JMIR Publications

ABSTRACT

Background:

Single-camera markerless pose estimation may support accessible knee movement assessment, but performance can depend on task, joint angle, and detection quality, and evidence across complementary measurement properties remains limited.

Objective:

To evaluate agreement of MediaPipe BlazePose Heavy knee-flexion estimates with goniometry for static positions and with Kinovea 2D video annotation for dynamic tasks, to assess inter-rater agreement of the Kinovea annotations, and to explore repeated-session performance in a small retest subsample.

Methods:

Fifteen healthy adults performed Squat, Sit-to-Stand, and bilateral Single-Leg Stance tasks during standardized lateral-view smartphone recording. Static knee flexion was assessed at four goniometer-set targets (30°, 60°, 90°, and 120°). Native hip–knee–ankle included angles were converted to clinical knee flexion (0° at full extension). Agreement was assessed using the intraclass correlation coefficient (ICC), specifically ICC(2,1), signed bias and 95% limits of agreement, mean absolute error (MAE), and root mean square error (RMSE). Dynamic primary estimates were conditional on BlazePose target-frame confidence ≥0.3; low-confidence observations were also reported as detection failures and included in an all-output stress test. Four participants returned after 21–29 days (median 25 days) for exploratory repeated-session analyses.

Results:

Of 210 dynamic target-frame comparisons, 33 (15.7%) failed the confidence criterion. Conditional task-specific ICCs were 0.773 (95% CI 0.225–0.910; 71/75 retained) for Squat, 0.632 (95% CI −0.081 to 0.872; 58/75 retained) for Sit-to-Stand, and 0.136 (95% CI −0.150 to 0.401; 48/60 retained) for combined Single-Leg Stance. In the all-output stress test, Squat ICC decreased to 0.428 and RMSE increased from 8.78° to 25.43°. Static agreement was ICC=0.874 (95% CI 0.834–0.904), with RMSE=15.49° and angle-dependent signed error. Inter-rater agreement for the Kinovea 2D annotations was high overall (ICC=0.999, 95% CI 0.997–0.999), although this does not establish absolute 3D accuracy. In the exploratory four-participant analysis, the ICC of the BlazePose-minus-Kinovea difference across sessions was 0.094 (95% CI −0.225 to 0.405), which was too imprecise to establish a stable fixed offset.

Conclusions:

Agreement between single-camera BlazePose and the study’s references depended on task, flexion angle, and detection quality. The findings support further feasibility and screening-level evaluation under controlled acquisition with per-recording quality checks; they do not support replacement of clinical goniometry, precise longitudinal decision-making, absolute 3D kinematic accuracy, or generalization to clinical populations or other smartphones. Clinical Trial: Not applicable.