Post-stroke hemiplegia rehabilitation evaluation method based on attention map control network

Abstract

The invention discloses a post-stroke hemiplegia rehabilitation evaluation method based on an attention map control network, and the method comprises the steps: firstly constructing a hemiplegia illness attention map control network framework HemiplegiaAGCN, carrying out the cleaning and classification of the clinical physical examination evaluation text data of a patient, and carrying out the analysis of the clinical physical examination evaluation text data; a symptom text analysis module and a scale analysis module are respectively designed to process a symptom text and a scale evaluation text, features of the symptom text and the scale evaluation text are analyzed in a targeted manner, then the features are fused by using a selective attention fusion unit designed in an illness state evaluation module, and finally classification evaluation is performed. The hemiplegia condition attention map control network designed by the method processes clinical physical examination evaluation text data of a patient in a targeted manner from different angles, extracts features most related to the condition of the patient and performs feature fusion to execute a condition evaluation classification task, so that the accuracy of hemiplegia condition evaluation is improved, doctors are assisted to perform rehabilitation evaluation, and the diagnosis efficiency is improved. The purposes of improving the efficiency of doctors, reducing the misdiagnosis rate and improving the utilization rate of medical resources are achieved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of rehabilitation medicine, and particularly relates to a post-stroke hemiplegia rehabilitation evaluation method based on an attention graph control network. BACKGROUND

[0002] In the existing clinical treatment of hemiplegia patients, the patient's condition is divided into flaccid stage, spastic stage and recovery stage. In order to avoid missing the best treatment period, the doctor first needs to use the patient's physical examination symptom data and evaluation scale results to perform condition staging evaluation. However, the staging diagnosis depends on the subjective experience of the doctor, and there is a possibility of misdiagnosis. With the help of deep learning technology, a large amount of patient symptom data and patient rehabilitation evaluation scale data can be processed and analyzed, which provides a new possibility for the transformation of the condition diagnosis and evaluation task of hemiplegia patients to high efficiency and intelligence.

[0003] In the prior art, patent CN202211005220.4 provides an automatic rehabilitation evaluation method based on LSTM, which reduces the influence of subjective factors of rehabilitation therapists on rehabilitation evaluation results, but ignores the possible situation that patients may not be able to perform actions and may not be able to wear sensor devices in clinical practice. Patients are difficult to collect effective data according to the planned arrangement, which may cause the rehabilitation evaluation task to be unable to be completed.

[0004] Patent CN202311163760.X discloses a stroke patient rehabilitation evaluation system, which collects patient body parameters and then performs rehabilitation evaluation, which can bring convenience to the diagnosis and treatment of the patient's condition, but ignores the existing clinical physical examination evaluation text data of the patient.

[0005] In summary, there is a need to study a hemiplegia rehabilitation evaluation method to solve the problems of the existing methods in collecting physiological data based on sensors for rehabilitation evaluation and using patient clinical physical examination evaluation text data for rehabilitation evaluation in clinical situations. SUMMARY

[0006] To solve the above technical problems, the application provides a post-stroke hemiplegia rehabilitation evaluation method based on an attention graph control network, which designs a hemiplegia condition attention graph control network to process patient clinical physical examination evaluation text data from different angles, extracts the most relevant features of the patient's condition, and performs feature fusion to perform condition evaluation classification tasks.

[0007] The technical scheme adopted by the application is as follows: a post-stroke hemiplegia rehabilitation evaluation method based on an attention graph control network, the specific steps are as follows:

[0008] S1, constructing a hemiplegia condition attention graph control network framework HemiplegiaAGCN;

[0009] The Hemiplegia AGCN comprises a symptom text analysis module, an assessment scale analysis module and a disease assessment module.

[0010] S2, data cleaning and classification are performed on the clinical examination evaluation text data to obtain symptom text and scale evaluation text;

[0011] The clinical examination evaluation text data is collected for data cleaning and classification, and the description of the current physical symptoms of the patient is divided into symptom text, and the quantitative evaluation of the physical condition of the patient is divided into scale evaluation text.

[0012] Among them, the symptom text uses descriptive vocabulary, including: "patient", "patient", "unable", "restricted", "pain"; the scale evaluation text uses quantitative score keywords, including: "score", "classification", "scale", "stage".

[0013] S3, input the symptom text obtained in step S2 into the symptom text analysis module of the framework constructed in step S1, and output the symptom features;

[0014] The symptom text analysis module comprises a symptom graph construction module, an adaptive edge weight adjuster, a symptom analysis unit and an associated semantic representation fusion module.

[0015] Among them, the symptom text obtained in step S2 is first graphically modeled by the symptom graph construction module, the adaptive edge weight adjuster is used to adjust the edge weight to obtain the optimal weight, and then the symptom analysis unit is used to output the symptom features and the associated semantic representation fusion module is used for fusion.

[0016] S4, input the scale evaluation text obtained in step S2 into the assessment scale analysis module of the framework constructed in step S1, and output the scale features;

[0017] The assessment scale analysis module comprises a hypergraph construction module based on the core classification rules of stroke ICF, a hypergraph convolution network HyperGCN and a hemiplegia function assessment attention unit.

[0018] Among them, the scale evaluation text obtained in step S2 is first constructed into a hypergraph by the hypergraph construction module based on the core classification rules of stroke ICF, then the features are extracted by the hypergraph convolution network HyperGCN, and the scale features are output by the hemiplegia function assessment attention unit.

[0019] S5, input the symptom features and scale features obtained in steps S3 and S4 into the disease assessment module to output the patient disease assessment features, perform feature classification, and complete the disease assessment;

[0020] The disease assessment module comprises a selection attention fusion unit and a feature classification module.

[0021] The symptom features and scale features obtained in steps S3 and S4 are input into a selection attention fusion unit for feature fusion, and patient condition assessment features are output, and the features are classified by a feature classification module.

[0022] The condition assessment module uses cross entropy as a loss function Loss, and the expression is as follows:

[0023]

[0024] where Batch represents the number of batch samples selected in batch stochastic gradient descent, G t represents the calculated prediction probability distribution of the tth sample, g t represents the true probability distribution of the tth sample. The condition assessment module uses a backpropagation algorithm to learn the network, updates the parameters by gradient descent, and iterates in turn until the loss function converges, and sets the hyperparameters learning rate β and batch_size.

[0025] Further, the step S3 is specifically as follows:

[0026] S31, graph relationship modeling of patient symptom text information is performed by a symptom graph construction module;

[0027] First, the symptom text information is segmented and standardized, and then entity recognition is performed to identify symptom text entities, causal connection word entities, daily activity text entities, and degree adverb entities.

[0028] Then, four rules are designed to convert the entity relationship of the symptom text information into structured information and construct a text graph, and the construction rules of the graph structure are as follows:

[0029] (1) Symptom recognition rule construction: symptom entity graph;

[0030] Node: symptom text entity. Edge: undirected edge, representing the coexistence relationship between symptoms.

