How do visual and smartphone camera-based shoulder ranges of motion compare?

Abstract

Background: Objective assessment of functional shoulder range of motion (ROM) is crucial for evaluating shoulder interventions and guiding rehabilitation. The goniometer is the clinical standard, but due to practicality, visual estimation is often used despite its lower reliability. Recently, smartphone video-based assessment using two-dimensional (2D) pose estimation models has emerged as a potential objective alternative. This study aimed to compare 2D-pose-based ROM assessment with visual estimation and examine inter-observer agreement.

Methods: Seventeen individuals (8 females, 9 males) with normal, pain-free shoulder function were assessed for active ROM in abduction, flexion, extension, external rotation (ER) in two positions (ERI & ERII), and functional internal rotation (FIR). 2D videos from three smartphones were used to estimate shoulder ROM, while two othopedic surgeons visually estimated ROM. For each movement, participants performed six repetitions, three at maximum and three less than maximum ROM (self-selected). Mixed effects models assessed the relationship between 2D-pose-based and visual-based ROM, with visual observer as fixed factor and visual estimates × observer interaction. The coefficient of determination (R²) from these mixed effects models assessed consistency, and smallest detectable difference was used to determine agreement.

Results: Consistency between 2D-pose and visual estimates was excellent for abduction (R² = 0.99), flexion (R² = 0.95), and ERII (R² = 0.86), good for extension (R² = 0.69) and ERI (R² = 0.73), and fair for FIR (R² = 0.52). Smallest detectable difference values ranged from 4.4° to 7.9°. Agreement varied by movement type and observer, with significant visual estimates × observer (P < .003) interaction effects for abduction and flexion: both observers reported higher ROM values than 2D-pose near end-ROM (∼3-4°) with observer 2 reporting lower values than observer 1 (∼15°) at smaller ROM (<60°). 2D-pose estimates were higher (∼20°) for extension at low ROM (<45°) than visual estimates. 2D-pose estimates were lower (∼30°) for ERI at high ROM (>45°) than visual estimates. Visual observers agreed on extension, ERI, ERII, and FIR estimates but disagreed on abduction and flexion estimates.

Conclusion: 2D-pose-based estimates of shoulder ROM were consistent with visual estimates for most movements, though discrepancies existed at specific ROM levels and between observers. The higher resolution estimates of 2D-pose suggests it could reduce observer variation, making it a promising alternative for clinical and research settings. However, further refinement is needed for movements like ERI and FIR using both methods. These findings highlight the importance of method consistency in assessing shoulder ROM and the potential benefits of automated methods for more consistent evaluations.

Keywords: Pose estimation; Range of motion assessment; Shoulder; Smartphone; Thoracohumeral; Validity.

Figure 1

Shoulder movements. From left to right: abduction, flexion, extension, ERI, and ERII. Participant positioning relative to the camera is shown: facing the camera for abduction and ERI; side view for flexion, extension, and ERII. In ERI, participants kept the elbow at 90° flexion and maintained elbow–torso contact while externally rotating; In ERII, participants held 90° shoulder abduction and elbow flexion. ERI angle was computed using inverse sine based on ratio of horizontal wrist–elbow projection over contralateral wrist–elbow length. ERII angle was calculated using inverse tangent relative to the global horizontal. All other shoulder angles were calculated using inverse tangent between relevant segments (shown in red). Note that the public version of the app does not display these landmarks or angle values. ERI, external rotation in position I; ERII, external rotation in position II.

Figure 2

Functional internal rotation. Definition of zones for functional internal rotation measured using 2D-pose–based wrist position are shown. Quadrants increase gradually with higher functional internal rotation range as measured by the Range of Motion Models in the mymobility App. 2D, two dimensional.

Figure 3

Measurement setup. Upper limb movements were concurrently recorded using 3 iPhones (mounted on a board) and viewed by two surgeons (visual observers).

Figure 4

Comparisons between 2D-pose and visual estimation of active shoulder ROM. The linear relationship and the 95% CI of the model (see statistical analysis for more detail) between 2D-pose and visual estimation of each observer is provided (blue = observer #1, red = observer #2), and individual data points (scatter) for abduction, flexion, extension, ERI ERII, and FIR. The diagonal dashed lines indicate perfect agreement (y = x). The BA plots show the difference (bias) (Delta = 2D-pose-visual) vs. visual, and the 95% CI of the model is displayed below each linear relationship between 2D-pose and visual plot and represents the LoA based on the 95% CI from the linear mixed model. The horizontal dashed line indicates perfect agreement (Delta = 0). ROM, range of motion; CI, confidence interval; ERI, external rotation in position I; ERII, external rotation in position II; FIR, functional internal rotation; BA, Bland–Altman; LoA, limits of agreement.

Figure 5

Agreement between the two observers for abduction (top panel) and flexion (bottom panel). Graphs reflect the mean difference between observers with 95% CIs (shaded area) derived from the mixed model (see statistical analysis for more detail). Observers agree when CI includes zero, indicated by the horizontal dashed line. CIs, confidence intervals.