From Anomaly Scores to Biomechanical Decision Support: An Evidential Audit in Rehabilitation Motion Analysis

Rehabilitation motion analysis often uses anomaly scores to flag trials that differ from typical movement. The practical problem is that a high score can reflect compensation, personal style, one poor repetition, or sensor noise. This study asked whether an unusual pattern in one task is repeated when the same person performs another task, in order to decide whether that anomaly should be prioritized for meaningful biomechanical review. We analyzed 2,589 trials from a public Kinect rehabilitation dataset with five task contexts. First, each task was scored against its own reference: chair trials against other chair trials, stand trials against other stand trials, and so on. We then focused on the 11 participants who had both chair and stand recordings. The audit asked whether the same type of unusual movement appeared in both tasks for the same person more strongly than it did after randomly pairing different people. Using only average trunk movement, the audit did not find chair-stand anomalies that deserved closer interpretation. When the analysis kept where in the body the unusual movement occurred, through body regions, joints, or simple biomechanical measures, it identified stronger candidates for meaningful biomechanical review. A second public rehabilitation dataset was used as an external check. It showed that the audit could detect repeated anomalies when the same task was repeated, while accepting cross-task support only selectively rather than automatically. The practical value is to help decide whether a trial identified as anomalous has enough repeated evidence to justify explanation, or should be documented without explanation until more evidence is available.