A depth camera used to measure movement after rotator cuff surgery agreed reasonably well with an inertial motion sensor for some actions, but it overstated how far patients could move their shoulders in two postoperative tests. The findings suggest the camera could help with broad screening or tracking recovery, yet should not be treated as a stand-alone precision measure for surgical outcomes.
The prospective observational study followed 12 patients with unilateral full-thickness rotator cuff tears smaller than 3 centimetres through preoperative and postoperative assessments. Four patients were male and eight were female, and their mean age was 54.1 years. Azure Kinect was compared with an inertial measurement unit, or IMU, recorded at the same time as the reference system.
Accuracy depended on the movement
Participants completed six simulated daily-living tasks and three shoulder range-of-motion tests on the affected side. Reach tasks used a standard 1-kilogram load, each task was done at a comfortable self-selected speed, and every task was repeated five times.
The clearest agreement came in the sagittal plane, which describes forward-and-back movement. Shoulder flexion showed intraclass correlation values from 0.708 to 0.907, while elbow flexion ranged from 0.804 to 0.846. The study also reported good-to-excellent similarity in the shape of several movement waveforms, meaning the two systems often followed the same pattern over the course of an action.
Performance fell for coronal and horizontal-plane movements, which include motion across the body and rotation around it. These measurements showed lower agreement and larger errors, and statistical tests found proportional bias in most joint-plane and task combinations. In practical terms, the size of the disagreement changed with the amount of movement being measured. The reported standard error of measurement for sagittal-plane movements was below 15 degrees, but the non-sagittal biases were larger.
The postoperative gap was hard to ignore
For three-dimensional shoulder range of motion, the systems showed reported agreement values of 0.837 for flexion, 0.729 for abduction and 0.735 for internal and external rotation. Before surgery, the differences between Azure Kinect and the IMU were not statistically significant for any of those measurements.
After surgery, however, Azure Kinect recorded substantially more flexion than the IMU: an average of 137.9 degrees compared with 97.1 degrees. It also recorded more external rotation, 48.6 degrees compared with 31.0 degrees. The reported comparisons were statistically significant, with p values of 0.002 for flexion and 0.038 for external rotation.
Both systems detected a significant change in flexion between the preoperative and postoperative assessments. Neither system detected a change in abduction or rotation. That means the camera broadly tracked the direction of change seen by the reference system, even though its postoperative absolute readings were higher for flexion and external rotation.
The authors noted that the average range-of-motion discrepancy exceeded a 10.4-degree benchmark for the minimal clinically important difference, or MCID. They also cautioned that this benchmark came from a composite shoulder score rather than a range-of-motion-only threshold, so applying it to measurement error is approximate.
Useful aid, not a replacement
The study was exploratory and preliminary. With only 12 participants, its estimates were imprecise, and several agreement measures had wide confidence intervals. The authors said this reduced precision, statistical power and generalizability, and called for larger multicentre studies with repeated sessions and independent cohorts.
The IMU was used as the reference even though its accuracy can vary by movement plane, making it an imperfect comparator for judging the camera. The study also examined only one postoperative time point, so later validity remains unknown.
The authors concluded that Azure Kinect has task-dependent validity. Its stronger performance for sagittal-plane movement may make it useful for screening or for following broad recovery trends, but the evidence does not support using it alone for precise surgical outcome assessment. Larger studies with repeated sessions, independent cohorts and stronger reference comparisons are needed.
Paper data and sources
Original title: Validation of Azure Kinect for upper limb kinematic assessment after arthroscopic rotator cuff repair: task-dependent accuracy and postoperative overestimation
Authors: Chen Y, Zhan C, Zhang L et al.
Journal/Repository: Frontiers in bioengineering and biotechnology
Status: Peer-reviewed
First online: 2026-10-07
DOI: Not available
Original paper · Full text