 studies using a combination of means, standard deviations, and sample sizes 14,15,20,22,24,[38][39][40][41][42]44 . A total of 20 effect sizes were calculated from two studies by converting standard errors to standard deviations. A total of five effect sizes were calculated from one study that reported medians and interquartile ranges 43 . In this case, we converted those values to means and standard deviations based on the Quantile estimation method 27 . Four studies did not report means, standard deviations, or both 6,21,38,44 . In this case, we extracted sample size and other statistical valuesindependent t-values 21 and p-values 6,38,44 -and calculated 28 effect sizes using combinations of these values. Pooled effect sizes of CoP movement for body sway and body sway velocity. Figure 2a shows body sway effect sizes between cLBP and HC groups. A total of 15 effect sizes were calculated from 15 studies. Within each study, effect sizes were averaged across all task conditions. Black squares represent effect sizes for each study and the whiskers represent the confidence interval. These values are also shown on the right of the Fig. 2a      ; the squares stand for the pooled effect sizes for results of CoP directions (AP, ML, and AP-ML); the circles stand for the pooled effect sizes for pain intensity (Low and High). All whiskers represent confidence intervals for each analysis. The rightmost column lists effect sizes and confidence intervals, which were calculated into Hedges' g. Figure 3a displays results for the firm surface with EO condition (red colors). Figure 3b represents results on quiet standing on the firm surface with EC (yellow and orange colors). Figure 3c represents results on quiet standing on the foam surface with EO (green colors). Figure 3d represents results on quiet standing on the foam surface with EC (blue colors). * p < .05; † p Q-test < .10.  Fig. 3b). In the low pain level represented by the darker yellow circle in Fig. 3b Figure 3c shows effect sizes for body sway on a foam surface with EO. All effect sizes were positive in favor of cLBP group. A medium and significant pooled effect size was found (p < 0.001, g = 0.68 [0.38, 0.97]) which is represented by the green diamond. I 2 value was 7.55%, and the PI was −0.05 to 1.41 (Table S4). No effect of movement direction (Q = 2.85; p = 0.240, I 2 = 0%, PI = NaN) was found. The green squares in Fig. 3c  Subgroup analyses for body sway velocity. Figure 4 indicates effect sizes for body sway velocity. No significant pooled effect size or moderator effect was found as shown in Fig. 4a and 4b. Publication bias. Publication biases were assessed using Egger's regression and Trim and Fill analysis. Table S5 in the supplementary material 1 represents the statistical results of Egger's regression test. The last column shows p-values for all meta-and subgroup analyses. Significant biases were found in the three conditions of body sway: the firm with EC (p = 0.033) and foam with EO (p = 0.006) and EC (p = 0.033). Those analyses were reassessed using the Trim and Fill analysis. Fig. S2 in the supplementary material 1 represents the results of the Trim and Fill analyses. The circles in the funnel plots represent standard errors by Hedges' g of each study. The white filled diamond under the x-axis represents the overall 95% confidence interval for each meta-analysis. For instance, Fig. S2a shows the funnel plot for the body sway meta-analysis when collapsing across all conditions within each study. Each study is represented by an open circle. The absence of a filled black circle means that no imputation had to be conducted for this meta-analysis. Open circles were evident for the firm surface with EO (Fig. S2b) and firm surface with EC (Fig. S2c). Fig. S2d and e show the funnel plots for body sway on a Foam surface with EO and EC. Black circles are included in each plot. Each black circle represents one imputed study, which makes the distribution of the values symmetric. The black diamond on the x-axis represents the mean confidence interval when including the imputed data. When the white and black diamonds overlap, we infer that publication bias is not affected. This was evident in each case. Figs. S2f-h show funnel plots for body sway velocity meta-analyses. Although black dots are shown in each case, overlap between the white and black diamond suggests no evidence of publication bias in either case.

Assessment of the quality in studies.
The results of QAT for every study are presented in Table S3. Five studies were evaluated as being good quality 13,14,39,41,43 , and six studies were of fair quality 15,[20][21][22][23][24] . Five studies were rated as poor quality 6,38,40,42,44 . GRADE assessment. GRADE was used to assess the evidence quality and the results are shown in Table 3.
