Low back pain (LBP) contributed to most years lived with disability to the global burden of disease in 1990 as in 2017 [1]. The impact of LBP ranges from causing minor inconvenience to substantial restrictions in daily activities and, in extreme cases, disability and early retirement. Although there may be improvements when considering population age, the overall years lived with disability from LBP is rising and needs to be addressed [1]. Standard treatment recommendations for LBP often incorporate exercising and advice regarding physical activity [2], and it has been demonstrated that exercises for chronic LBP improve outcomes such as pain or disability to a certain degree [3-5]. Differences in effects between exercises with a distinct training focus have been appraised as negligible [4,5]. The limited effectiveness motivates the exploration of new ways for enhancing these treatments, as it has already been outlined by other authors [6]. Considering that changes of motor control of the lumbar region are discussed as a plausible cause for recurrence of LBP [7] and given that feedback plays a central role in motor learning [8], digital tools that make movement patterns more visible could be one such way to enhance exercise treatments. These interventions could be used together with other treatments as implemented in previous studies [6,9-11] or independently as needed, as a form of self-management. Many people with LBP do not request treatment, especially those with mild disability [12]. Therefore, supportive technology that can provide some degree of guidance at home, while maintaining independence, and could be especially interesting for this group of people.

Motor control can be described “...as the way in which the nervous system controls posture and movement to perform a specific motor task, and includes consideration of all the associated motor, sensory, and integrative processes” [7]. Physical characteristics and movement behaviors assessed to derive insights into deviations in people with LBP concerning these processes have revealed many new insights but still demand further clarification [7,13]. Examples of movement differences such as the limitations in range of motion (ROM) of the lumbar spine were found in all movement planes [14] and limited ROM of the trunk in the frontal plane but not in the sagittal plane seem to precede the occurrence of LBP [15]. Differences in the trunk region are thought to relate to differences in postural balance [7], which have been found by many studies [13,16,17]. Consequently, practicing movement tasks that focus on movement of the lumbar spine and hip could have the potential to enhance postural balance. Nevertheless, as highlighted earlier, the associations of LBP and movement behavior have not been fully clarified [7].

Altered movement behaviors seem to extend to tracing tasks, which work with feedback on trunk movements and have been used as an indicator for motor control in laboratory settings [18,19]. These studies used tasks that required participants tracing a circular pattern [18] or performing flexion movements [19] with their trunk, while receiving concurrent feedback. Results regarding the accuracy were conflicting, as one study found a difference between people with and without LBP in the accuracy of the tracing [18], whereas the other did not [19]. However, the latter study confirmed differences with respect to timing relative to the feedback between the groups [19]. Similar tasks may serve not only as a proxy measure of trunk motor control but also as training opportunity. It was recently found that practice to keep the lumbar spine constantly neutral during a box lift task was more successful when participants obtained digital feedback than when the participants used a mirror [20]. As described earlier, people with LBP seem to show movement patterns that deviate from those in people without LBP in different manners, such as displaying a high degree of rigidity or little control at all on their movement [7]. In tracing tasks, where movement becomes visible in relation to a target, people with LBP may develop well-coordinated trunk movements that are neither rigid nor loose. Furthermore, proprioception seems to be affected in people with LBP [21], which might relate to difficulties with fine-tuned trunk movements [18]. Early results indicate that there may be improvements in proprioception through interventions using feedback on trunk movement in people with LBP [22]. In short, we assume that such training may reduce postural balance impairments by restoring more adequate movement behaviors of the trunk.

The effects of exercising interventions directly on postural balance have been studied previously. Meta-analyses on intervention studies with older people suggest that balance training [23] and Pilates [24] but not programs focusing on strength or mixing different kinds of exercises [23] can enhance postural balance. Although these results originate from older participants, studies in people with LBP imply similar development of pain [25] and treatment results [26], largely independent of age. Furthermore, altered trunk characteristics have also been reported in younger people with LBP [27]. Multiple interventional studies with people with LBP found an impact of exercising interventions on at least one of the investigated criteria describing balance [28-30], whereas in another study no differences in postural balance were detected [31]. However, different tasks with varying requirements were used; for example, standing on moving ground [28], standing on a single leg [29], and assessments in squat positions [30].