[0031] Construction method: extract all symptom text entities from the symptom text, and each entity is a node. If two symptoms appear in the same text paragraph, a edge is created between the corresponding nodes.

[0032] (2) Symptom causal rule construction: symptom causal graph;

[0033] Node: symptom text entity. Edge: directed edge, representing the direct causal relationship.

[0034] Construction method: extract all symptom text entities and causal connection word entities from the symptom text, and each entity is a node. If the symptom text entity appears before and after the causal connection word entity in a sentence, create an edge between the symptom text entity nodes.

[0035] (3) Symptom impact rule construction: symptom impact graph;

[0036] Node: symptom text entity, daily activity text entity. Edge: directed edge, indicating direct symptom impact on daily life.

[0037] Construction method: extract all symptom text entities and daily activity text entities from the symptom text, and each entity is a node. If the symptom text entity appears before the daily activity text entity in a sentence, create an edge between the symptom text entity node and the daily activity text entity node.

[0038] (4) Symptom degree rule construction: symptom degree graph;

[0039] Node: symptom text entity, degree adverb entity. Edge: directed edge, indicating the combination relationship between the symptom and the degree adverb.

[0040] Construction method: identify the combination relationship between the symptom text entity and the degree adverb entity, and if the symptom text entity and the degree adverb entity appear in the same sentence, create an edge between the corresponding nodes.

[0041] S32, based on the graph relationship constructed in step S31, design an adaptive edge weight adjuster to adjust the edge weight to obtain the optimal weight according to the needs of the feature classification task in the downstream disease condition evaluation module;

[0042] The initial value of the edge weight is determined by randomly sampling from a uniform distribution of [0, 1], and for the edge e between node i and node j ij , the learning process expression of the weight is as follows:

[0043]

[0044] Where ReLU(·) = max(0, ·) is the activation function; a ij is a learnable parameter.

[0045] S33, construct a symptom analysis unit, combine GCN and BERT, integrate the association information of the entity relationship structure graph and the character-level semantic details provided by the pre-trained BERT model through the attention mechanism, and output the symptom features;

[0046] The symptom analysis unit comprises a GCN module, a BERT module, an attention mechanism fusion module, and a feature fusion module.

[0047] The GCN module is used to extract the entity relationship structure information obtained after step S32, to obtain the feature representation L GCN , and the expression is as follows:

[0048] L GCN = GCN(G) (3)

[0049] wherein G represents the graph structure adjusted by the adaptive edge weight adjuster, and step S31 constructs four kinds of graphs, so the symptom analysis unit needs to be executed four times, and each time G corresponds to a different graph.

[0050] The BERT module is used to encode the symptom text obtained in step S2 to obtain the context embedding representation f BERT of each word, and the expression is as follows:

[0051] f BERT = BERT(Input) (4)

[0052] wherein Input represents the symptom text input after tokenization. The feature vector at the [CLS] position is extracted from the BERT output as the semantic representation F cls of the entire text, and the expression is as follows:

[0053] F cls = f BERT [:,0,:] (5)

[0054] wherein [:,0,:] represents the feature vector at the first position of the BERT output, which is used as the semantic representation of the entire text.

[0055] Then the attention mechanism fusion module takes L GCN and f BERT after weight matrix processing as query matrix Q G and key matrix K B respectively, and value matrix V is set to f BERT , and the weighted semantic representation is calculated through the attention mechanism, and the expression is as follows:

[0056] Q G = L GCN W Q ,K B = f BERT W k ,V = f BERT (6)

[0057]

[0058] wherein W Q and W k represent weight matrices, and dk denotes the dimension of the key vector, a denotes the feature representation of fusing the structure information extracted by GCN and the semantic information extracted by BERT, and T denotes the transpose operation.

[0059] The semantic representation F cls of BERT is further fused with the attention-fused feature a through a feature fusion module to obtain the final feature representation logits s , which is expressed as follows:

[0060] logits s = MLP(m·F cls +(1-m)·a) (8)

[0061] where F cls denotes the output feature of BERT, m denotes an interpolation parameter, and 0≤m≤1.

[0062] S34, fusing the feature representation obtained in step S33 through a correlation semantic representation fusion module to output a symptom feature;

[0063] After passing through the feature fusion module, four symptom correlation semantic feature representations logits s are obtained, and the correlation semantic representation fusion module generates a symptom feature matrix representation of the patient symptom text by splicing logits s , which is expressed as follows:

[0064]

[0065] where || denotes a connection operator for splicing features, s denotes an identification serial number of the symptom correlation semantic feature representation, H∈R 1×c×4 denotes a symptom feature matrix, R denotes a real number field, and c denotes a feature dimension.

[0066] Further, the step S4 is specifically as follows:

[0067] S41, constructing a hypergraph based on the stroke ICF core classification rule;

[0068] First, the standard evaluation scale used is constructed into a hypergraph based on the stroke ICF core classification rule, i.e., directly according to the stroke ICF core classification rule, the scale items are divided into four dimensions: body function assessment C bf , body structure assessment C bs , activity participation assessment C ap , and environmental factor assessment C ef . For each dimension r∈{bf, bs, ap, ef}, G r =(V r ,E r ,Wr ).

[0069] Among them, V r Represents the node set under this dimension, corresponding to the items in the scale; E r Represents the hyperedge set under this dimension, corresponding to the set of items with certain correlation in the scale; W r :E r →R + represents the weight function, i.e., the associated weight of the hyperedge, R + represents the set of positive real numbers.

[0070] For each item v i and v j , calculate their similarity sim(v i ,v j ), that is, calculating the cosine similarity of the feature vector, the expression is as follows:

[0071]

[0072] Among them, x i ,x j Indicates project v i ,v j The eigenvector of .

[0073] Then set a similarity threshold to determine whether to add a hyperedge. If the similarity between two items is greater than the threshold, they will be added to the hyperedge. For each dimension r, traverse all items in that dimension, generate hyperedges based on the similarity between items, and combine the hyperedges to form the hyperedge set E of that dimension. r .

[0074] For each hyperedge, calculate the hyperedge weight w(e), which is expressed as follows:

[0075]

[0076] Here, |e| represents the number of items in hyperedge e.

[0077] S42, using the hypergraph convolutional network HyperGCN to perform feature extraction on the hypergraph obtained in step S41;

[0078] The four-dimensional hypergraph G constructed in step S41 r =[G bf ,G bs ,G ap ,G ef ] Input HyperGCN for feature extraction and obtain the feature representation Z of four dimensions r , the expression is as follows:

[0079] Zr =HyperGCN(G r ) (12)

[0080] Among them, Z r =[Z bf ,Z bs ,Z ap ,Z ef ],Z bf ∈R n×d ,Z bs ∈R n×d ,Z ap ∈R n×d ,Z ef ∈R n×d , n represents the number of hypergraph nodes, and d represents the feature dimension.