The first column in Table 3 lists the different outcome measures and conditions, and the second column represents the number of studies included in the corresponding meta-analysis as well as the total sample size and CoP variables included. The third to seventh columns indicate the ratings for subcategories of GRADE. The eight column ("other") shows decisions on publication bias and other considerations that can upgrade the quality of evidence, which is followed by estimate of outcome values. The final column represents the final decision for rating evidence certainty. As a result, the evidence of CoP outcomes on Firm with EC in body sway, CoP outcome in body sway velocity, and CoP outcome on Firm with EO in body sway velocity were rated as low certainty. The rest of evidence was rated as very low certainty.

Discussion
Compared with controls, individuals with cLBP showed increased postural sway during quiet standing. Subgroup analyses revealed larger effect sizes when vision was removed and when self-reported pain intensity was higher but only when subjects completed the task on a firm surface. In contrast, we did not find strong support for an effect of cLBP on postural sway velocity.
When averaging data across all experimental conditions within each study, we found strong evidence that cLBP is associated with greater postural sway during quiet standing. Our findings suggest that cLBP is associated with altered postural control 7,45 . Several explanations may account for this change in control including reduced proprioceptive acuity 46 , restrictive trunk movement 5 , and protective trunk muscle strategies 47,48 . Reduced proprioception means that larger movements must be made for a change in internal estimate of CoP position to be updated 11 . Once perceived, a correction in the opposite direction may also then be disproportionality large 22 . Restrictive trunk movement and co-contractions across trunk muscles may also contribute to changes in posture by shifting control from the trunk to the ankles. LBP has been associated with co-activation of left and right gluteus medius muscles 49,50 , sustained activation of back and abdominal muscles during quiet standing 43 , leading to a restriction in trunk movement reflective of a protective strategy 12,47 . In addition to changes in trunk muscle activation, a shift in motor control away from the trunk has also been evidenced, with the ankle being most commonly observed 5,19,51,52 . This shift in control to the ankle can be problematic given that perturbation studies show that trunk control is most effective for maintaining postural equilibrium 53 , with the ankles contributing minimally to posture. Taken together, the greater CoP sway in cLBP evident in our findings may result from altered proprioception and restrictive and protective trunk strategies, which combine to shift postural control to the ankles.
Previous systematic reviews on back pain and postural sway have reported inconsistent results 5,18,54 . In the current meta-analysis twelve out of 16 articles reported greater CoP sway in cLBP compared to healthy controls. The difference in findings between previous reviews and the present meta-analysis may be due to different inclusion criteria for the back pain group. Our meta-analysis included studies that recruited individuals with cLBP, while previous reviews included individuals with both recurrent and cLBP 5,16,18 . Smaller CoP movements have been reported in LBP studies (e.g., Mok et al. 55 and Salavati et al. 56 ), these were excluded from our metaanalysis because they did not meet our inclusion criteria for cLBP. Models of motor adaptation have been proposed for muscular pain 7,57,58 , such that motor strategies are altered based on pain chronicity. For instance, the progression from acute to chronic neck-shoulder pain is associated with a progressive decrease in variability of trunk movement with a corresponding increase trunk stiffness 57 . In the context of LBP, it is plausible, therefore, and eyes closed [EC]); the squares stand for the pooled effect sizes for results of CoP directions (AP, ML, and AP-ML); the circles stand for the pooled effect sizes for pain intensity (Low and High). All whiskers stand for confidence intervals for each analysis. The rightmost column represents effect sizes and confidence intervals, which were calculated into Hedges' g. Figure 4a indicates results on the condition on firm surface with EO (red colors). Figure 4b represents results on quiet standing on the firm surface with EC (yellow and orange colors). www.nature.com/scientificreports/ that acute LBP is associated with an increase in variability of trunk movement and a decrease in trunk stiffness as individuals search for optimal solutions to minimize the pain. This could potentially lead to reduced postural sway. As chronicity develops, a progressive decrease in variability and increase in stiffness would lead to an increase in postural sway. We therefore suggest that distinguishing between acute, recurrent and chronic phases of LBP is critical when assessing the impact of low back pain on postural sway. Removing vision during quiet standing increased effect sizes by ~ 0.2 irrespective of which surface individuals were standing on. Effect sizes were similar for both eyes open conditions (0.64 and 0.68) and both eyes closed conditions (0.98, 0.89). Changing the surface on which subjects were standing had a minimal impact on effect size. If a deficit in proprioception is a dominant factor in altered postural control in cLBP, then one would have expected effect sizes to be largest in the eyes closed plus foam surface condition. This was not the case in our data, suggesting that