Comprehensive Summary
Bai et al. researched the functionality of traditional rehabilitation exercise assessments and compared these models to more recent machine learning approaches, which utilize covariance matrices to calculate skeletal motion data. Universally, rehabilitation exercise assessment is an essential component in order to ensure that patients are correctly carrying out their prescribed recovery movements from their injuries or pre-existing medical conditions. However, it came to the researchers’ attention that the traditional assessment methods, which include rule and template based systems, especially have difficulty with complex motion variations and generalizations across larger populations of patients. In response to these specific limitations, the researchers proposed a new rehabilitation assessment framework, using covariance matrices with a Symmetric Positive Definite (SPD) manifold, which emphasizes these complex motions and spatial relationships with the patient to the environment. The framework specifically consists of K-Nearest Neighbors (KNN) on the SPD manifold, Tangent Space Linear SPD Support Vector Machine (SVM) with gradient descent (stochastic optimization), and neural networks which have multi scale feature extraction for vectorized SPD data on the manifold. The researchers utilized Riemannian geometry and Log Euclidean mapping to distinguish the correct and the incorrect movements in the model. The accuracies were reported to be 92.40%, 85.18%, and 87.59% across three different rehabilitation datasets. Covariance matrices were divided at three anatomical scales, including arms, legs, and the entire body. These three components were combined to form a block diagonal SPD matrix, displaying independent joint dynamics with the overall movement structure of the patient. The results had great accuracy in the model provided, and further research with the SPD manifold will be carried out.
Outcomes and Implications
In the study, the researchers proposed a new approach for assessing the rehabilitation exercises using their SPD manifold framework with the utilization of covariance matrices. Upon testing the SPD manifold framework across three datasets, the accuracy had improved significantly as more trials were performed. Considering the success of the model, the methodology presented could be used in a clinical setting. For example, the researchers may be able to improve patient outcomes by providing the clinicians with detailed feedback of the patients’ rehabilitation based on the covariance matrix results. This would ensure correct execution and performance of the rehabilitation exercises, which would also facilitate and reinforce patient recovery. In the future, the model can be integrated into mobile devices for more practical and accessible daily usage. Further optimization is necessary, as noted by the researchers, but this model has great potential to be implemented into a clinical environment to foster trust among clinicians and patients and to allow for accurate rehabilitation exercise metrics reading.