康复学报2025，Vol.35Issue(4)：370-377,8.DOI:10.3724/SP.J.1329.2025.04006
图-时空卷积神经网络在脑卒中患者运动意图识别中的应用
Application of Graph Spatio-Temporal Convolutional Neural Network in Motor Intention Recognition of Stroke Patients
摘要
Abstract
Objective To evaluate the decoding accuracy and model performance of a graph spatio-temporal convolutional neural network(G-STCNN)in motor intention recognition of stroke patients.Methods We developed a novel G-STCNN model by integrating graph convolution and spatio-temporal convolution modules to simultaneously capture the local topological structure and global spatiotemporal dynamics of EEG signals.A total of 2 000 EEG samples from 50 patients with acute ischemic stroke(AIS)were obtained from the public"EEG dataset of stroke patients"provided by Xuanwu Hospital,Capital Medical University for valida-tion.The confusion matrix was used to compare the decoding accuracy[accuracy(ACC),positive predictive value(PPV),sensitivity(Sens.)and Kappa coefficient]of the four deep learning models,including deep convolutional neural network(Deep ConvNet),com-pact convolutional neural network(compact ConvNet),shallow convolutional neural network(shallow ConvNet)and G-STCNN.Ablation studies were conducted to assess the performance of the G-STCNN model after removing the graph convolution module(Module 2)and the spatio-temporal convolution module(Module 3).Results(1)Confusion matrix:the classification accuracies for the left and right hands of the deep convolutional neural network model were 57.44%and 50.89%,respectively;the misclassifica-tion rates were 42.56%and 43.56%,respectively.The classification accuracies for the left and right hands of the compact ConvNet model were 57.56%and 54.78%,respectively,and the misclassification rates were 42.44%and 39.67%,respectively.The classifica-tion accuracies for the left and right hands of the shallow convolutional neural network model were 56.56%and 49.56%,respectively,and the misclassification rates were 43.44%and 44.89%,respectively.The classification accuracies for the left and right hands of the graph spatio-temporal convolutional neural network model were 73.55%and 76.33%,respectively,and the misclassification rates were 26.45%and 23.67%,respectively.The graph spatio-temporal convolutional neural network exhibited the largest diagonal values and the darkest colors,in contrast with the smallest off-diagonal values and the lightest colors.(2)Decoding accuracy:Compared with the deep convolutional neural network,the compact convolutional neural network,and the shallow convolutional neural net-work,the ACC,PPV,Sens.,and Kappa coefficient of the graph spatio-temporal convolutional neural network were significantly higher,and the differences were statistically significant(P<0.05).(3)Module performance:Feature visualization in two-dimensional space revealed distinct and well-separated clusters for left-and right-hand tasks using the full G-STCNN model.Removing graph convolution Module 2 resulted in mixed left and right hands class distributions and blurred cluster boundaries,while removal of the spatiotemporal convolution Module 3 led to high overlap and reduced inter-class separability.Conclusion The G-STCNN model demonstrates superior decoding performance for motor imagery tasks in stroke patients and holds promise for providing data-driven,stratified rehabilitation interventions.关键词
脑卒中/运动功能障碍/脑电图/运动意图识别/深度学习/图-时空卷积神经网络Key words
stroke/motor dysfunction/electroencephalography/motor intention recognition/deep learning/graph spatio-tem-poral convolutional neural network引用本文复制引用
徐慧,马骏,张晶晶,张弢,赵志伟,单春雷..图-时空卷积神经网络在脑卒中患者运动意图识别中的应用[J].康复学报,2025,35(4):370-377,8.基金项目
国家自然科学基金面上项目(82272612) （82272612）