Reviews so far revealed mixed success of digital tools for exercising [32] and encouraging early results for virtual reality applications [33] in people with LBP. The present literature has been described as heterogeneous with respect to the interventions [32] and methods [33]. Technical progress is considered as a factor to possibly impact intervention effectiveness, although so far, no clear difference for later studies was observed [32]. Thus, consistent research effort is necessary. Digital tools can be designed to influence the way movements are performed during exercises, for instance, to vary the speed [34]. Investigated tools include exergames readily available in the market; for example, exercising with the well-known Wii balance board (Nintendo of America, Inc) [35,36] or the Valedo Motion (Hocoma AG) [6,11]. Sensor technology has been used to intervene on movement characteristics of the trunk directly; for example, in studies [37,38] where warning participants from performing extreme back movements during everyday work was investigated. Other interventions have specifically encouraged movement of the lumbar spine in an exercising context [9,11,22,39] or are otherwise dedicated to providing feedback on lumbar spine movement [9-11,22]. For different kinds of tools used, only a small amount of research has been conducted [32]. Therefore, such digitally supported training modalities should further be investigated. Different systems and technological setups have been explored, for instance, cameras [40], wearable sensors [6,41], and sensors readily available in mobile phones [42], sometimes in combination with virtual reality headsets [41,42]. However, only few studies provide first insights in the effects of these interventions on movement quality in people with LBP [9-11]. In all, 2 studies suggested such interventions might have positively affected trunk ROM, but in one study, it remained ambiguous whether there was a significant difference to the standard care control group [9] and in the other study only a single group was investigated [22]. In a third study no effect on ROM was found [10]. Motor control impairment was not different in a study, where patients in the intervention group received access to additional exercises with sensor-based feedback other than the control group [11]. To our knowledge, the effect of such exercises on postural balance in people with LBP has not yet been investigated.

The primary aim of this study was to examine whether exercising with feedback on trunk movements can enhance postural balance, indicated by the change in anterior-posterior (AP) postural sway between the assessments before and after the intervention. Additional parameters to quantify postural balance were explored. As secondary outcomes, movement of the lumbar spine and hip during 2 different movement tasks and participant-reported outcomes were included. A further aim was to analyze adherence to the intervention.

This manuscript was based on a study protocol [44] that included a 2-arm randomized controlled trial. Figure 1 shows the assessment and intervention schedule. The study took place at University Hospital Zurich between May 2019 and October 2020. Except for an extension of the study period of 3 Months to compensate for a pause due to the COVID-19 pandemic, the study was completed as planned, and interim analyses of intervention effects were not conducted. Outcomes were assessed twice at T1 and T2, before an intervention was given. Further assessments were taken after another 3-week period with a fixed exercising schedule for the intervention group (T3) and a subsequent 6-week period without specified exercising schedule (T4). Participants were randomized during the assessments at T2, and those assigned to the intervention group received an introduction to the exercising program right after the assessment. After T3, participants in the intervention group retained the Valedo Home exercising system (Hocoma AG), without being required to follow a specific schedule or to complete any exercises at all. This period was introduced to observe further adherence to the exercising program, without commitment to a schedule provided by a therapist or to complete a schedule for research purposes. Participants who were randomized to the control group did not receive a sham intervention.

Block randomization (blocks of 2 and 4), stratification by body height, and 1:1 allocation were implemented through the randomization tool in REDCap (Research Electronic Data Capture; Vanderbilt University) [45] hosted at Eidgenössische Technische Hochschule Zurich. AM generated random sequences with the dedicated R package blockrand (version 1.3; [49]) and randomized the participants using REDCap. The staff conducting assessments of the outcomes was blinded, and randomization occurred as late as possible (at T2) to reduce the risk of accidental unblinding.

The published study protocol [44] contained a further research question involving an additional patient group. We intended to compare the effect of the intervention between a group of patients and other participants who did not receive other treatments than the exercise intervention with postural feedback. In this manuscript, we report only the research questions that could be investigated based on the collected data, as insufficient patients were enrolled at the study site.

Participants were recruited through different web-based bulletin boards, websites, distribution of flyers, and personal communication. Recruitment was completed 3 months before the planned end date, to allow all participants to finish in time. Eligibility was ascertained in an interview-like setting that allowed the participants to describe their situation. Participants were considered eligible if they provided informed consent, they were aged at least 18 years, reported to the investigator that they had nonspecific LBP, and did not receive therapy or medical treatment for LBP within the past 6 months. There was no questionnaire-based assessment or formalized cutoff scores for pain intensity during the eligibility check. If it was not possible to clarify the eligibility of a participant by asking standard questions alone, one of the trained physiotherapists of the team decided on the criterion in question. The criterion of no recent treatment was relaxed from 12 months to 6 months during the study to improve recruitment rates. Participants reporting specific LBP or radicular syndrome were excluded from participation. Participants were also excluded if they indicated to the investigator that they would not be able to complete the movements required by the exercise intervention owing to high pain. Other reasons for exclusion were pregnancy, taking medication that impairs postural balance, severely impaired vision, allergic reactions to adhesive strips, and insufficient proficiency in German or English.