[0081] S43, inputting the features extracted in step S42 into the designed hemiplegia function assessment attention unit for feature enhancement, and finally outputting the enhanced scale features;

[0082] First, calculate the query vector Query and the key vector Key, and then transform the body function evaluation hypergraph feature Z bf and body structure evaluation hypergraph feature Z bs Add and combine them as internal evaluation factor features, and generate the query matrix Q and key matrix K of the internal factor features through linear transformation. The expressions are as follows:

[0083] Q=(Z bf +Z bs )W q1 , K=(Z bf +Z bs )W k (13)

[0084] Among them, W q1 ,W k ∈R d×d represents the learnable weight matrix.

[0085] Then calculate the global Query weight, the expression is as follows:

[0086]

[0087] Among them, w a ∈R d represents a learnable vector, γ∈R n represents the global attention weight.

[0088] Then the activity participates in evaluating the hypergraph feature Z ap and environmental factors to evaluate the hypergraph feature Z efAdd the combination as external evaluation factor features, after dimension reduction through the full connection layer, and add and fuse the internal evaluation factor query vector calculated through the global Query weight, to obtain the global Query vector q' ∈ R d , The expression is as follows:

[0089]

[0090] Where, i is an index variable, used to traverse all feature vectors, n represents the number of feature vectors, FC(·) represents the full connection layer operation.

[0091] Let the global Query vector and the key Key matrix interact to generate the global context representation through element-by-element multiplication calculation, and the expression is as follows:

[0092] C=K⊙q′ (16)

[0093] Where, ⊙ represents element-by-element multiplication, C ∈ R n×d Represents the global context representation.

[0094] Finally, the output quantity table feature X is calculated as follows:

[0095] X=Φ(C) (17)

[0096] Where, Φ represents a linear transformation, used to further extract features.

[0097] Further, the step S5 is specifically as follows:

[0098] S51, the symptom features and the quantity table features obtained in steps S3 and S4 are dimensionally unified using full connection operation to obtain the quantity table feature representation and the symptom feature representation;

[0099] The calculation expression of the quantity table feature representation Y X And the symptom feature representation Y H Is as follows:

[0100] Y X =FC X (X) (18)

[0101] Y H =FC H (H) (19)

[0102] Where, Y X ∈R n×l , Y H ∈R n×l , FC X (·) and FC H (·) represent the full connection layer operation performed on the two kinds of features, and l represents the feature dimension.

[0103] S52, the scale feature representation Y X , the symptom feature representation Y H , the input is designed to select attention fusion unit for feature fusion, and the output is a patient condition assessment feature;

[0104] First, the scale feature representation Y X , the symptom feature representation Y H , the feature Y Θ is obtained by bit-by-bit addition fusion Θ , the input of the selection attention gate unit SAGU in the selection attention fusion unit is obtained, and the corresponding feature output O is obtained, and the expression is as follows:

[0105] O=SAGU(Y Θ ) (20)

[0106] Where, SAGU(·) represents the selection attention gate unit, and O represents the output feature representation of SAGU.

[0107] SAGU inputs Y Θ respectively to different fully connected layers to perform fully connected operation, and then performs activation to obtain corresponding feature matrix U, L, Z, which are respectively used to capture the relationship of different aspects of the feature and provide input for the attention mechanism, and the expression is as follows:

[0108] U=φ U (Y Θ W U ), L=φ L (Y Θ W L ), Z=φ Z (Y Θ W Z ) (21)

[0109] Where, W U ,W L ∈R l×j , W Z ∈R l×Ω is a learnable weight matrix, Ω represents the head_size of attention, φ U , φ L , φ Z represent the activation function, and j represents the dimension of the intermediate feature, which is used to control the dimension of the transformed feature.

[0110] Then the SAGU output feature can be represented as follows:

[0111] O=(U⊙AL)W o (22)

[0112] Where, W o∈R l×|e| is a learnable weight matrix. A ∈ R n×n represents an Attention matrix, which is responsible for fusing information between tokens, and the specific expression is as follows:

[0113]

[0114] wherein, represents an affine transformation, relu represents an activation function, and 1 / n represents a normalization factor.

[0115] Finally, the output I of the selection attention fusion unit is obtained in combination with SAGU G , and the expression is as follows:

[0116] I G = ((SAGU(Y Θ )+Y H )+Y X ) (24)

[0117] wherein, I G represents the patient condition evaluation feature extracted by the model.

[0118] S53, based on step S52, performing feature classification through a feature classification module to complete a condition evaluation classification task;

[0119] The patient condition evaluation feature I G is fed into a fully connected layer, and then the output of the fully connected layer is sent to a Softmax function for classification to obtain probability values p i of different condition symptom stages, and finally the symptom stage with the highest probability is the diagnosis and evaluation result y of the hemiplegia attention graph convolutional neural network framework HemiplegiaAGCN, and the expression is as follows:

[0120] q=W·I G +b (25)

[0121]

[0122] y=argmax 1≤i≤τ (p i ) (27)

[0123] wherein, q represents a feature vector obtained through a fully connected layer, W represents a weight matrix of the fully connected layer, I G represents a patient condition evaluation feature, and b represents a bias vector. q i represents the i-th element of the fully connected layer output, i.e., the unnormalized score of the i-th class, τ=3, q k represents the k-th element of the fully connected layer output. p idenotes the predicted probability value of the i-th class. argmax 1≤i≤τ (p i ) denotes the index of selecting the maximum value in the probability distribution p to select the highest probability of the symptom staging category.

[0124] Advantages of the present application: The method of the present application first constructs a hemiplegia AGCN (Attention Graph Control Network) framework, cleans and classifies patient clinical examination evaluation text data, respectively designs a symptom text analysis module and an assessment scale analysis module to process symptom text and scale evaluation text, analyzes the features of the symptom text and scale evaluation text, and then uses the selected attention fusion unit in the disease evaluation module to fuse the features, and finally performs classification evaluation. The hemiplegia AGCN (Attention Graph Control Network) designed by the method of the present application processes patient clinical examination evaluation text data from different angles, extracts the most relevant features of the patient's condition, and performs feature fusion to perform the disease evaluation classification task, which helps to improve the accuracy of hemiplegia evaluation, assists doctors in rehabilitation evaluation, improves doctor efficiency, reduces misdiagnosis rate, and improves medical resource utilization rate. BRIEF DESCRIPTION OF DRAWINGS

[0125] Figure 1 A flowchart of a hemiplegia rehabilitation evaluation method based on an attention graph control network for stroke according to the present application.

[0126] Figure 2 A hemiplegia AGCN (Attention Graph Control Network) framework structure diagram in an embodiment of the present application. DETAILED DESCRIPTION

[0127] The method of the present application will be further described below in conjunction with the drawings and examples.

