Method for evaluating upper limb motor function of stroke patient

Abstract

The application discloses a method for evaluating upper limb motor function of a stroke patient, which comprises the following steps: firstly, collecting near-infrared signals of the brain of a subject, and preprocessing the collected signals through band-pass filtering and cubic spline interpolation, so as to convert the optical signals into oxyhemoglobin concentration data; secondly, calculating the Pearson correlation coefficient between the preprocessed multi-channel signals, constructing a functional connection matrix under different sparsity thresholds, observing the brain network connection, and performing graph theory analysis on the brain network connection; thirdly, performing b-clustering, classifying to obtain disease subtypes, and taking the functional connection matrix of the healthy people as the input of a convolutional neural network; and finally, fusing the output reliability of the convolutional neural network with demographic information features, performing support vector regression, and selecting an optimal model as a scoring model; the application constructs a brain network under different sparsity thresholds, and determines the optimal threshold through subsequent analysis, so that the data is more accurate and the space-time resolution is higher.

Description

Technical Field

[0001] This invention belongs to the field of near-infrared signal deep learning technology, and specifically relates to a method for assessing upper limb motor function in stroke patients that integrates demographic information and fNIRS multimodal deep learning methods. Background Technology

[0002] Stroke is an acute cerebrovascular disease caused by the sudden rupture or blockage of blood vessels in the brain, resulting in brain tissue damage. It includes ischemic stroke and hemorrhagic stroke. Stroke has a very high disability rate, causing various functional impairments, among which upper limb and hand dysfunction significantly impacts the ability to live independently. Currently, the Fugl-Meyer Upper Limb Dysfunction Rating Scale (FMA-UL) is commonly used to assess upper limb function in stroke patients. However, this method is time-consuming and easily influenced by the experience and proficiency of professionals, thus possessing a degree of subjectivity. Therefore, quantitative assessment of hand function rehabilitation has significant clinical importance.

[0003] Near-infrared functional brain imaging (fNIRS) can measure changes in the concentrations of oxyhemoglobin, deoxyhemoglobin, and total hemoglobin in specific brain tissues, thereby examining neuronal activity, cellular energy metabolism, and hemodynamic-related functions, reflecting the state and processing of the brain. fNIRS has advantages such as insensitivity to motion artifacts and high spatiotemporal resolution. Therefore, the application of fNIRS in the study of brain functional networks in stroke patients, and its guidance in rehabilitation, has become a focus of many current scientific research efforts. It can, to a certain extent, improve assessment efficiency, significantly reduce the workload of physicians, and make assessments more objective, efficient, and accurate.

[0004] Convolutional Neural Networks (CNNs) are widely used in deep learning image processing. They offer advantages such as shared convolutional kernels, the ability to process high-dimensional data, and automatic feature extraction. The K-means algorithm is a classic clustering algorithm, a type of unsupervised learning without labels, and is widely used in data classification.

[0005] The invention, with application number CN202210966098.0 and titled "A Method for Stroke Rehabilitation Assessment Based on Brain-Muscle Function Network Characteristics," discloses a method that collects EEG and EMG signals from stroke patients under a grip strength test, preprocesses them to construct variational mode decomposition (VMD) transfer entropy, quantifies the coupling between nodes using VMD-TE, selects the top 15% of the connection strength values ​​in the connection matrix as the threshold for binarization of the CMFN, and finally calculates brain-muscle function network characteristics. It compares the network characteristics of patients in three rehabilitation stages to assess stroke rehabilitation. However, this method has several drawbacks: first, EEG signals are easily interfered with by head movements and have low spatial resolution; second, when using variational mode decomposition to quantify coupling values, bandwidth overlap occurs when the signal length is too large, resulting in poor calculation accuracy; and third, it can only analyze the patient's rehabilitation stage and cannot quantitatively assess the rehabilitation level, thus having significant limitations. Summary of the Invention

[0006] To overcome the shortcomings of the existing technology, the present invention aims to provide a method for assessing upper limb motor function in stroke patients. This method integrates demographic information and a multimodal deep learning approach using fNIRS. It constructs a functional connectivity matrix by calculating the Pearson correlation coefficient. To remove spurious connections, two large-scale graph theory metrics are calculated for the functional connectivity matrix at different sparsities: the average weighted clustering coefficient and global efficiency. K-means binary clustering is performed on patients using the area under the graph theory metric-sparseness threshold curve as a feature to identify two stroke subtypes with different characteristics in global network information flow. Finally, the reliability of the convolutional neural network output is fused with demographic features to perform support vector regression, and the optimal model is selected as the scoring model. The method of the present invention provides more accurate data and higher resolution.

