Quantitative Taxonomy of Hand Kinematics Based on Long Short-Term Memory Neural Network

Abstract

Assessment of hand motor function is crucial to stroke patients. However, the commonly used Fugl-Meyer assessment (FMA) scale requires 11 hand and wrist movements. To simplify these movements, this study proposes a hand motion classification framework based on deep learning to achieve the quantitative taxonomy of hand kinematics. To the best of our knowledge, this is the first study to use deep learning for the quantitative taxonomy of the hand kinematics. First, we use the Long Short-Term Memory (LSTM) neural network to extract deep features of 20 hand movements (including 11 FMA movements) from 37 healthy subjects, and rank the LSTM neural network output value (predicted probability) of each sample. The similarity between the movements obtained by the nonlinear transformation can be used to draw the confusion matrix. Then the confusion matrix is taken as the category feature to obtain the clustering dendrogram to show the similarity between different hand movements intuitively. Next, the 20 hand movements are divided into four groups by hierarchical clustering. The silhouette coefficient of the clustering results is 0.81, which is close to the ideal value of 1, indicating the validity of the clustering result. Finally, the clustering center is calculated to find the corresponding movement as the representative movement for motor function assessment. As a result, we reduced the 20 movements to 5 movements, allowing for a faster quantitative assessment of hand motor function than the FMA scale. This also lays the foundation of the assessment paradigm for our follow-up research on evaluation algorithm.

Project supported by the Beijing Municipal Natural Science Foundation under Grant JQ19020, National Natural Science Foundation of China under Grant 62025307, Grant U1913209, Grant62073319, Grant62103412 and China Postdoctoral Science Foundation under Grant 2021M693403.