[0128] As shown in the flowchart of a hemiplegia rehabilitation evaluation method based on an attention graph control network for stroke according to the present application, the specific steps are as follows: Figure 1 S1, a hemiplegia AGCN (Attention Graph Control Network) framework is constructed;

[0129] As shown in the flowchart of a hemiplegia rehabilitation evaluation method based on an attention graph control network for stroke according to the present application, the specific steps are as follows:

[0130] Figure 2 As shown, the HemiplegiaAGCN includes a symptom text analysis module, an assessment scale analysis module, and a disease evaluation module.

[0131] S2, data cleaning and classification are performed on the clinical examination evaluation text data to obtain symptom text and scale evaluation text; ​

[0132] The collected clinical examination evaluation text data is cleaned and classified, and the description of the current physical symptoms of the patient is divided into symptom text, and the quantitative evaluation of the physical condition of the patient is divided into scale evaluation text.

[0133] The patient clinical examination evaluation text data used in this embodiment is shown in Table 1. It can be found that the clinical examination evaluation data is composed of two types of data. One is the description of the current physical symptoms of the patient, including: “The patient has right hemiplegia, upper limb function is limited, common movement mode is dominant, the patient's daily life ability is limited, moderate dependence, and needs help.” The other is the quantitative evaluation of the physical condition of the patient, including: “Modified Ashworth classification: ankle plantar flexor muscle tension 1 degree. Brunnstrom stage: lower extremity stage III.” The two types of data actually describe the patient's condition from different angles. The symptom text is often more specific and detailed, while the quantitative evaluation text often only describes the situation through related scores.

[0134] Table 1

[0135]

[0136] Among them, the symptom text adopts descriptive vocabulary, including: “patient”, “patient”, “cannot”, “limited”, “pain”; the scale evaluation text adopts quantitative score keywords, including: “score”, “classification”, “scale”, “stage”. The two types of data describe the clinical manifestations of the patient from different angles. In this embodiment, the two types of data are divided for subsequent model targeted processing. The divided patient clinical data is shown in Table 2.

[0137] Table 2

[0138]

[0139] S3, input the symptom text obtained in step S2 into the symptom text analysis module of the framework constructed in step S1, and output symptom features;

[0140] As shown in Figure 2 , the symptom text analysis module includes a symptom graph construction module, an adaptive edge weight adjuster, a symptom analysis unit, and a related semantic representation fusion module.

[0141] Among them, the symptom text obtained in step S2 is first subjected to graph relationship modeling by the symptom graph construction module, the adaptive edge weight adjuster is used to adjust the edge weight to obtain the optimal weight, and then the symptom analysis unit is used to output the symptom features and the related semantic representation fusion module is used for fusion.

[0142] S4, input the scale evaluation text obtained in step S2 into the evaluation scale analysis module of the framework constructed in step S1, and output scale features;

[0143] As shown in the Figure 2 The assessment scale analysis module comprises a hypergraph construction module based on the stroke ICF core classification rule, a hypergraph convolution network HyperGCN and a hemiplegia function evaluation attention unit.

[0144] The scale evaluation text obtained in step S2 is first subjected to hypergraph construction by the hypergraph construction module based on the stroke ICF core classification rule, then subjected to feature extraction by the hypergraph convolution network HyperGCN, and finally subjected to output of scale features by the hemiplegia function evaluation attention unit.

[0145] S5, inputting the symptom features and the scale features obtained in steps S3 and S4 into a disease condition evaluation module to output patient disease condition evaluation features, performing feature classification, and completing disease condition evaluation;

[0146] As shown in the Figure 2 The disease condition evaluation module comprises a selection attention fusion unit and a feature classification module.

[0147] The symptom features and the scale features obtained in steps S3 and S4 are inputted into the selection attention fusion unit to perform feature fusion, output patient disease condition evaluation features, and perform feature classification by the feature classification module.

[0148] The disease condition evaluation module uses cross entropy as a loss function Loss, and the expression is as follows:

[0149]

[0150] Wherein, Batch represents the number of batch samples selected in batch stochastic gradient descent, G t represents the calculated prediction probability distribution of the tth sample, g t represents the true probability distribution of the tth sample. The disease condition evaluation module adopts a back-propagation algorithm to learn the network, updates the parameters by gradient descent, and iterates in turn until the loss function converges. In this embodiment, the hyperparameters learning rate β = 0.001 and batch_size = 50 are set.

[0151] In this embodiment, step S3 is specifically as follows:

[0152] S31, performing graph relationship modeling of patient symptom text information by a symptom graph construction module;

[0153] In order to model the language level features of patient symptom text information and capture deep semantic information, the symptom text information is first subjected to word segmentation and standardization processing, and then subjected to entity recognition to identify symptom text entities, causal connection word entities, daily activity text entities and degree adverb entities.

[0154] Then four rules are designed to convert the entity relationship of the symptom text information into structured information and build a text graph, so as to enhance the understanding of the model to the symptom text. The construction rules of the graph structure are as follows:

[0155] (1) Symptom recognition rule construction: symptom entity graph;

[0156] Node: symptom text entity, including: “right limb hemiplegia”, “upper limb function limited”. Edge: undirected edge, representing the coexistence relationship between symptoms.

[0157] Construction method: extract all symptom text entities from the symptom text, and each entity is a node. If two symptoms appear in the same text paragraph, a edge is created between the corresponding nodes.

[0158] (2) Symptom causality rule construction: symptom causality graph;

[0159] Node: symptom text entity, including: “shoulder and elbow control is poor”, “unable to use the affected hand to wash the healthy side upper limb”. Edge: directed edge, representing the direct causal relationship.

[0160] Construction method: extract all symptom text entities and causal conjunction word entities from the symptom text, and each entity is a node. If the symptom text entity appears before and after the causal conjunction word entity in a sentence, a edge is created between the symptom text entity nodes.

[0161] (3) Symptom impact rule construction: symptom impact graph;

[0162] Node: symptom text entity, daily activity text entity, including: “hand utility decreased”, “unable to complete bathing independently”. Edge: directed edge, representing the direct impact of symptoms on daily life.

[0163] Construction method: extract all symptom text entities and daily activity text entities from the symptom text, and each entity is a node. If the symptom text entity appears before the daily activity text entity in a sentence, a edge is created between the symptom text entity node and the daily activity text entity node.

[0164] (4) Symptom degree rule construction: symptom degree graph;

[0165] Node: symptom text entity, including: “hemiplegia”; degree adverb entity, including: “weak strength”, “low grip strength”. Edge: directed edge, representing the combination relationship between symptoms and degree adverbs.

[0166] Construction method: identify the combination relationship between symptom text entities and degree adverb entities, if the symptom text entity and the degree adverb entity appear in the same sentence, then create an edge between the corresponding nodes.