[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0008] A method for assessing upper limb motor function in stroke patients includes the following steps:

[0009] Step 1: Acquire multi-channel near-infrared data, perform filtering and artifact removal preprocessing, and convert it into concentration data;

[0010] Step 2: Construct a functional connectivity matrix from the preprocessed concentration data, calculate the graph theory AUC index of the functional connectivity matrix under different sparsities, and calculate the area under the graph theory index-sparseness threshold curve.

[0011] Step 3: Perform K-means clustering on the graph theory AUC index to divide the patients into two subtypes;

[0012] Step 4: Use a convolutional neural network model to perform binary classification on the functional connectivity matrix of a patient of one of the two subtypes at a certain sparsity threshold and the functional connectivity matrix of a healthy person, and perform leave-one-out cross-validation. Multiply the classification probability value output by the CNN model of each fold by the classification accuracy of the validation set of that fold as the hemodynamic phase response feature obtained based on the fNIRS signal, which is defined as HFS (Hemodynamic Feature Score).

[0013] Step 5: Fuse HFS with patient demographic features, input the fused features into the support vector regression model, and perform leave-one-out cross-validation. The data partitioning in the SVR model and the CNN model in Step 4 is completely consistent and corresponds one-to-one, to obtain the fitted FMA-UL and calculate its evaluation index. Finally, process the data of each subtype at each sparsity threshold in Step 4 and Step 5, and finally select the optimal threshold.

[0014] Step one specifically involves:

[0015] Acquire multi-channel near-infrared data, with a data acquisition time of t and a number of patient groups of l. p The number of healthy individuals is l h The number of channels is m, the sampling frequency is f, and the near-infrared data of the i-th patient is represented as X. i (n)={x1(n),x2(n),…,x m Similarly, for healthy individuals, the patient dataset is obtained as follows: (n)} Healthy people dataset

[0016] The acquired data is preprocessed, and the raw light intensity data X is converted into X. i (n)={x1(n),x2(n),…,x m (n)} is converted into optical density data Y i (n)={y1(n), y2(n),…,y m (n)}; A bandpass filter of 0.01-0.20Hz is used to eliminate physiological noise, namely heart rate and respiration, and cubic spline interpolation is used to eliminate motion artifacts; finally, the denoised optical density data is converted into hemoglobin concentration data; the patient dataset is converted into The same applies to healthy people.

[0017] Step two specifically involves:

[0018] To expand the sample size and ensure sufficient model training, the data of duration t was segmented into q segments, with some overlap in length between segments. Each patient's data set after segmentation contained q subsets. The brain network was considered to consist of nodes and edges connecting these nodes, with each channel considered a node. The Pearson correlation coefficient was used as a measure of connectivity between two nodes. The Pearson correlation coefficients for the oxyhemoglobin concentration data of m channels were calculated, yielding q×l. p An m×m correlation coefficient matrix R p The formula for calculating the Pearson correlation coefficient is as follows:

[0019]

[0020] Among them, z i This represents the time series of N points in channel i; z i，n ρ represents the nth point in the time series of channel i; i，j This represents the Pearson correlation coefficient between channel i and channel j, with a value ranging from -1 to 1, ρ. i，j When ρ = 1, the two channels have a perfect positive correlation; i，j，j When ρ = -1, the two channels have a completely negative correlation. i，j When the correlation coefficient is 0, the two channels are not correlated.