Participants were not compensated and provided informed consent in writing before any study procedure was started. The trial was approved by the Cantonal Ethics Committee Zurich (BASEC: 2018-02132) and registered in ClinicalTrials.gov (NCT04364243).

The intervention is described in Textbox 1. Multimedia Appendix 3 contains a video on the exercises used in this study (published with permission from Hocoma AG).

Data preparation and analysis were conducted in MATLAB R2018a (The MathWorks, Inc) and R (version 4.0.4; R Foundation for Statistical Computing). The simultaneously recorded data of the force platform and the IMUs were time-synchronized based on aligning a sideways leaning movement of the participants, shifting their weight to the left and the right, which was performed before and after each repetition. For this time synchronization, the movement had to be clearly distinct from the tasks and identifiable in both sources of data. This parallel recording was important for comparing force plate and IMU data. The beginning and the end of each balance and movement task were defined based on marker time stamps set in the IMU data during the assessment. The markers were inspected visually and corrected by hand before further analysis, as placement during the assessment was sometimes not optimal and occurred too early during the time-synchronization movement or too late during the task.

To assess the equivalence among the treatment groups at study entry (T1), participant characteristics were compared. Welch t test (2-tailed) or alternatively Wilcoxon rank-sum tests were used, if the data appeared to be not normally distributed based on Normal QQ plots or Shapiro-Wilk tests within groups. Dependent group t tests (2-tailed) were used to test whether change had occurred between T1 and T2, or if the assumptions were not met, Yuen tests, as provided by the R package WRS2 [64], were used. The hypotheses regarding the intervention effects were tested by comparing the change of the respective outcome (Δ outcome: T3-T2) between the intervention and the control group, predicting the more favorable outcome for the intervention group. These comparisons were performed each as intention-to-treat (ITT) and per-protocol (PP) analyses. In the ITT analyses, all participants who had been randomized at T2 were included. Missing values at T2 and T3 were replaced with the mean of the previous 2 assessments (T1 and T2) of the participant. For the PP analyses, participants who either had incomplete data or had been randomized to the intervention group but exercised <1 hour, between T2 and T3, were excluded. Comparisons were performed using independent group t tests (1-tailed), when the data were normally distributed according to Shapiro-Wilk tests, and a Levene test did not show heterogeneity of variances. Otherwise, Wilcoxon rank-sum tests were used. The a priori power analysis is reported within the study protocol [44]. We calculated post hoc power for the primary comparison using G*Power (Heinrich Heine Universität Düsseldorf) [65]. The effect size r observed for the ITT comparison of the primary outcome was converted to d using a web-based tool [66]. Power was calculated for a directed Wilcoxon rank-sum test with an α level set to .05 and the distribution menu set to normal.

Additional exploratory analyses to compare the absolute scores across all assessment visits including the second intervention period were conducted using mixed 2-way ANOVA. Only participants who completed the study (n=20) were included in these analyses. Missing data were replaced by mean scores of the previous assessments of that participant. Generalized η² was used as effect size [48], and calculations were made using the r package rstatix (version 0.7.0; [50]). Shapiro-Wilk tests and Levene tests were used to test the assumptions of normality and homogeneity of variances. If the data did not fulfill the assumptions of normality and homogeneity of variances, different data transformations were explored. In cases where no suitable transformation was found, Friedman ANOVA was conducted across the assessment visits for each group separately, and group differences were compared at each assessment visit using the Bonferroni-corrected Wilcoxon rank-sum tests. Data on adherence were analyzed using descriptive statistics and graphs.