[0167] S32, based on the graph relationship constructed in step S31, the adaptive edge weight adjuster adjusts the edge weight to obtain the optimal weight according to the needs of the feature classification task in the downstream disease condition evaluation module;

[0168] The initial value of the edge weight is determined by randomly sampling from a uniform distribution of [0, 1], and for the edge e between node i and node j ij , the weight learning process expression is as follows:

[0169]

[0170] where ReLU(·) = max(0, ·) is an activation function to ensure non-negative weights.a ij is a learnable parameter.

[0171] S33, construct a symptom analysis unit, combine GCN and BERT, integrate the associated information of the entity relationship structure graph and the character-level semantic details provided by the pre-trained BERT model through an attention mechanism, and output symptom features;

[0172] In order to model the semantic relationship between different levels of information in the symptom text and further understand the real clinical situation of the patient, a symptom analysis unit is designed. This unit combines the advantages of GCN and BERT, and integrates the associated information of the entity relationship structure graph and the character-level semantic details provided by the pre-trained BERT model through an attention mechanism.

[0173] The symptom analysis unit comprises a GCN module, a BERT module, an attention mechanism fusion module, and a feature fusion module.

[0174] The GCN module is used to extract the entity relationship structure information obtained after step S32, to obtain the feature representation L GCN , the expression is as follows:

[0175] L GCN = GCN(G) (3)

[0176] Where G represents the graph structure adjusted by the adaptive edge weight adjuster, and step S31 constructs four graphs, so the symptom analysis unit needs to be executed four times, and G corresponds to different graphs each time.

[0177] The BERT module is used to encode the symptom text obtained in step S2 to obtain the context embedding representation f BERT of each word, the expression is as follows:

[0178] f BERT = BERT (Input) (4)

[0179] where Input denotes the symptom text input after tokenization. The feature vector at the [CLS] position is extracted from the output of BERT as the semantic representation F cls of the whole text, and the expression is as follows:

[0180] F cls = f BERT [:, 0, :] (5)

[0181] where [:, 0, :] denotes the feature vector at the first position (i.e., the [CLS] position) of the BERT output, which is used as the semantic representation of the whole text.

[0182] Then the attention mechanism fusion module will L GCN and f BERT after weight matrix processing respectively as query matrix Q G and key matrix K B , value matrix V is set to f BERT , through attention mechanism calculation weighted semantic representation, expression as follows:

[0183] Q G = L GCN W Q , K B = f BERT W k , V = f BERT (6)

[0184]

[0185] where W Q and W k denote the weight matrix, d k denotes the dimension of the key vector, a denotes the feature representation of the structural information extracted by the fusion GCN and the semantic information extracted by BERT, and T denotes the transpose operation.

[0186] Then through the feature fusion module BERT semantic representation F cls and the feature a after attention fusion are linearly combined to obtain the final feature representation logits s , and the expression is as follows:

[0187] logits s = MLP (m · F cls + (1-m) · a) (8)

[0188] where F clswherein y represents the output feature of BERT (extracted from the [CLS] position), m represents an interpolation parameter, and 0≤m≤1.

[0189] S34, the feature representation obtained in step S33 is fused by an associated semantic representation fusion module, and a symptom feature is output;

[0190] After passing through the feature fusion module, four symptom associated semantic feature representations logits s are obtained. s The associated semantic representation fusion module generates a symptom feature matrix representation of the patient's symptom text, and the expression is as follows:

[0191]

[0192] wherein || represents a connection operator for splicing features, s represents an identification serial number of the symptom associated semantic feature representation, H∈R 1×c×4 represents a symptom feature matrix, R represents a real number field, and c represents a feature dimension.

[0193] In the embodiment, the step S4 is specifically as follows:

[0194] S41, construction of a hypergraph based on a stroke ICF core classification rule;

[0195] In an actual disease assessment task, many different standard assessment scales are used, but there is a large amount of redundant information between the assessment scale items, which makes it difficult to effectively extract accurate information. The present embodiment proposes an assessment scale analysis module. First, the standard assessment scale used is constructed into a hypergraph based on the stroke ICF core classification rule, i.e., the scale items are directly divided into four dimensions according to the stroke ICF core classification rule: body function assessment C bf , body structure assessment C bs , activity participation assessment C ap , and environmental factor assessment C ef . For each dimension r∈{bf, bs, ap, ef}, G r =(V r ,E r ,W r ).

[0196] wherein V r represents a node set under the dimension, corresponding to the items in the scale; E rRepresents the hyperedge set under this dimension, corresponding to the set of items in the scale that have certain associations (the association here is defined as whether the items are centered around a specific function of the patient. For example, under the physical function assessment dimension, there are several items that describe the patient's muscle tension function. Then these items are associated and can be constructed into a hyperedge). In actual construction, the association between items is determined by calculating the similarity between them; W r :E r →R + represents the weight function, i.e., the associated weight of the hyperedge, R + represents the set of positive real numbers.

[0197] For each item v i and v j , calculate their similarity sim(v i ,v j ), that is, calculating the cosine similarity of the feature vector, the expression is as follows:

[0198]

[0199] Among them, x i ,x j Indicates project v i ,v j The eigenvector of .

[0200] Then set a similarity threshold to determine whether to add a hyperedge. If the similarity between two items is greater than the threshold, they will be added to the hyperedge. For each dimension r, traverse all items in that dimension, generate hyperedges based on the similarity between items, and combine the hyperedges to form the hyperedge set E of that dimension. r .

[0201] For each hyperedge, calculate the hyperedge weight w(e), which is expressed as follows:

[0202]

[0203] Here, |e| represents the number of items in hyperedge e.

[0204] S42, using the hypergraph convolutional network HyperGCN to perform feature extraction on the hypergraph obtained in step S41;

[0205] The four-dimensional hypergraph G constructed in step S41 r =[G bf ,G bs ,G ap ,G ef ] Input HyperGCN for feature extraction and obtain the feature representation Z of four dimensions r , the expression is as follows:

[0206] Z r = HyperGCN(G r ) (12)

[0207] where Z r = [Z bf , Z bs , Z ap , Z ef ], Z bf ∈ R n×d , Z bs ∈ R n×d , Z ap ∈ R n×d , Z ef ∈ R n×d , n denotes the number of hypergraph nodes, and d denotes the feature dimension.

[0208] S43, input the features extracted in step S42 into the designed hemiplegia function evaluation attention unit to enhance the features, and finally output the enhanced scale features;

[0209] First, calculate the query vector Query and the key vector Key, combine the body function evaluation hypergraph feature Z bf and the body structure evaluation hypergraph feature Z bs to obtain the internal evaluation factor feature, and generate the query Query matrix Q and the key Key matrix K of the internal factor feature through linear transformation, the expression is as follows:

[0210] Q = (Z bf + Z bs ) W q1 , K = (Z bf + Z bs ) W k (13)

[0211] where W q1 , W k ∈ R d×d denote the learnable weight matrix.

