What if recovery from chronic pain could be tracked not only by asking how someone feels, but by measuring how their body moves in virtual reality? This study finds a promising movement signal—but also shows why it is not yet a finished biomarker.

Chronic low back pain is a major public health problem, and reliably measuring its effects on movement and activity—and changes due to treatment—is a healthcare challenge. A novel digital therapeutic was previously shown to reduce fear of movement and increase quality of life in adults with moderate to severe chronic low back pain.

The study therefore examines whether wearable-device data collected during that trial could be used as a digital measure for chronic low back pain studies. Measuring pain and its harmful effects across multiple areas of life remains a clinical and research challenge.

Pain is a private mental event that cannot be observed except by the person experiencing it. There is no objective measure of a subjective experience, but objectively defined correlates can be measured physiologically, through observation, or by capturing changes in gross or fine movement of the body or limbs.

Digital biomarkers in chronic pain can focus on cognitive, emotional, relational, or physical domains. Chronic low back pain is often characterized by avoiding movements believed to create further pain or re-injury. Capturing movement restriction and its possible restoration in treatment is a valid mechanistic target for a clinical endpoint in trials and clinical practice.

The analysis used the VIRPI pilot clinical trial of a digital therapeutic tested in adults with chronic low back pain, high disability, and high fear of movement and re-injury. The study was conducted fully remotely in twenty-twenty and twenty-twenty-one, and the VR intervention was effective in reducing fear of movement and improving subjectively reported quality of life compared with sham VR and treatment as usual.

Adults experienced about thirty daily treatment sessions, creating an opportunity to explore movement data collected by VR devices and two wearables during repetitive exposure to movement. The second goal was to identify movement behaviors that correlate with clinical data and consider their use as digital endpoints.

Because there were no a-priori hypotheses about which movements might matter, the analysis used a data-driven approach. The VIRPI study was registered on ClinicalTrials.gov as NCT zero four two two five eight eight four. It was a double-blind, three-arm, prospective, randomized controlled trial comparing a digital therapeutics software solution, a sham placebo comparator, and standard care.

Adults with chronic low back pain were recruited from the community between January and October twenty-twenty, screened, and randomly allocated to a study arm; participants and study personnel were blind to allocation. The digital therapeutic was a fully immersive VR experience with twenty-four modules, or individual intervention days, delivered over six to eight weeks.

It used a lakeshore environment, a virtual disembodied mentor providing behavior change content, and physiotherapeutically designed gamified tasks. The sham placebo used the same VR environment but contained no behavior change content, and participants were advised to relax there.

The sham intervention controlled for non-specific influences such as treatment expectations, duration of exposure, environment, and immersion. Digital therapeutics and sham participants received an Oculus Quest VR head-mounted device and two handheld controllers, while ten participants were allocated to standard care.

All thirty-nine participants received Empatica Embrace two and Garmin Vivosmart four wearables. Thirty-nine participants were randomized to the treatment arms; they included thirty-four women and five men, with an average age of fifty-four point seven years.

Thirty of thirty-nine participants had more than five years of low back pain, and mean pain intensity at trial start was two point eight on a zero-to-five scale. Mean Oswestry Disability Index was thirty-six point one, and mean Tampa Scale of Kinesiophobia was forty-one point eight, indicating high disability and a strong belief that movement would lead to further pain and reinjury.

The VR controllers recorded x, y, and z coordinates, with recording intervals of approximately zero point zero three seconds and some variation over time. The timepoints were combined into movement segments based on metadata. The controller coordinates were transformed into velocity in three-dimensional space, with velocity at each timepoint calculated as the change from the previous timepoint.

Linear regression lines were fitted to segment averages over study days, and the slope was assessed as the direction of change over time. A positive slope was considered an increase in velocity over time, while a negative slope was considered a decrease. The linear fit was not optimal, but it was kept for simplicity and to obtain a single measure describing the direction of change.

Because study days could span multiple calendar days and included planned pauses, the diurnal aspect was not naturally linear, so regression results were treated only as a proxy for natural within-day progression. The analysis correlated changes in velocity from head, left, and right controllers with changes in the Tampa Scale of Kinesiophobia, Overall Health visual analogue scale, and EQ-five-D-five-L quality-of-life score.

For controllers, the regression slope measured change, and the slope was multiplied by the participant’s number of VR study days to estimate movement change over the study. Clinical change was measured from baseline to the end of treatment after thirty study days.