[0021] By applying the Fisher-Z transform to the Pearson correlation coefficient to make it conform to a normal distribution, a new functional connectivity matrix R is obtained. px Similarly, the functional connectivity matrix R of a healthy person can be obtained. hx The Fisher-Z transform formula is as follows:

[0022]

[0023] Since some spurious connections exist in the functional connectivity matrix, graph theory analysis is performed using different sparsity thresholds. When the modified Pearson correlation coefficient between nodes i and j in the network is greater than the threshold TH, a connection between the nodes is considered to exist, and the correlation coefficient is used as the weight w of the connection between network nodes. i，j Thus, a functional connectivity matrix with a sparsity threshold of TH is constructed. Within the sparsity range of 10-50%, with a step size of 5%, nine functional connectivity matrices are obtained. For each functional connectivity matrix under a sparsity threshold, graph theory metrics are calculated, including the average weighted clustering coefficient and global efficiency. The specific explanations and calculation formulas are as follows:

[0024] (1) For an undirected network, the clustering coefficient of the i-th node is:

[0025]

[0026] Where, k i Let i be the degree of node i;

[0027] The average weighted clustering coefficient of the network is as follows, where N represents the number of nodes:

[0028]

[0029] (2) For an undirected network, its global efficiency is expressed as:

[0030]

[0031] In the formula, d ij Let the shortest path length from node i to node j be expressed as:

[0032] d i，j =min(w i，j -1 )

[0033] For n TH After calculating the graph theory indices for the functional connectivity matrices at each sparsity threshold, a graph theory index-sparse threshold graph is plotted, with the graph theory index on the ordinate and the sparsity threshold on the abscissa. The area under the curve is calculated for each of the two graph theory indices, yielding A. C A E .

[0034] The clustering in step three specifically involves calculating A of the functional connectivity matrix of all samples. C A E The unsupervised learning K-means clustering algorithm without labels was used, with a cluster size of 2; the binary clustering results were two disease subtypes.

[0035] Step four specifically involves: using functional connectivity at different sparsity thresholds as features, a convolutional neural network method is used to perform classification tasks on each subtype and healthy individuals, thereby extracting features based on brain hemodynamic response. One-hot encoding is used, and leave-one-out cross-validation is employed. The data is divided into training, testing, and validation sets. The testing and validation sets each contain all subsets of a sample, and the remaining samples form the training set. The classification probability value output by each fold of the CNN model is multiplied by the classification accuracy of the validation set for that fold, and the result is used as the hemodynamic response feature obtained based on the fNIRS signal, defined as HFS (Hemodynamic Feature Score).

[0036] The convolutional neural network model in step four consists of two convolutional layers and four fully connected layers. Its structure is as follows: first convolutional layer, first pooling layer, second convolutional layer, second pooling layer, Dropout layer, BN layer, and fully connected layer. The output layer is a Softmax classification layer. The pooling layer performs max pooling, the loss function is classification cross-entropy, and the Adam optimizer is used for optimization.

[0037] Step five involves feature fusion of HFS and demographic information. Specifically, the demographic information includes gender, age, affected side, and mobility. The fused features serve as input to the support vector regression model, with the model label being the FMA score. The FMA score is assessed by an experienced rehabilitation physician on the day of fNIRS data collection, with a perfect score for healthy individuals. Leave-one-out cross-validation is used for training. To avoid data leakage, the test and validation sets in both the SVR and CNN models should correspond one-to-one to obtain the fitted FMA-UL, and the coefficient of determination R is calculated. 2 The mean squared error (RMSE) is used as the basis for the final determination coefficient R. 2 The scoring model is selected based on the root mean square error (RMSE).

[0038] Compared with the prior art, the present invention has the following advantages:

[0039] (1) The multi-channel near-infrared data obtained in step one is simple and fast in data measurement, less affected by head movement and electromagnetic interference, and the data is more accurate with higher spatiotemporal resolution.

[0040] (2) Based on the large-scale network index AUC, the stroke patients were clustered to identify two stroke subtypes from the perspective of the global information flow efficiency of the brain. The model was trained and predicted separately for patients with different disease types to avoid the influence of large differences in the brain networks of different types of patients on the results. This enabled the hierarchical assessment of hand dysfunction in stroke patients and improved the accuracy of the model.

[0041] (3) By integrating the hemodynamic response of the brain with demographic characteristics, the brain network characteristics and basic personal information of the subjects are fully considered, and multimodal information fusion modeling is realized, which improves the accuracy, robustness and generalization ability of the model.