On the basis of visual inspection, orientation data from the IMU sensors were corrected in 2 cases where axes were flipped (15 trials of 2 participants). The data from a participant at T1 and another participant in the control group at T2 were discarded, because misplacement of the sensors was suspected. For a participant in the intervention group, no data for the T3 assessment were available, as the sensors had not been sufficiently charged. In the ITT analysis, all randomized participants were analyzed and missing values were replaced as described in the Methods section. For the ITT analysis of the balance and questionnaire data, 6 replacements (5 control and 1 intervention) were made for the T3 assessments. For the movement tasks, 1 replacement in the control group was made for the T2 assessments, and 7 replacements (5 control and 2 intervention) were made for the T3 assessments. For the PP analysis, participants for whom replacements had to be made were removed from the analysis. In addition, data from the participants in the intervention group (3 balance and questionnaires; 2 movement tasks), who had exercised <1 hour within the 3-week period and were excluded from the PP analysis, if this was not already the case when the participant was also a dropout, or the data had already been removed owing to insufficiently charged sensors.

During data analysis, it was discovered that some of the items of the WHOQOL-Bref at T3 in the German version had been collected with response options ranging from 1 to 4 instead of 1 to 5. Data collected with the affected items (items 3-9) were discarded for all assessment visits, and the scores of the scales were calculated without those items.

As not all data fulfilled the requirements for the 2-way mixed ANOVA for the analysis across all 4 assessment visits, the data were transformed where necessary. Transformations applied are reported in Table 1.

The effective duration between T1 and T2 had a median of 21 days (IQR 5; minimum 17, maximum 97); between T2 and T3, 23 days (IQR 3; minimum 19, maximum 36); and between T3 and T4, 44 days (IQR 7.75; minimum 38, maximum 112). For a participant, the time span between T1 and T2 was extended to 97 days, and for 2 participants, the period between T3 and T4 was extended to 112 and 99 days, respectively, because of an interruption in the study owing to the COVID-19 pandemic. The period between T2 and T3 was not affected by extensions due to the COVID-19 pandemic.

As presented in Figure 3, a total of 93 participants made an initial contact and requested information regarding the study. Of those 93 participants, 38 (41%) provided written informed consent. At T1, a total of 32 participants, without recent treatment for LBP, were eligible for the study. In all, of 32 participants, 5 (16%) dropped out before randomization at T2 (n=27, 84%). Tables 2 and 3 show the participant characteristics at baseline. Between the participants randomized to the intervention and the control group, there were no significant differences in any outcome measure at T1 or T2, but analyses of change between T1 and T2 revealed that there was a significant reduction in pain intensity across participants who had been randomized (T1: median 3.00; mean 3.26, SD 1.56 and T2: median 2.00; mean 2.59, SD 1.34). Descriptive statistics on all outcomes at all assessment visits are reported in Table S1 in Multimedia Appendix 4 and comparisons of outcomes at T1 and T2 are shown in Table S2 in Multimedia Appendix 4.

Exploratory analyses were conducted across all 4 assessment visits among a subset of participants who remained in the study until T4 (n=20).

Participants in the intervention group were instructed to complete a fixed set of 90 exercises between the assessments T2 and T3. Of these exercises, a median of 61% (55/90; range 2%-99%) were completed. As not all exercises were performed with the specified duration and frequency, and some participants performed the exercises that were provided from the device but were not intended as part of the program, effective time spent exercising differed from the completion of the program. In this period with a predefined schedule (T2-T3), participants exercised a median of 77.2% (139/180; range 3%-202%) of the targeted exercising duration of 180 minutes. The exercising time of 4 participants exceeded 180 minutes. During the intervention period with a schedule, a total of 7 participants performed a median of 9 exercises (minimum 1, maximum 41), equivalent to 17 minutes (minimum 2, maximum 109) that were not part of the program. In the intervention period without a schedule, of the 11 participants who had remained in the study, 4 (36%) participants performed a median of 27 exercises (minimum 1, maximum 29), equivalent to 82 minutes (minimum 2, maximum 101). An overview of the number of any exercise performed is provided in Figure 6.

There were no unintended effects that were related to the intervention. Although reasons for not adhering to the protocol were not assessed systematically and participants had been encouraged to contact the investigators with any difficulties, some participants in the intervention group reported problems with handling the devices. This included difficulties such as finding the right icon on the tablet and difficulties with the calibration of the IMUs and program failures of the tablet. These issues likely contributed to the low adherence of some participants.