[0212] Then calculate the global Query weight, the expression is as follows:

[0213]

[0214] where w a ∈ R d denotes the learnable vector, and γ ∈ R n denotes the global attention weight.

[0215] Then combine the activity participation evaluation hypergraph feature Z apand environmental factor evaluation hypergraph feature Z ef Additive combination as external evaluation factor feature, after dimension reduction through full connection layer, add and fuse with internal evaluation factor query vector calculated through global query weight, to obtain global query vector q' ∈ R d , the expression is as follows:

[0216]

[0217] Wherein, i is an index variable, used to traverse all feature vectors, n represents the number of feature vectors (the same as the number of hypergraph nodes), FC(·) represents the full connection layer operation.

[0218] Let the global query vector and the key matrix interact to generate the global context representation through element-by-element multiplication calculation, the expression is as follows:

[0219] C=K⊙q′ (16)

[0220] Wherein, ⊙ represents element-by-element multiplication, C ∈ R n×d Represents the global context representation.

[0221] Finally, the output quantity table feature X is calculated as follows:

[0222] X=Φ(C) (17)

[0223] Wherein, Φ represents a linear transformation, used to further extract features.

[0224] In this embodiment, the step S5 is specifically as follows:

[0225] S51, the symptom feature and the quantity table feature obtained in steps S3 and S4 are dimensioned by full connection operation, to obtain the quantity table feature representation and the symptom feature representation;

[0226] The calculation expression of the quantity table feature representation Y X And the symptom feature representation Y H Is as follows:

[0227] Y X =FC X (X) (18)

[0228] Y H =FC H (H) (19)

[0229] Wherein, Y X ∈R n×l , Y H ∈R n×l , FC X (·) and FC H(·) represents a full connection layer operation performed on two features, and l represents the feature dimension.

[0230] S52, the scale feature representation Y X , the symptom feature representation Y H , the input is designed to select an attention fusion unit for feature fusion, and the output is a patient condition assessment feature;

[0231] First, the scale feature representation Y X , the symptom feature representation Y H , the feature Y Θ is obtained by bit-by-bit addition fusion Θ , the Y Θ is input into the select attention gated unit SAGU (Select Attention Gated Unit) in the select attention fusion unit, to obtain the corresponding feature output O, the expression is as follows:

[0232] O=SAGU(Y Θ ) (20)

[0233] Wherein, SAGU(·) represents the select attention gated unit, and O represents the output feature representation of SAGU.

[0234] The operation details of SAGU are as follows:

[0235] SAGU inputs Y Θ into different full connection layers respectively to perform full connection operation, and then performs activation to obtain the corresponding feature matrix U, L, Z, which are respectively used to capture the relationship of different aspects of the feature and provide input for the attention mechanism, the expression is as follows:

[0236] U=φ U (Y Θ W U ), L=φ L (Y Θ W L ), Z=φ Z (Y Θ W Z ) (21)

[0237] Wherein, W U ,W L ∈R l×j , W Z ∈R l×Ω are learnable weight matrices, Ω represents the head_size of attention, Ω=128 in the embodiment, φ U , φ L , φ Z represent the activation function, and j represents the dimension of the intermediate feature, which is used to control the dimension of the transformed feature.

[0238] The SAGU output feature can be represented as follows:

[0239] O = (U AL) W o (22)

[0240] where W o ∈ R l×|e| is a learnable weight matrix. A ∈ R n×n represents an Attention matrix responsible for fusing information between tokens, and the specific expression is as follows:

[0241]

[0242] where, represents an affine transformation. relu is a commonly used activation function to introduce nonlinearity, and relu 2 is a square operation on the output of relu, which can further amplify the nonlinearity effect. 1 / n is a simple normalization factor to eliminate the length effect.

[0243] Finally, the output I G of the selection attention fusion unit is obtained by combining SAGU, and the expression is as follows:

[0244] I G = ((SAGU(Y Θ )+Y H )+Y X ) (24)

[0245] where I G represents the patient condition assessment feature extracted by the model.

[0246] S53, based on step S52, feature classification is performed by a feature classification module to complete the condition assessment classification task;

[0247] The patient condition assessment feature I G is fed into a fully connected layer, and then the output of the fully connected layer is sent to a Softmax function for classification to obtain probability values p i of different condition symptom stages. Finally, the symptom stage with the highest probability is the diagnostic evaluation result y of the hemiplegia attention graph convolutional network framework HemiplegiaAGCN, and the expression is as follows:

[0248] q = W I G +b (25)

[0249]

[0250] y = argmax 1≤i≤τ (p i) (27)

[0251] wherein q represents a feature vector obtained through the fully connected layer, W represents a weight matrix of the fully connected layer, I G represents a patient condition evaluation feature, and b represents a bias vector. q i represents the i-th element of the fully connected layer output, i.e., the unnormalized score of the i-th class, τ = 3, q k represents the k-th element of the fully connected layer output. p i represents the probability value of the i-th class predicted. argmax 1≤i≤τ (p i ) refers to selecting the index of the maximum value in the probability distribution p to select the highest probability of the symptom staging class, y, i.e., the symptom staging evaluation result predicted by the model.

[0252] In summary, the method of the present application designs a hemiplegia condition attention graph control network to process patient clinical examination evaluation text data from different angles, extract the most relevant features of the patient condition, and perform feature fusion to perform condition evaluation classification tasks, which helps to improve the accuracy of hemiplegia condition evaluation, assist doctors in rehabilitation evaluation, improve doctor efficiency, reduce misdiagnosis rate, and improve the use rate of medical resources.

[0253] Those skilled in the art will realize that the embodiments described herein are for the purpose of helping the reader understand the principles of the present application and should be understood as not limiting the scope of protection of the present application to such specific statements and embodiments. Those skilled in the art can make various other specific modifications and combinations according to the technical inspiration disclosed in the present application without departing from the essence of the present application, and these modifications and combinations are still within the scope of protection of the present application.

