毕业生知识库

脑卒中、脊髓损伤等疾病往往会造成下肢运动功能障碍，需要通过康复训练来恢复患肢的运动功能，而训练方案需要依据患者的运动功能水平进行个性化设计。为此，需要准确评价患者的运动功能水平。目前，临床的运动功能评价主要采用各种功能量表，由治疗师人工完成，存在评价过程繁琐、评价结果依赖治疗师主观经验等缺点。近年来，如何基于电生理学、运动学与动力学等数据客观分析和评价人体的运动功能逐渐成为康复医疗领域的研究热点问题。本文针对基于表面肌电信号、运动学数据、足底压力数据的下肢运动功能分析与评价方法展开研究，主要的研究内容和创新工作如下：

1. 提出了一种基于表面肌电信号的运动功能分析与评价方法。采用非负矩阵分解算法提取下肢表面肌电信号的肌肉协同模块，并基于协同模块之间的余弦相似度，提出了肌肉协同模块的稳定性分析和一致性分析方法。针对运动功能评价问题，提出了一种基于健康人肌电信号的参考协同模块计算方法，并将被试的肌肉协同模块与参考协同模块的相似度作为运动功能的评价指标。最后，基于脊髓损伤患者和健康人踏车训练时采集的肌电信号数据，对肌肉协同模块的稳定性、不同被试间肌肉协同模块的一致性以及上述运动功能评价指标进行了实验验证。

2. 设计了一种基于运动学数据的运动功能量化评价模型。基于动作捕捉系统采集的运动学数据，应用卷积神经网络从信号的通道分布和时间序列两个层面分别获取信号的空间特征和时间特征，进而引入注意力模块，分别计算了通道和时间序列点的权重，并将权重和两类特征进行聚合，经过长短期记忆网络层和全连接层，给出运动功能的量化评价指标。最后，基于两个公开数据集验证了所提方法的有效性，结果表明本文的方法具有较低的平均绝对偏差。

3. 提出了基于足底压力数据的运动功能障碍检测和运动功能评价方法。基于随机森林、支持向量机、K近邻以及CNN-LSTM方法分别建立了运动功能障碍检测模型。同时，根据UPDRSM量表的评分将运动功能障碍分为六个等级，并基于CNN-LSTM建立了运动功能障碍等级识别模型。最后，在帕金森患者的足底压力数据集上开展了实验验证，结果表明，上述运动功能障碍定性检测模型均可获得较高的准确率，其中CNN-LSTM模型准确率最高。同时，基于上述数据集的三个子集进行了运动功能量化评价实验，结果表明应用本文所提的方法可以在三个子集上分别获得86.15%、88.91%、90.97%的评价准确率。

4. 建立了一种基于运动学数据与足底压力数据融合的运动功能障碍检测模型。基于髋关节骨关节炎患者步行时采集的髋关节角度数据以及足底压力数据，分别构建了基于运动学数据和基于足底压力数据的运动障碍识别模型，进而融合上述两种基于单模态数据的识别模型输出，构建了基于数据融合的运动障碍识别模型。实验结果表明，该模型的运动功能障碍检测准确率达92.86%，比基于单模态数据的识别模型提高了2.38%。

Stroke, spinal cord injury and other diseases often cause lower limb motor dysfunction, which needs rehabilitation training to restore the motor function of the affected limb. And the training program needs to be personalized according to the level of motor function of the patient. Therefore, it is necessary to accurately evaluate the motor function level of patients. At present, clinical motor function evaluation mainly adopts various functional scales, which are completed manually by therapists. There are some shortcomings, such as the tedious evaluation process and the dependence of evaluation results on subjective experience of therapists. In recent years, how to objectively analyze and evaluate the motor function of human body based on the electrophysiology, kinematics and kinetics data has gradually become a hot research topic in the field of rehabilitation medicine. This paper focuses on the analysis and evaluation methods for lower limb motor function based on surface electromyographic signal, kinematics data and plantar pressure data. The main research contents and innovative work are as follows:

1. An analysis and evaluation method of motor function based on surface electromyographic signal is proposed. Non-negative matrix decomposition algorithm was used to extract the muscle synergy modules of lower limb surface EMG signal. And based on the cosine similarity between the synergy modules, the stability analysis and consistency analysis for the muscle synergy modules were proposed. Aiming at the problem of motor function evaluation, a calculation method of reference synergy module based on electromyographic signal of healthy subjects was proposed. The similarity between the muscle synergy modules and reference synergy modules was used as the evaluation index of motor function. Finally, based on the electromyographic data collected from patients with spinal cord injury and healthy subjects during cycling training, the stability of muscle synergy modules, the consistency of muscle synergy modules between different subjects and the above motor function evaluation indexes were verified experimentally.

2. A quantitative evaluation model of motion function based on kinematic data is designed. Based on the kinematic data collected by the motion capture system, the convolutional neural network is applied to obtain the spatial and temporal characteristics of the signal from the channels and time series of the signal. And the attention module is introduced to calculate the weight of the channel and time series points respectively. The weight and two kinds of features are aggregated. After the long short-term memory layer and the full connection layer, The quantitative evaluation index of motor function is given. Finally, the proposed method is validated based on two public datasets, and the results show that the proposed method has a low mean absolute deviation.

3. A method of motor dysfunction detection and motor function evaluation based on plantar pressure data was proposed. Based on random forest, support vector machine, K nearest neighbor and CNN-LSTM method, motion dysfunction detection models were established respectively. At the same time, the motor dysfunction was divided into six levels according to the UPDRSM scale. The recognition model of motor dysfunction levels was established based on CNN-LSTM. Finally, the experimental verification was carried out on the plantar pressure data set of Parkinson's patients. The results showed that the above qualitative detection models of motor dysfunction could achieve high accuracy, among which the CNN-LSTM model had the highest accuracy. At the same time, the motor function quantitative evaluation experiment was carried out based on three subsets of the above dataset. The results show that the evaluation accuracy is 86.15%, 88.91% and 90.97% on the three subsets by using the method proposed in this paper.

4. A motor dysfunction detection model based on the fusion of kinematics data and plantar pressure data was established. Based on the hip angle data and plantar pressure data collected during walking of patients with hip osteoarthritis, the recognition models were respectively constructed. Then, the output of the above two recognition models based on single modal data was integrated to construct the movement disorder recognition model based on data fusion. Experimental results show that the accuracy of motion dysfunction detection of this model is 92.86%, which is 2.38% higher than that of recognition model based on single mode data.