毕业生知识库

康复始于评价，止于评价。评价自始至终贯穿于脑卒中患者的康复周期，对于患者的手部康复至关重要。然而现阶段的手部运动功能评价仍然存在着许多不足和局限性：缺少对患者手部运动数据的深度挖掘，手部运动功能评价精度较差；手部运动功能评价依赖于治疗师的观测，主观性强；缺少对患者手部运动功能的定量评价；量表存在“天花板效应”，对于手功能较好的患者无法区分其康复等级的差异。针对上述问题，本文围绕脑卒中患者的手部运动功能智能化定量评价开展了研究，论文的主要研究内容和创新点如下：
1. 提出了一种基于经验模态分解融合集成长短期记忆神经网络的患者数据增广方法(EMD-ILSTM)，并开展了临床实验验证。首先，利用经验模态分解方法，将1 维运动信号分解为多维信号矩阵，并利用LSTM 网络对整体的多维信号矩阵进行训练和增广。之后，在手部运动公开数据集Ninaweb 中进行分类测试，测试结果表明，在分类任务中使用增广数据可以提高5.2% 的分类准确率。此外，在中国康复研究中心附属北京博爱医院进行了临床测试。通过对患者的手部运动数据降维处理，验证了降维后的增广数据与原始数据的类内距离更小，类间距离更大，证明了数据增广的有效性。
2. 提出了一种基于集成学习的手部运动功能半定量评价方法，并开展了临床实验验证。首先，为了确定临床实验范式，提出一种基于LSTM 神经网络的手部运动学分析方法，利用层次聚类将公开数据集中的20 种手部运动分为4 类，并确定了5 个聚类中心动作。其次，设计了基于Leap Motion 传感器的手部运动数据采集系统，并基于该系统开展了临床实验，采集了50 名脑卒中患者的手部运动数据。最后，利用基于子空间的集成学习方法对患者进行分类，并与三种传统分类方法(加权K 最近邻、决策树、支持向量机) 进行了对比。同时，利用Wilcoxon 符号秩次检验、Kappa 检验两种非参数统计指标对该评价方法的有效性进行了验证。
3. 提出了一种基于高斯混合模型的手部运动功能定量评价方法，并开展了临床实验验证。首先，根据多种手部运动功能评价量表和多位治疗师的临床经验，设计了手部运动功能评价范式和基于数据手套和肌电手环的多模态数据采集系统。其次，基于该系统开展了临床实验。采集了35 名患者受试者和35 名健康受试者的手部运动数据和表面肌电信号(sEMG) 数据，并建立了脑卒中患者数据集和健康人对照组数据集。然后提取了11 种手部运动特征值，并进行特征组合。最后，完成了基于高斯混合模型的脑卒中患者的手部运动功能定量评价，并利用斯皮尔曼系数验证了该评价方法的有效性。
4. 提出了一种基于手部运动功能评价图网络(HAGCN) 的多模态手部运动精细评价框架，并开展了临床实验验证。首先，提出了HAGCN 网络结构，不仅在空间域对手部的关节点进行空间结构方面的架构和注意力机制的分析，并且在时间域上设计了卷积计算算法。其次，利用LSTM 神经网络处理表面肌电信号数据，自动提取表面肌电信号特征。然后，提出加权融合评价方法，并利用该框架进行手部运动功能评价。最后，开展临床实验，利用斯皮尔曼系数验证了该评价框架的有效性。该框架不仅可以完成手部运动功能的精细评价，同时可以避免一些传统量表的“天花板效应”。
关键词：手部运动功能定量评价，运动学分析，数据增广，高斯混合模型，图卷积网络

Rehabilitation begins and ends with assessment. Assessment accompanies stroke patients’ rehabilitation cycle and is significant for their hand rehabilitation. However, there are still many shortcomings and limitations in assessing hand motor function: lack of further mining of patient’s data and poor accuracy of hand motor function assessment; the assessment of hand motor function depends on the observation of the therapist and is highly subjective; lack of quantitative assessment of the patient’s hand motor function; there is a “ceiling effect” in the scale, and it is impossible to distinguish the difference in rehabilitation grade of patients with better hand function. In response to the above problems, this paper focuses on the intelligent quantitative assessment of the hand motor function of stroke patients. The main research contents and innovations of this paper are as follows:
1. A patient data augmentation method based on empirical mode decomposition and integrated long short-term memory neural network (EMD-ILSTM) is proposed and validated in clinical trials. This method decomposes the 1-dimensional motor signal into a multi-dimensional signal matrix by EMD, and the whole multidimensional signal matrix is trained and augmented by the LSTM network. After that, the classification test is conducted on the public dataset of hand movements called “Ninaweb.” The test results show that the classification accuracy could be improved by 5.2% using the enhanced data in the classification task. In addition, clinical tests are conducted at Beijing Bo’ai Hospital, affiliated with the China Research and Rehabilitation Center. Through dimensionality reduction processing of the collected patient hand movements data, it is verified that the intra-class distance between the augmented data and the original data after dimensionality reduction is smaller, and the inter-class distance is larger, which proves the effectiveness of data augmentation.
2. A semi-quantitative assessment method of hand motor function based on ensemble learning is proposed and validated in clinical trials. Firstly, a method of hand kinematics analysis based on LSTM is proposed to determine the clinical trial paradigm. Twenty hand movements in the public data set are divided into four classes by hierarchical clustering, and five cluster center movements are determined. Secondly, a data acquisition system based on Leap Motion is designed, and a clinical trial is carried out based on the system to collect the hand movement data of 50 stroke patients. Finally, the ensemble learning method based on ensemble subspace is used to classify patients. The effectiveness of the proposed method is verified by comparing the classification accuracy with other traditional classification methods (W-kNN, Decision Tree, SVM) and utilizing two non-parametric statistical indexes, Wilcoxon signed rank test
and Kappa test.
3. A quantitative assessment method of hand motor function based on the Gaussian mixture model is proposed and validated in clinical trials. Firstly, according to various hand function assessment scales and the clinical experience of multiple therapists, a hand function assessment paradigm and a multi-modality data acquisition system based on data gloves and sEMG bracelets is designed. Based on the system, 35 patient subjects and 35 healthy subjects’ hand movements and sEMG data are collected, and the data sets of stroke patients and healthy subjects are established. Then, 11 kinds of feature values are extracted manually according to the movement characteristics of the activities, and the feature combination is carried out. Finally, the qualitative assessment of hand motor function based on the Gaussian mixture model is completed, and the effectiveness of the assessment method is proved by the Spearman coefficient.
4. A multi-modality hand motion fine assessment framework based on HAGCN (Hand Motor Function assessment Graph Network) is proposed and validated in clinical trials. First of all, the HAGCN network structure is proposed, which not only analyzes the structure of the spatial structure and the attention mechanism of the joint nodes of the hands in the spatial domain but also designs the convolution calculation algorithm in time. Then, the sEMG data is processed by the LSTM network, and the features
of sEMG are extracted automatically. Then, this framework proposes a weighted fusion method to assess the hand motor function. Finally, with the help of the clinical trial data, the validity of the assessment framework is verified by using the Spearman's coefficient. This framework can not only complete fine assessment but also avoid the “ceiling effect” of some traditional scales.
Key Words: Quantitative Assessment of Hand Motor Function, Kinematic Analysis, Data Augmentation, Gaussian Mixture Model, Graph Convolutional Network