[0042] (4) Compared with other existing FMA assessment methods, this invention includes more patients and can achieve basic coverage of high and low FMA scores, thereby improving the accuracy of model prediction. Attached Figure Description

[0043] Figure 1 This is a flowchart of the algorithm used in this invention.

[0044] Figure 2This is a schematic diagram of a CNN network structure.

[0045] Figure 3 The image shows the prediction results of the SVR model.

[0046] Figure 4 The result is shown in the figure. Figure 4 (a) in the equation represents the coefficient of determination R. 2 (b) shows the mean square error (RMSE). Detailed Implementation

[0047] The process and advantages of the present invention will be described in detail below with reference to the accompanying drawings.

[0048] The specific experiments of this invention were conducted under the Windows 10 (64-bit) operating system. The brain network was constructed based on Matlab R2020b, and the network was trained based on the TensorFlow 2.6.0 framework, with Keras as the front-end API and Python 3.9 as the compiler virtual environment.

[0049] The near-infrared signal dataset used in this invention comes from experimental acquisition. In a resting state, the subject maintains a relaxed sitting posture with palms facing up on their lap for 180 seconds, and near-infrared signals from 48 channels are acquired.

[0050] Based on the aforementioned near-infrared dataset, the following method was adopted: Figure 1 The algorithm flow shown, using partial data as an example, illustrates the specific implementation steps of this invention as follows:

[0051] Step 1: Acquire multi-channel near-infrared data, perform filtering and artifact removal preprocessing, and convert it into concentration data.

[0052] Data was collected from 94 patient groups and 12 healthy individuals. The number of acquisition channels was 48, and the acquisition time was 180 seconds. The near-infrared data of the i-th patient is represented as X. i (n)={x1(n),x2(n),…,x 48 (n)}, the same applies to healthy individuals. The resulting patient dataset is X. p =X1(n), X2(n), ..., X 94 (n), the dataset of healthy people is X h =X1(n), X2(n), ..., X 12 (n).

[0053] The above data is preprocessed, and the modified Lambert-Beer law is used to transform the original light intensity data X. i (n)={x1(n),x2(n),…,x 48 (n)} is converted into optical density data Y i(n)={y1(n), y2(n),…,y 48 (n)}. A bandpass filter of 0.01–0.20 Hz was used to eliminate physiological noise (heart rate and respiration), and cubic spline interpolation was used to eliminate motion artifacts. Finally, the denoised optical density data was converted into hemoglobin concentration data. Since oxyhemoglobin concentration is considered to have a better signal-to-noise ratio than deoxyhemoglobin concentration, oxyhemoglobin concentration was used for subsequent analysis. The patient dataset was converted to Zn0. p =Z1(n), Z2(n), ..., Z 94 (n), the same applies to healthy people.

[0054] Step 2: Construct a functional connectivity matrix from the preprocessed concentration data, calculate the graph theory AUC index of the functional connectivity matrix under different sparsities, and calculate the area under the graph theory index-sparseness threshold curve.

[0055] The 180-second data set was segmented into 20 segments with 50% overlap between segments. Each patient sample after segmentation contained 20 subsamples. Pearson correlation coefficients were calculated for the 48-channel oxyhemoglobin concentration data, resulting in 1880 48×48 correlation coefficient matrices R0. p The formula for calculating the Pearson correlation coefficient is as follows:

[0056]

[0057] Then, the Pearson correlation coefficient is corrected by Fisher-Z transformation to make it conform to a normal distribution, resulting in a new functional connectivity matrix R. px Similarly, the functional connectivity matrix R of a healthy person is obtained. hx The Fisher-Z transform formula is as follows:

[0058]