the change in surface had a minimal impact. We offer three explanations. First, the number of studies contributing to the foam condition was smaller as compared to the firm condition, and second, altered proprioceptive feedback in cLBP may have reached a ceiling in the eyes open plus firm surface condition (effect size ~ 0.6), such that changing the surface to foam had a minimal additional effect. Other factors, such as deconditioning in cLBP 59 , pain-related fear, and catastrophizing 60 , may also be associated with greater sway, but we did not assess these factors in the current study. In contrast to the surface condition, changing visual input had a larger impact on effect size. Our finding is consistent with other studies in chronic pain that have manipulated vision [61][62][63] . For instance, individuals with cLBP show greater variability in stride time and minimum Table 3. GRADE assessment for the certainty of evidence. The Grading of Recommendations Assessment, Development and Evaluation (GRADE) assesses certainty of evidence body based on four criteria: High, Moderate, Low, and Very low certainty. The High certainty represents very confidence that the true effect lies close to that of the estimate of the effect; the Moderate certainty represents moderately confidence that the true effect is likely to be close to the estimated effect; the Low certainty represents limited confidence in the estimate of the effect; the Very low certainty represents very little confidence in an estimated effect. Firm EO The firm surface with eyes open; Firm EC The firm surface with eyes closed; Foam EO The foam surface with eye open; Foam EC The foam surface with eyes closed; CoP Center of pressure; g, Hedges' g; CI confidence intervals; PI prediction intervals. a The I 2 was higher than 50% that means a high level of heterogeneity. b The result of this meta-analysis showed a high and pooled effect size. In this case, we upgraded the level of certainty of evidence 36 . c The sample size for the meta-analysis was lower than an optimal sample size (n = 244) calculated based on 5% of margin of error, 95% CI, and the total sample size (n = 663). www.nature.com/scientificreports/ foot clearance during gait when vision was removed 61 . Other evidence comes from individuals with chronic neck pain 62,63 . Underestimating the extent of neck rotation using virtual reality led to greater movement prior to the onset of pain 62 . During standing, greater body sway was found in individuals with chronic neck pain compared to controls in eyes open condition, with the between group difference becoming larger when vision was removed 63 .
With eyes closed, we also found evidence that effect sizes were greatest when comparing controls and individuals who self-reported higher as compared to lower pain intensity. This effect was not evident in the eyes open conditions, suggesting that vision may mask or compensate for the effect of pain intensity on postural sway. Our findings are consistent with other studies that have found that individuals who report higher pain intensity (VAS ≧ 5) show greater CoP sway compared to individuals with lower pain intensity (VAS = 1.47) 64 and controls 14,38,39,41 . Importantly, individuals with lower LBP intensity (VAS < 2) were found to have smaller and slower CoP sway compared to healthy controls 55,56 , which suggests that pain and intensity in addition to pain chronicity should be considered when associating cLBP with postural sway.
We note several limitations of this meta-analysis. First, we restricted our analysis to studies which recruited individuals with cLBP since previous evidence revealed that there is a different impact of the LBP disease stages on a postural control strategy 7,57,58 . Our findings, therefore, can only generalize to individuals with cLBP. Second, we did not consider the impact of specific vs. non-specific cLBP on postural sway. Third, we identified 16 studies that included an eyes-open condition, but the number of studies included in the different analyses and subanalyses decreased. Fourth, the certainties of the outcomes included in our meta-analysis were rated as low or very low. In part this is because all included studies were observational studies and by default these are initially rated as low certainty. A high magnitude of heterogeneity across CoP outcomes also contributed to the low or very low ratings. High heterogeneity in body sway has been noted in previous systematic reviews [16][17][18] . We move the field forward by demonstrating that a small portion of the heterogeneity can be accounted for by pain intensity. Nevertheless, we acknowledge that even when accounting for pain intensity and CoP direction, there was still a considerable amount of variance unexplained. Other factors such as different sampling rates, filters, trial durations, number of trial repetitions, CoP equations, and instructions for the quiet standing tasks may help explain this variance, but our sample size was not large enough to address these potential factors. As additional studies are published, future meta-analyses can address these important factors. Nevertheless, we adhered to strict inclusion guidelines and our findings do suggest that vision and pain intensity play important roles in the impact of cLBP on postural sway.

Conclusion
Impact of this meta-analysis is that cLBP is associated with greater sway, which likely comes from a visual occlusion and a change in control strategy. Interventions that target both of these factors should therefore be explored.

Data availability
All data generated or analyzed during this study are included in this published article (and its Supplementary Information files).