Claims

1. A post-stroke hemiplegia rehabilitation assessment method based on an attention graph control network, the specific steps are as follows: S1. Construct HemiplegiaAGCN, an attention graph control network framework for hemiplegia. The HemiplegiaAGCN includes: Symptom text parsing module, rating scale parsing module, and condition assessment module; S2. Clean and classify the clinical examination assessment text data to obtain symptom text and scale assessment text; Collect clinical physical examination and assessment text data for data cleaning and classification, dividing the description of the patient's current physical symptoms into symptom text, and dividing the quantitative evaluation of the patient's physical condition into scale assessment text; The symptom text uses descriptive words, including "patient", "unable", "limited", and "pain"; the scale assessment text uses quantitative scoring keywords, including "score", "grading", "scale", and "stage"; S3. Input the symptom text obtained in step S2 into the symptom text parsing module of the framework constructed in step S1, and output symptom features; The symptom text parsing module includes: a symptom graph construction module, an adaptive edge weight regulator, a symptom parsing unit, and an associated semantic representation fusion module; The symptom text obtained in step S2 is first modeled using the symptom graph construction module, and the edge weights are adjusted using the adaptive edge weight regulator to obtain the optimal weights. The symptom features are then output using the symptom parsing unit and fused using the associated semantic representation fusion module. S4, inputting the scale evaluation text obtained in step S2 into the rating scale parsing module of the framework constructed in step S1, and outputting the scale features; The rating scale parsing module includes: a hypergraph construction module based on the ICF core classification rules of stroke, a hypergraph convolutional network HyperGCN, and a hemiplegia function assessment attention unit; The scale evaluation text obtained in step S2 is first constructed through a hypergraph construction module based on the core classification rules of the stroke ICF, and then features are extracted through the hypergraph convolutional network HyperGCN, and the scale features are output through the hemiplegia function assessment attention unit; S5. Input the symptom characteristics and scale characteristics obtained in steps S3 and S4 into the condition assessment module to output the patient's condition assessment characteristics, perform feature classification, and complete the condition assessment; The condition assessment module includes: a selection attention fusion unit and a feature classification module; The symptom features and scale features obtained in steps S3 and S4 are input into the attention fusion unit for feature fusion, and the patient's condition assessment features are output, and feature classification is performed through the feature classification module; The condition assessment module uses cross entropy as the loss function Loss, which is expressed as follows: Among them, Batch represents the number of batch samples selected in batch stochastic gradient descent, G t represents the calculated predicted probability distribution of the tth sample, g t Represents the true probability distribution of the tth sample; and the condition assessment module uses the back error propagation algorithm to learn the network, updates the parameters by gradient descent, iterates in sequence until the loss function converges, and sets the hyperparameters learning rate β and batch_size.

2. A post-stroke hemiplegia rehabilitation assessment method based on attention graph control network according to claim 1, characterized in that: The step S3 is specifically as follows: S31. Modeling the graph relationship of patient symptom text information through a symptom graph construction module; First, the symptom text information is segmented and standardized, and then entity recognition is performed to identify symptom text entities, causal connective entities, daily activity text entities, and degree adverb entities. Then, four rules are designed to convert the entity relationships of the symptom text information into structured information and construct a text graph. The rules for constructing the graph structure are as follows: (1) Construction of symptom identification rules: symptom entity graph; Node: symptom text entity; Edge: undirected edge, indicating the coexistence relationship between symptoms; Construction method: Extract all symptom text entities from the symptom text, and use each entity as a node; if two symptoms appear in the same text paragraph, create an edge between the corresponding nodes; (2) Construction of symptom causal rules: symptom causal diagram; Node: symptom text entity; Edge: directed edge, indicating direct causal relationship; Construction method: Extract all symptom text entities and causal connective entities from the symptom text, with each entity as a node; if a symptom text entity appears before or after a causal connective entity in a sentence, create an edge between the symptom text entity nodes; (3) Construction of symptom impact rules: symptom impact diagram; Nodes: Symptom text entities, daily activity text entities; Edges: directed edges, representing the direct impact of symptoms on daily life; Construction method: Extract all symptom text entities and daily activity text entities from the symptom text, with each entity as a node; if a symptom text entity appears before a daily activity text entity in a sentence, create an edge between the symptom text entity node and the daily activity text entity node; (4) Symptom degree rule construction: symptom degree diagram; Nodes: Symptom text entity, degree adverb entity; Edges: Directed edges, representing the combination relationship between symptoms and degree adverbs; Construction method: Identify the combination relationship between the symptom text entity and the degree adverb entity. If the symptom text entity and the degree adverb entity appear in the same sentence, create an edge between the corresponding nodes. S32. Based on the graph relationship constructed in step S31, an adaptive edge weight regulator is designed to adaptively adjust the edge weights to obtain the optimal weight according to the needs of the feature classification task in the downstream disease assessment module; The initial value of the edge weight is determined by random sampling from a uniform distribution in [0,1]. For the edge e between node i and node j, ij , its weight The learning process expression is as follows: Among them, ReLU(·)=max(0,·) is the activation function; a ij is a learnable parameter; S33. Build a symptom parsing unit, combining GCN and BERT, integrating the association information of the entity relationship structure graph with the character-level semantic details provided by the pre-trained BERT model through the attention mechanism, and output symptom features; The symptom analysis unit includes: GCN module, BERT module, attention mechanism fusion module, and feature fusion module; The GCN module is used to extract the entity relationship structure information obtained after step S32 and obtain the feature representation L GCN , the expression is as follows: L GCN =GCN(G) (3) Wherein, G represents the graph structure after being adjusted by the adaptive edge weight regulator, and step S31 constructs four graphs, so the symptom analysis unit needs to be executed four times, and each time G corresponds to a different graph; The BERT module is used to encode the symptom text obtained in step S2 to obtain the contextual embedding representation f of each word. BERT , the expression is as follows: f BERT =BERT(Input) (4) Among them, Input represents the symptom text input after word segmentation; the feature vector of the [CLS] position is extracted from the output of BERT as the semantic representation F of the entire text cls , the expression is as follows: F cls =f BERT [:,0,:] (5) Among them, [:,0,:] means taking the feature vector at the first position of BERT output, which is used as the semantic representation of the entire text; Then the attention mechanism fusion module will L GCN and f BERT After being processed by the weight matrix, they are respectively used as the query matrix Q G and bond matrix K B , the value matrix V is set to f BERT , the weighted semantic representation is calculated through the attention mechanism, and the expression is as follows: Q G =L GCN W Q ,K B =f BERT W k ,V=f BERT (6) Among them, W Q and W k represents the weight matrix, d k represents the dimension of the key vector, α represents the feature representation that integrates the structural information extracted by GCN and the semantic information extracted by BERT, and T represents the transposition operation; Then, the semantic representation F of BERT is transformed into cls Linearly combine the feature α after attention fusion to obtain the final feature representation logits s , the expression is as follows: logits s =MLP(m·F cls +(1-m)·α) (8) Among them, F cls represents the output feature of BERT, m represents an interpolation parameter, and 0≤m≤1; S34, fusing the feature representation obtained in step S33 through an associated semantic representation fusion module to output symptom features; After the feature fusion module, four symptom-related semantic feature representations logits are obtained s , the input association semantic representation fusion module is constructed by concatenating logits s Generate the symptom feature matrix representation of the patient's symptom text, the expression is as follows: Among them, || represents the connection operator used to splice features, s represents the identification number of the symptom-related semantic feature representation, H∈R 1 ×c×4 Represents the symptom feature matrix, R represents the real number domain, and c represents the feature dimension.