[0059] Because some spurious connections exist in the functional connectivity matrix, a thresholding method is used to remove them. Specifically, different sparsity thresholds are applied for graph theory analysis. Sparsity is the proportion of important connections among all connections; a higher correlation coefficient indicates a higher importance for the connection. Sparsity thresholds ranging from 10% to 50% are used, with a step size of 5%, resulting in sparse networks with nine thresholds. When the Pearson correlation coefficient between nodes i and j in the primitive brain network is greater than the threshold TH, a connection between the nodes is considered to exist, and this correlation coefficient is used as the weight w of the connection between network nodes. i，j This allows us to construct a functional connectivity matrix with a sparsity threshold of TH. Based on the selected threshold range and step size, the number of new functional connectivity matrices is n. THFor each sparsity threshold, graph theory metrics are calculated for the functional connectivity matrix, including the average weighted clustering coefficient and global efficiency. The specific explanations and calculation formulas are as follows:

[0060] (1) The average weighted clustering coefficient characterizes the degree to which nodes in a graph tend to cluster. It is a measure of local separation in the graph. The higher the average weighted clustering coefficient, the higher the modularity of the network and the higher the local efficiency. For an undirected network, the clustering coefficient of the i-th node is:

[0061]

[0062] Where, k i Let be the degree of node i.

[0063] The average weighted clustering coefficient of the network is as follows, where N represents the number of nodes:

[0064]

[0065] (2) Global efficiency is the inverse property of shortest path length, representing the ability of nodes to transmit information. The higher the global efficiency, the stronger the network's ability to transmit information. For an undirected network, its global efficiency is expressed as:

[0066]

[0067] In the formula, d ij Let be the shortest path length from node i to node j. The larger the correlation coefficient between two nodes, the shorter their distance is considered to be. The weight of the connection between nodes is mapped to the distance length, using the reciprocal mapping. That is, the shortest distance length between nodes is expressed as:

[0068] d i,j =min(w i,j -1 )

[0069] After calculating the graph theory indices for the functional connectivity matrices at nine sparsity thresholds, a graph theory index-sparse threshold graph was plotted, with the graph theory index on the ordinate and the sparsity threshold on the abscissa. The area under each of the four curves was calculated to obtain A. C A E .

[0070] Step 3: Perform K-means clustering on the graph theory AUC index to divide the patients into two subtypes; each subtype is then processed in the subsequent steps.

[0071] Calculate A of all functional connectivity matrices C A E An unsupervised K-means clustering algorithm without labels was used, with a cluster size of 2. In this example, the number of patients in the two subtypes was 76 and 18, respectively.

[0072] Step 4: Use a convolutional neural network to perform binary classification of the functional connectivity matrix of a subtype of patients at a certain sparsity threshold with the functional connectivity matrix of healthy individuals, and perform leave-one-out cross-validation. Multiply the classification probability value output by the CNN model at each fold by the classification accuracy of the validation set at that fold, and use the result as the hemodynamic phase response feature obtained based on the fNIRS signal, defined as HFS (Hemodynamic Feature Score).

[0073] A convolutional neural network model consists of two convolutional layers and four fully connected layers, such as... Figure 2 As shown, its structure consists of: a first convolutional layer, a first pooling layer, a second convolutional layer, a second pooling layer, a Dropout layer, a BN layer, and a fully connected layer. The output layer is a Softmax classification layer. The pooling layers use max pooling, and the loss function is the classification cross-entropy, optimized using the Adam optimizer to improve convergence speed.

[0074] Using functional connectivity at different sparsity thresholds as features, the aforementioned convolutional neural network was used to perform classification tasks on each subtype and healthy individuals, thereby extracting features based on cerebral hemodynamic responses. One-hot encoding was used, with patient labels [0,1] and healthy individual labels [1,0]. The data was divided into training, test, and validation sets. The test and validation sets each contained subsets of a single sample, and the remaining samples formed the training set. The model was trained to obtain the Hemodynamic Frame of Action (HFS).

[0075] Step 5: Fuse the HFS with patient demographic characteristics, input the data into the Support Vector Regression (SVR) model, and perform leave-one-out cross-validation. To prevent data leakage, the data partitioning in the SVR model is completely identical and corresponds one-to-one with that in Step 4. Obtain the fitted FMA-UL and calculate its evaluation index. The evaluation index is the coefficient of determination R. 2 The mean squared error (RMSE) is used as the basis for the final determination coefficient R. 2 The scoring model is selected based on the root mean square error (RMSE).

[0076] The data under each sparsity threshold is processed in steps four and five, and the optimal threshold is finally selected.