For robustness, correlations were assessed using the full data and after removing a participant with two-thirds of study days missing, using both parametric and non-parametric correlations. Figure three shows boxplots of average segment velocity for the head-mounted device and the left- and right-hand controllers, separated into action and three no-action categories.

The black dots mark means, while the boxes and whiskers show the distributions with outliers removed. The authors use this visual comparison as a segmentation validity check: activity segments should reflect movement, whereas recordings also include preparation and end-of-day periods when movement continues without the target activity.

After the initial processing, only action segments were used to assess average velocity, and regression lines were fitted to average velocity during the study days. The velocity of the left-hand, right-hand, and head controllers for every participant is shown in Figures four through six.

For most participants, velocity appears to increase over the study, although this could have many causes; one suggested factor is that participants were moving better over time. Variance and obvious outliers remain, but most regression lines have an upward tendency, with a similar trend in left-hand and head-controller movements.

Figure four plots average left-hand-controller velocity during action-task segments across thirty study days for each DTxP arm participant, with black regression lines summarizing change over time. This matters because the authors use these patterns to examine progression, while noting that missing later data for participant 1054 could introduce bias.

Figure five plots average right-hand controller velocity across action-task segments over thirty daily sessions for each DTxP participant. Each blue point represents a segment, while the black line summarizes its change over study days; most panels show a modest upward slope, although participant one thousand forty shows a slight decline and participant one thousand fifty-four has a shorter observation range.

The broad scatter highlights substantial within-participant variability, making velocity a useful exploratory measure alongside the reported clinical correlations. Figure six shows average head-controller velocity for action-task segments across thirty daily sessions, with separate panels for the DTxP arm participants.

Each blue point represents a session-level observation, while the black regression line summarizes how velocity changes over study days. The lines generally slope upward, although the scatter varies across participants, making this visual useful for examining changes in movement speed over the study period.

For the Tampa Scale of Kinesiophobia and EQ-five-D-five-L quality-of-life score, increasing movement velocity correlated with a negative change; for Overall Health visual analogue scale change, the correlation was positive. This indicates that increased velocity correlated with improving clinical measures, and the association was similar for parametric and non-parametric correlations.

When the full data were examined, correlations between controllers were strong, but correlations between controllers and clinical results were weaker than in the cleaned data. Even so, Spearman correlations between the head device and clinical measurements remained at a moderate level.

Table one lists participant-level changes in clinical assessments alongside estimated velocity changes for the head, left, and right controllers. The controller values are calculated as the daily slope multiplied by study days in virtual reality; for example, participant one thousand sixteen has a right-controller change of zero point zero nine five meters per second.

This matters because it places movement-derived measures and changes in kinesiophobia, overall-health VAS, and EQ-5D-5L quality-of-life scores side by side for correlation analysis. Electrodermal-activity peaks per minute were compared between the digital therapeutics and sham arms by study day.

There were more peaks per minute early in the study overall, followed by some higher distributions in the digital therapeutics arm closer to the end. However, the association was not clear. Figure seven shows boxplots of electrodermal activity peaks per minute for each participant, arranged from lower to higher median values.

The blue and white boxes distinguish the DTxP and Sham arms, while the box heights and outlying points show substantial variation within and between participants. The authors note that the DTxP arm tends toward higher medians with greater variability, whereas the Sham arm more often shows narrower, lower-median distributions—important context given that the electrodermal data were considered noisy and unstable between sessions.

Movement data came from twelve adults with chronic low back pain, disability, and fear of pain and reinjury who completed at least thirty VR rehabilitation sessions. It proved possible to manage large datasets from multiple sources collected across participants and across time, but not all data sources were useful.

Electrodermal data were noisy and unstable between sessions, making them unsuitable for further analyses with the limited study population, while heart-rate data lacked precision because they were aggregated over a daytime. Positive correlations between velocity and self-report clinical outcomes are only the first step in establishing a movement-based digital biomarker.

Replication is necessary, followed by extension to different interventions and different clinical groups. The interpretation of velocity needs further consideration because movement-data segmentation was not based on specifically known movements, so it is unclear which movements show the change best.

More precise task metadata is needed, and movement data may support measures beyond velocity, including multidimensional descriptors enabled by machine learning methods. Across repeated VR rehabilitation sessions, movement velocity generally increased and was associated with better clinical measures, especially for head and right-hand sensors.

But the small sample and noisy non-movement data mean replication is essential.