3. A post-stroke hemiplegia rehabilitation assessment method based on attention graph control network according to claim 1, characterized in that: The step S4 is specifically as follows: S41. Hypergraph construction based on ICF core classification rules for stroke; Firstly, a hypergraph was constructed for the standard assessment scale based on the ICF core classification rules of stroke, that is, the scale items were divided into four dimensions directly according to the ICF core classification rules of stroke: physical function assessment C bf , body structure assessment C bs , activity participation evaluation C ap , Environmental Factor Assessment C ef ; For each dimension r∈{bf,bs,ap,ef}, G r =(V r ,E r ,W r ); Among them, V r Represents the node set under this dimension, corresponding to the items in the scale; E r Represents the hyperedge set under this dimension, corresponding to the set of items with certain correlation in the scale; W r :E r →R + represents the weight function, i.e., the associated weight of the hyperedge, R + represents the set of positive real numbers; For each item v i and v j , calculate their similarity sim(v i ,v j ), that is, calculating the cosine similarity of the feature vector, the expression is as follows: Among them, x i ,x j Indicates project v i ,v j The eigenvector of Then set a similarity threshold to determine whether to add a hyperedge. If the similarity between two items is greater than the threshold, they will be added to the hyperedge. For each dimension r, traverse all items in that dimension, generate hyperedges based on the similarity between items, and combine the hyperedges to form the hyperedge set E of that dimension. r ; For each hyperedge, calculate the hyperedge weight w(e), which is expressed as follows: Where |e| represents the number of items in hyperedge e; S42, using the hypergraph convolutional network HyperGCN to perform feature extraction on the hypergraph obtained in step S41; The four-dimensional hypergraph G constructed in step S41 r =[G bf ,G bs ,G ap ,G ef ] Input HyperGCN for feature extraction and obtain the feature representation Z of four dimensions r , the expression is as follows: FROM r =HyperGCN(G r ) (12) Among them, Z r =[Z bf ,Z bs ,Z ap ,Z ef ],Z bf ∈R n×d ,Z bs ∈R n×d ,Z ap ∈R n×d ,Z ef ∈R n×d , n represents the number of hypergraph nodes, d represents the feature dimension; S43, inputting the features extracted in step S42 into the designed hemiplegia function assessment attention unit for feature enhancement, and finally outputting the enhanced scale features; First, calculate the query vector Query and the key vector Key, and then transform the body function evaluation hypergraph feature Z bf and body structure evaluation hypergraph feature Z bs Add and combine them as internal evaluation factor features, and generate the query matrix Q and key matrix K of the internal factor features through linear transformation. The expressions are as follows: Q=(Z bf +Z bs )IN q1 ,K=(Z bf +Z bs )IN k (13) Among them, W q1 ,W k ∈R d×d represents the learnable weight matrix; Then calculate the global Query weight, the expression is as follows: Among them, w a ∈R d represents a learnable vector, γ∈R n represents the global attention weight; Then the activity participates in evaluating the hypergraph feature Z ap and environmental factors to evaluate the hypergraph feature Z ef The sum of the external evaluation factors is combined and fused with the internal evaluation factor query vector calculated by the global Query weight after dimensionality reduction in the fully connected layer to obtain the global Query vector q′∈R d , the expression is as follows: Where i is an index variable used to traverse all feature vectors, n represents the number of feature vectors, and FC(·) represents the fully connected layer operation; Then, the global Query vector and the key matrix are multiplied element by element to generate the global context representation, which is expressed as follows: C=K⊙q′ (16) Where ⊙ represents element-wise multiplication, C∈R n×d Represents the global context representation; Finally, the scale feature X is output, and the calculation expression is as follows: X=Φ(C) (17) Among them, Φ represents a linear transformation, which is used to further extract features.

4. A post-stroke hemiplegia rehabilitation assessment method based on attention graph control network according to claim 1, characterized in that: The step S5 is specifically as follows: S51, unifying the dimensions of the symptom features and scale features obtained in steps S3 and S4 using a full-connect operation to obtain scale feature representation and symptom feature representation; Scale characteristic representation Y X and symptom characteristics indicate Y H The calculation expression is as follows: Y X =FC X (X) (18) Y H =FC H (H) (19) Among them, Y X ∈R n×l , Y H ∈R n×l , FC X (·) and FC H (·) represents the fully connected layer operation performed on two features, and l represents the feature dimension; S52. Express the scale characteristics as Y X , symptom characteristics indicate Y H , input the designed selective attention fusion unit for feature fusion, and output the patient's condition assessment features; First, the scale characteristic Y X , symptom characteristics indicate Y H The feature Y is obtained by bit-by-bit addition fusion Θ , change Y Θ Input the selective attention gating unit SAGU in the selective attention fusion unit to obtain the corresponding feature output O, which is expressed as follows: O=SAGU(Y Θ ) (20) Among them, SAGU(·) represents the selective attention gating unit, and O represents the output feature representation of SAGU; SAGU will enter the Y Θ Input different fully connected layers to perform fully connected operations, and then activate to obtain the corresponding feature matrices U, L, and Z, which are used to capture the relationship between different aspects of the features and provide input for the attention mechanism. The expressions are as follows: U=φ U (Y Θ W U ),L=φ L (Y Θ W L ),Z=φ Z (Y Θ W Z ) (21) Among them, W U ,W L ∈R l×j , W Z ∈R l×Ω is a learnable weight matrix, Ω represents the head_size of attention, φ U ,φ L ,φ Z represents the activation function, j represents the dimension of the intermediate feature, which is used to control the dimension after feature transformation; Then the SAGU output features can be expressed as follows: O=(U⊙AL)W o (22) Among them, W o ∈R l×|e| is a learnable weight matrix; A∈R n×n Represents the Attention matrix, which is responsible for fusing information between tokens. The specific expression is as follows: in, represents affine transformation, relu represents activation function, and 1 / n represents normalization factor; Finally, we combine SAGU to get the output of the selective attention fusion unit I G , the expression is as follows: Yo G =((SAGU(Y Θ )+Y H )+Y X ) (24) Among them, I G Represents the patient condition assessment features extracted by the model; S53: Based on step S52, feature classification is performed through a feature classification module to complete the disease assessment classification task; Patient's condition assessment characteristics I G Feed it to the fully connected layer, and then send the output of the fully connected layer to the Softmax function for classification to obtain the probability value p of different disease symptom stages i , the symptom stage with the highest probability is the diagnostic evaluation result y of the hemiplegia attention graph control network framework HemiplegiaAGCN, which is expressed as follows: q=W·I G +b (25) y=argmax 1≤i≤τ (p i ) (27) Among them, q represents the feature vector obtained by the fully connected layer, W represents the weight matrix of the fully connected layer, and I G represents the patient's condition assessment feature, b represents the bias vector; q i represents the i-th element of the fully connected layer output, i.e., the unnormalized score of the i-th category, τ = 3, q k represents the kth element of the fully connected layer output; p i Represents the predicted probability value of the i-th category; argmax 1≤i≤τ (p i ) indicates selecting the index of the maximum value in the probability distribution p to select the symptom stage category with the highest probability.

Images