Self-directed home exercising with feedback on trunk movements for a period of approximately 3 weeks did not enhance postural balance during quiet standing in study participants with LBP or significantly affect any other of the investigated outcomes. Comparisons of the groups with respect to the movement tasks showed, descriptively, a tendency toward slightly increased motion of the lumbar spine during both tasks in the intervention group, combined with a small reduction in the control group, which contradicted our predictions. Adherence to the scheduled exercising program was low. After the participants were no longer provided with a schedule to complete, only some participants kept using the training device repeatedly without specific instructions. Despite not showing intervention effects in this trial, it cannot be excluded that these interventions may still be beneficial when integrated into a therapy setting with patients. Furthermore, for other exercising interventions, it has been demonstrated that exercising could have more pronounced effects in patients than in other study participants with LBP [3]. A review showed that the results were positive for exercising with digital systems for LBP, when these exercises were delivered together with another intervention, but otherwise not [32].

A particular strength of this work is the combination of an investigation of the exercising program at home in a period with a set training schedule and in a second interval, where participants could exercise as they wished. Comparison with studies on related interventions in home-based settings, which are considered similar, are difficult, as in a study a combined value including other exercises was investigated [11], and in another study, self-report methods had failed [6]. Furthermore, in a study that investigated exercising with the Nintendo Wii, completion of 71% of the advised time was achieved [36], which is comparable with the median of 77% obtained in this study. Nevertheless, the schedule provided was much more demanding and additional measures were used to improve adherence in the other study [36]. The comparatively short 3-week exercise period in combination with the low adherence may have limited the effects of the intervention. This assumption is in line with the findings of a review on virtual reality interventions, which indicated that interventions with more sessions may be more successful [33].

Results on adherence, considering time spent exercising, was more favorable than the number of exercises performed as requested by the investigators. Some participants exercised even more than required but did not follow the instructions precisely. In some cases, participants may have forgotten to reset the play time from the default 4 to 2 minutes, or the game may have motivated the participants to explore additional contents and may have provided stronger guidance than the instructions from the investigators. The 6-week period with flexible exercising opportunity resembled more closely to the conditions under which participants would be using such tools without the connection to a therapeutic setting. Only few participants kept exercising after T3. These results may imply that such interventions only get adopted by a small number of people or might rather be integrated within supervised programs on site. Within the setup of this study, it could not be determined whether the provided schedule or the participants’ commitment to comply with the study protocol resulted in higher amounts of exercises between T2 and T3. Future studies should investigate if and how automated scheduling options can help improve adherence and how they should be integrated. Although the Valedo app offers the option to generate an exercising plan, such functions could be placed more prominently. In addition, the context and the kind of assistance required should receive more attention. Recently, for example, blended therapy setups for people with LBP have been explored [75].

The low number of participants who could be recruited is an important limitation of this study. Although the individual components of the study protocol may not have been too time consuming, the overall effort associated with study participation, including diary methods and activity tracking not discussed in this manuscript, may have been a cause for low recruitment and retention rates. These assessments were included to answer additional research questions beyond the scope of a single manuscript but contributed to the effort by the study participants. In line with this presumption, reasons given for withdrawal were frequently related to time investment or perceived benefit and effort. This might also have contributed to the low adherence to the intervention. Time intervals between assessments were slightly stretched owing to frequent requests from participants to reschedule appointments, as the study participation was not part of an official treatment program and therefore often had to take place often outside of the working hours of the participants. The assessment of the movement tasks was preceded and followed by the participants shifting their weight to the sides and back, to time-synchronize the data from the IMUs with data collected simultaneously from the force platform. Although supporting analyses of change between T2 and T3, where data from trials that appeared to be performed from an unstable starting position were removed, appeared similar, this setup could have influenced the results. The study participants could not be blinded, and with most assessments conducted in the field, it could not be ruled out that the participants completed all exercises themselves. In a case with a particularly high number of exercises, it was suspected that other people may have completed some of those exercises. We did not record the use of pain medication during the study; therefore, confounding effects of pain medication could not be ruled out. Further, although we consider the availability of the questionnaires in different languages as a strength, this setup may have caused inconsistencies between the questionnaires that different participants received.

The results obtained in this study indicate that exercising with feedback on trunk movements alone may not influence postural balance during quiet standing in people with only moderate LBP intensity and disability. No significant intervention effects on lumbar spine and hip movement, pain intensity, disability, QOL domains, and fear of movement were observed. These results must be seen within the context of a small sample size and low adherence to the intervention, resulting in low doses of exercise. More work in this field is required; for example, to establish the effect of interventions using feedback on trunk movements in people more severely affected by LBP and clarify more proximal effects on trunk movement properties. As the amount of exercising dropped substantially in the intervention period without a schedule, future studies should investigate the impact of different scheduling options and explore such interventions in combination with other therapeutic settings or other strategies to improve adherence.