Expert-derived Biomechanical Features Improve Cross-Dataset Generalisation in Sit-to-Stand Movement Classification

Abstract

Providing feedback for unsupervised physiotherapy by classifying erroneous motions can play a vital role in improving rehabilitation for motion-impaired patients. Inertial Measurement Unit (IMU) based rehabilitation classifiers show strong in-domain performance but rarely generalize beyond their training dataset, limiting real-world clinical applicability. Biomechanically grounded feature representations offer a promising path toward more interpretable and transferable models. We test this principle on sit-to-stand compensatory movement classification, deriving 16 features directly from expert rule-based decision logic. A Decision Tree trained on these features was used to verify that the learned boundaries closely resemble expert-defined clinical thresholds. A Random Forest over the same feature space achieved 68.7% zero-shot and 79.3% one-shot cross-dataset accuracy, improving generalisability over generic-feature baselines and providing initial evidence for the value of biomechanically grounded features in STS rehabilitation assessment.