576 FO27 – Evaluation of a machine learning models to assess postural orientation errors during a single-leg squat

Background Measures of postural orientation (i.e., the ability to align body segments in relation to each other) are needed because undesirable postural orientation is a potential risk factor for lower extremity injury. Clinical and biomechanical assessment of postural orientation is time-consuming, and thus automating this process, e.g., with artificial intelligence, may streamline the process. Objectives To evaluate the ability of machine learning (ML) models to accurately classify postural orientation during the single-leg squat (SLS). Design Methodological study Settings Clinical setting Participants Videos (n=147) of 30 participants performing the SLS in the frontal plane. Several different cohorts (healthy (n=15), patients with hip/knee injury (n=15)) were included for variability. Intervention Postural Orientation Errors (POEs) of the lower extremity and trunk were assessed from videos and scored as good, fair, or poor. Three experts (physical therapists) visually assessed each repetition for each POE independently and then they reached consensus for each repetition and POE. ML-models, one for each POE, was trained to assess POEs using supervised learning on a subset of videos from the consensus assessment (n=48). The remaining 99 videos from the consensus assessment were used to evaluate the validity of the ML-models by comparing the prediction of ML-models with the consensus scores, using quadratic weighted kappa, (κ), Spearman's correlation coefficient (rs), and accuracy. Main Outcome Measures Postural Orientation Error (POE) scores Results The agreement between the AI-models and the consensus score showed moderate to substantial agreement for knee and thigh (κ=0.58–0.7, rs=0.566–0.702), fair agreement for pelvis and trunk (κ=0.28–0.4, rs=0.324–0.432), and no agreement for foot (κ=-0.042, rs=-0.05). ML-models predicted the expert consensus score in 53–78% of the cases. Conclusions The use of ML-models as a fast and comprehensive assessment of POEs during the SLS shows promising results. Modifications to some ML-models may lead to improvements in model performance.