[0077] Demographic information includes sex (male = "1", female = "0"), age, affected side (left = "1", right = "2", healthy = "3"), and mobility information (mobile = "1", immobile = "0"). After feature fusion, the data is input into the support vector regression model, labeled with the FMA score. The FMA score is assessed by an experienced rehabilitation physician on the day of fNIRS data collection, with healthy individuals receiving a perfect score. Leave-one-out training is used, and to avoid data leakage, the test and validation sets in both the SVR and CNN models should correspond one-to-one. The evaluation coefficient R is calculated... 2 The model is evaluated using the root mean square error (RMSE), such as... Figure 3 As shown. The model evaluation index is the coefficient of determination R. 2 The mean squared error (RMSE) is used as the basis for the final determination coefficient R. 2 The scoring model is selected based on the root mean square error (RMSE). The final result in this example is as follows: Figure 4 As shown, considering all factors, the model prediction performance is optimal when the sparsity threshold is set to 10%.

[0078] The FMA assessment features of this invention are derived from two sources: features extracted based on cerebral hemodynamic response and features based on demographic information. Using functional connectivity at different sparsity thresholds as features, a convolutional neural network (CNN) method is employed to perform classification tasks on each subtype and healthy individuals, thereby extracting features based on cerebral hemodynamic response. The output of the CNN model is then fused with demographic information and input into a support vector regression (SVR) model to fit the FMA-UL of different subtypes of patients. Based on the evaluation coefficients of the regression model, a model with an appropriate sparsity is selected as the final scoring model.

Claims

1. A method for assessing upper limb motor function in stroke patients, characterized in that, Includes the following steps: Step 1: Acquire multi-channel near-infrared data, perform filtering and artifact removal preprocessing, and convert it into concentration data; Step 2: Construct a functional connectivity matrix from the preprocessed concentration data, calculate the graph theory AUC index of the functional connectivity matrix under different sparsities, and calculate the area under the graph theory index-sparseness threshold curve. Step 3: Perform K-means clustering on the graph theory AUC index to divide the patients into two subtypes; Step 4: Use a convolutional neural network model to perform binary classification on the functional connectivity matrix of a patient of one of the two subtypes at a certain sparsity threshold and the functional connectivity matrix of a healthy person, and perform leave-one-out cross-validation. Multiply the classification probability value output by the CNN model of each fold by the classification accuracy of the validation set of that fold as the hemodynamic phase response feature obtained based on the fNIRS signal, which is defined as HFS (Hemodynamic Feature Score). Step 5: Fuse HFS with patient demographic features, input the fused features into the support vector regression model, and perform leave-one-out cross-validation. To prevent data leakage, the data partitioning in the SVR model and the CNN model in Step 4 is completely consistent and corresponds one-to-one. Obtain the fitted FMA-UL and calculate its evaluation index. Finally, process the data of each subtype at each sparsity threshold as described in Step 4 and Step 5, and finally select the optimal threshold. Step two specifically involves: To expand the sample size and ensure the model is fully trained, a training period of [duration missing] was conducted. The data is split into segments. There are several segments, with some overlap in length; after segmentation, each patient's data contains... Subdatasets; The brain network is considered to consist of nodes and edges connecting these nodes, with each channel considered a node. The Pearson correlation coefficient is used as a measure of connectivity between two nodes. The Pearson correlation coefficients for oxyhemoglobin concentration data from m channels are calculated respectively. indivual Correlation coefficient matrix R p The formula for calculating the Pearson correlation coefficient is as follows: Among them, z Indicates channel of Time series of points; z , Indicates channel The first time series One point; , Indicates channel With channel The Pearson correlation coefficient, its value is 1~1, At that time, the two channels have a perfectly positive correlation; At that time, the two channels have a completely negative correlation. At that time, the two channels are not correlated; By applying the Fisher-Z transform to the Pearson correlation coefficient to make it conform to a normal distribution, a new functional connectivity matrix R is obtained. px Similarly, the functional connectivity matrix R of a healthy person can be obtained. hx The Fisher-Z transform formula is as follows: Since some spurious connections exist in the functional connectivity matrix, graph theory analysis is performed using different sparsity thresholds. When the modified Pearson correlation coefficient between nodes i and j in the network is greater than the threshold TH, a connection between the nodes is considered to exist, and the correlation coefficient is used as the weight of the connection between network nodes. i,j Thus, a functional connectivity matrix with a sparsity threshold of TH is constructed. Within the sparsity range of 10-50%, with a step size of 5% sparse functional connectivity matrix, nine functional connectivity matrices are obtained. For each functional connectivity matrix under the sparsity threshold, graph theory metrics are calculated, including the average weighted clustering coefficient and global efficiency. The specific explanations and calculation formulas are as follows: (1) For an undirected network, the clustering coefficient of the i-th node is: in, Let i be the degree of node i; The average weighted clustering coefficient of the network is as follows, where N represents the number of nodes: (2) For an undirected network, its global efficiency is expressed as: In the formula, For nodes To the node The shortest path length is expressed as: For n TH After calculating the graph theory indices for the functional connectivity matrices at each sparsity threshold, a graph theory index-sparse threshold graph is plotted, with the graph theory index on the ordinate and the sparsity threshold on the abscissa. The area under the curve is calculated for each of the two graph theory indices, yielding A. C A E .

