Assessment of dynamic knee angle deviations in the frontal plane in physiotherapy clinical practice: intra- and inter-rater reliability of an application and agreement with two AI-models

BACKGROUND: Functional leg axis deviations, particularly dynamic knee valgus (DKV), are considered key risk factors for lower-extremity injuries. In physiotherapy practice there are no standardized tools to quantitatively assess functional leg axis deviations within a treatment session. While application-based 2D analyses offer accessible options, the agreement between manual and automated, AI-based Human Pose Estimation (HPE) methods remains underexplored. Here, the aims were to investigate the intra- and inter-rater reliability of a manual application-based method (PhysioMaster®); to quantify the agreement between two AI-based HPE models (OpenPose, BlazePose); and to explore which functional test, Single-Leg Squat (SLS), Single-Leg Hop for Distance (SLH), or Single-Leg Landing (SLL), elicits the most pronounced leg axis deviations. METHODS: Sixteen healthy adults (8 females) performed three standardized single-leg tasks (SLS, SLH, SLL) with each leg. Knee angles in the frontal plane were measured from monocular video images using the PhysioMaster® application (two raters × two ratings) and compared with automated 2D analyses using OpenPose and BlazePose. Reliability and agreement were evaluated using Intraclass Correlation Coefficients, Lin's Concordance Correlation Coefficient (CCC), and Bland-Altman analyses. Linear Mixed Model (LMM) analyses were carried out to explore which functional task would be the most effective. RESULTS: The PhysioMaster® application showed excellent intra- and inter-rater reliability across all tests (ICC ≥ 0.95). OpenPose demonstrated excellent agreement with manual application-based measurements (CCC = 0.95; Bias = 0.87°; LoA = -4.32° to 6.05°), while BlazePose achieved good agreement (CCC = 0.89; Bias = -0.67°; LoA = -8.76° to 7.42°). LMM analyses revealed that all three tasks significantly evoked knee angle deviations. The SLL evoked the largest valgus angles (Estimate = -8.29°, p < .001), while the SLS elicited the highest varus angles (Estimate = 7.92°, p < .001). CONCLUSION: Both, manual application-based and AI-based 2D methods, provided reliable and largely consistent assessments of knee valgus angles in the frontal plane during functional testing. OpenPose achieved the best agreement with manual application-based reference values, and the SLL was the most suitable task for detecting DKV. While the results support the integration of manual application-based and AI-assisted posture analysis into clinical physiotherapy, further research on concurrent validity of all three methods used against 3D motion capture is needed.