A deep learning-based hand motion classification for hand dysfunction assessment in cervical spondylotic myelopathy

Abstract

<p>Hand motor function is found to be associated with cervical spondylotic myelopathy (CSM). This study aimed to explore a deep learning-based hand motion classification to quantify the severity. A hand dysfunction assessment method based on hand motion data and clinical information variables was proposed. A squeeze-and-excite multivariate long short-term memory fully convolutional network (SE-MLSTM-FCN) was developed to extract feature map from hand motion data. Feature map was then combined with clinical information variables to determine the severity by various classifiers. To evaluate the method, kinematic data of the left and right hands during rapid grip and release and basic clinical information were collected from 186 CSM patients with different severity levels. The results show that when only hand motion data was used for severity assessment, SE-MLSTM-FCN achieved an accuracy of 79.97 %, an F1 score of 0.7937 and a Kappa score of 0.6837, outperforming the baseline models. Moreover, it had a 15 % lower computational complexity, indicating that the proposed feature extraction model is superior in both performance and efficiency. By combining clinical information variables with the feature map from SE-MLSTM-FCN, random forest (RF) achieved the best classification result, with an accuracy of 83.06 %, an F1 score of 0.8282, a Kappa score of 0.7332 for three-class severity assessment. Therefore, the deep learning approach based on hand motion data and clinical information is promising for assessing the severity of hand dysfunction in CSM, which would be of great value in monitoring CSM progression and evaluating treatment effect. <br></p>