2. The method for assessing upper limb motor function in stroke patients according to claim 1, characterized in that, Step one specifically involves: Acquire multi-channel near-infrared data, with a data acquisition time of t and a number of patient groups. The number of healthy people is The number of channels is m, the sampling frequency is f, and the near-infrared data of the i-th patient is represented as: Similarly, for healthy individuals, the patient dataset is obtained as follows: The healthy people dataset is ; The acquired data is preprocessed, and the raw light intensity data is... Convert to optical density data A bandpass filter of 0.01–0.20 Hz was used to eliminate physiological noise, namely heart rate and respiration, and cubic spline interpolation was used to eliminate motion artifacts. Finally, the denoised optical density data was converted into hemoglobin concentration data; the patient dataset was then converted into… The same applies to healthy people. .

3. The method for assessing upper limb motor function in stroke patients according to claim 1, characterized in that, The clustering in step three specifically involves calculating A of the functional connectivity matrix of all samples. C A E The unsupervised learning K-means clustering algorithm without labels was used, with a cluster size of 2; the binary clustering results were two disease subtypes.

4. The method for assessing upper limb motor function in stroke patients according to claim 1, characterized in that, Step four specifically involves: using functional connectivity at different sparsity thresholds as features, a convolutional neural network method is used to perform classification tasks on each subtype and healthy individuals, thereby extracting features based on brain hemodynamic response. One-hot encoding is used, and leave-one-out cross-validation is employed. The data is divided into training, testing, and validation sets. The testing and validation sets each contain all subsets of a sample, and the remaining samples form the training set. The classification probability value output by the CNN model at each fold is multiplied by the classification accuracy of the validation set at that fold, and the result is used as the hemodynamic response feature obtained based on the fNIRS signal, defined as HFS (Hemodynamic Feature Score).

5. The method for assessing upper limb motor function in stroke patients according to claim 4, characterized in that, The convolutional neural network model in step four consists of two convolutional layers and four fully connected layers. Its structure is as follows: first convolutional layer, first pooling layer, second convolutional layer, second pooling layer, Dropout layer, BN layer, and fully connected layer. The output layer is a Softmax classification layer. The pooling layer performs max pooling, the loss function is classification cross-entropy, and the Adam optimizer is used for optimization.

6. The method for assessing upper limb motor function in stroke patients according to claim 1, characterized in that, Step five involves feature fusion of HFS and demographic information. Specifically, the demographic information includes gender, age, affected side, and mobility. The fused features serve as input to the support vector regression model, with the model label being the FMA score. The FMA score is assessed by an experienced rehabilitation physician on the day of fNIRS data collection, with a perfect score for healthy individuals. Leave-one-out cross-validation is used for training. To avoid data leakage, the test and validation sets in both the SVR and CNN models should correspond one-to-one to obtain the fitted FMA-UL, and the coefficient of determination (COD) is calculated. The root mean square error (RMSE) is used as the basis for determining the coefficient of determination. The scoring model is selected based on the root mean square error (RMSE).

Images