Personalised healthcare and exercise rehabilitation based on upper-limb metrics

Rapid and accurate upper-limb motion analysis and metrics are essential for enhancing exercise medicine knowledge graph to drive personalised medicine. However, current studies face limitations in understanding multi-timescale and multi-target, compounded by discrepancies in physiological structure. This study proposes Upper-Limb Dynamic Warping (UP-Ldw), which effectively responds to motion discrepancies and adapts to the physiological characteristics. UP-Ldw constructs a geometric model by parameterizing features and incorporating physiological structure parameters to accommodate individual variations. Dynamic temporal regularization is integrated to accommodate motion sequences across multiple time scales. Ultimately, the motion similarity among various targets is outputted to facilitate comparison and metrics. Furthermore, two datasets are developed: Upper-Limb 3-Dimensional Dataset (UP-L-3D), and Upper-Limb Geometric Modeling Dataset (UP-L-GM), both utilized for validation. Comparison experiments employed convolutional neural network (CNN), principal component analysis (PCA), and random forests. Results demonstrate that UP-Ldw achieves the highest accuracy of 97.92 % using metrics 20 as the discriminant criterion, with a short running time of 1–8 ms. UP-Ldw aligned with CNN confusion matrices and the PCA downscaling, validating its precise motion analysis. The Random Forest model attained an average accuracy of 91.1 %, confirming the validity of the geometric model. A generalization experiment was conducted using the public dataset Arm-CODA, further validating UP-Ldw's ability to adapt to physiological structures and effectively metricize upper-limb motion. Overall, UP-Ldw employs artificial intelligence to metricize motion, facilitating mirror rehabilitation. This advancement contributes significantly to the engineering applications of personalised healthcare and exercise rehabilitation.