Intelligent evaluation method for hemiplegia rehabilitation based on dynamic neighborhood firework optimization

Abstract

The invention discloses a hemiplegia rehabilitation intelligent evaluation method based on dynamic neighborhood firework optimization. The method comprises the following steps: generating a standardized multi-source hemiplegia rehabilitation data set; constructing a to-be-self-supervised training rehabilitation heterogeneous graph based on the standardized multi-source hemiplegia rehabilitation data set; using an improved decoding enhanced mask self-supervised graph learning device to carry out mask coding processing on the rehabilitation heterogeneous graph to be self-supervised and trained and the mask configuration set to obtain a masked rehabilitation heterogeneous graph representation; obtaining an initial rehabilitation score mapping model; forming a multi-target fitness function; generating a candidate spark parameter set; outputting an optimal parameter configuration set; and performing scoring reasoning on the time sequence consistent potential representation set by using the optimized rehabilitation scoring model. According to the method, comprehensive characterization which is highly sensitive to medium and long-term rehabilitation trend change and insensitive to accidental abnormity can be output, and the stability is obviously improved under real clinical multi-modal data.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of medicine, in particular to a hemiplegia rehabilitation intelligent evaluation method based on dynamic neighborhood fireworks optimization. BACKGROUND

[0002] With the rapid development of intelligent rehabilitation evaluation and wearable multi-modal sensing technology, the demand for quantitative analysis of joint function and individualized intervention of hemiplegia patients in the clinic is increasing. The existing hemiplegia joint rehabilitation evaluation method is mostly based on single sensor data, single index or statistical features, combined with linear regression and traditional machine learning methods to output scores, which is difficult to capture the dynamic evolution of multi-joint coordination, muscle coupling and complex biomechanical chain. In the face of common problems such as sensor transient packet loss, electromyographic signal noise and partial data loss, traditional methods have obvious shortcomings in score stability, cross-cycle trend discrimination and clinical interpretability.

[0003] The existing technology mainly relies on static feature extraction and simple fusion strategy in the spatio-temporal correlation fusion and robust representation of multi-source rehabilitation data, which is difficult to effectively handle high-order coupling and missing noise in heterogeneous data, resulting in scores being sensitive to occasional abnormalities and lacking accurate tracking of medium and long-term trends. At the same time, the traditional multi-index weighted evaluation method uses fixed parameter setting and lacks dynamic individualized adjustment mechanism for different patient stages and rehabilitation goals, making it difficult to balance evaluation accuracy, stability and clinical interpretability. SUMMARY

[0004] One object of the present application is to provide a hemiplegia rehabilitation intelligent evaluation method based on dynamic neighborhood fireworks optimization. The present application can output a comprehensive representation that is highly sensitive to medium and long-term rehabilitation trend changes and not sensitive to occasional abnormalities, with significant stability improvement in real clinical multi-modal data.

[0005] According to the hemiplegia rehabilitation intelligent evaluation method based on dynamic neighborhood fireworks optimization of the present application, the method comprises the following steps: Collecting a set of multi-source hemiplegia rehabilitation raw data and preprocessing to generate a set of standardized multi-source hemiplegia rehabilitation data; Constructing a self-supervised training rehabilitation heterogeneous graph based on the set of standardized multi-source hemiplegia rehabilitation data; Using an improved decoding enhanced mask self-supervised graph learner to perform mask coding processing on the self-supervised training rehabilitation heterogeneous graph and the mask configuration set to obtain a masked rehabilitation heterogeneous graph representation; Reconstructing and decoding the masked rehabilitation heterogeneous graph representation through a multi-level cross-layer attention decoding module to obtain an initial rehabilitation score mapping model; Forming a multi-objective fitness function by weighted summation of rehabilitation score prediction accuracy index, stage discrimination sensitivity index and structural explanation sparsity index; Set parameter optimization set, based on multi-objective fitness function and initial parameter individual set, use dynamic neighborhood fireworks algorithm to perform population initialization and wide area exploration, and perform explosion generation, displacement disturbance and boundary correction to generate candidate spark parameter set; Load the candidate spark parameter set into the initial rehabilitation score mapping model respectively and perform multi-objective evaluation, select the winning spark and update the population, iterate optimization until the condition is met, output the optimal parameter configuration set; Apply the optimal parameter configuration set to the initial rehabilitation score mapping model to obtain the optimized rehabilitation score model, and use the optimized rehabilitation score model to score the time sequence consistent potential representation set.

[0006] Optionally, the preprocessing operation comprises: Collect a plurality of source hemiplegia rehabilitation original data sets; Perform initial synchronization calibration on the plurality of source hemiplegia rehabilitation original data sets, establish a unified timestamp sequence, and perform resampling alignment on different sampling frequency signals to obtain a time-aligned multi-source hemiplegia rehabilitation data set; Perform noise suppression processing, artifact rejection processing, abnormal peak detection and truncation processing, and missing segment marking processing on the time-aligned multi-source hemiplegia rehabilitation data set to form a marked and purified multi-source hemiplegia rehabilitation data set; Perform interpolation reconstruction and local time slice padding processing on the marked and purified multi-source hemiplegia rehabilitation data set, and perform normalization processing and gait cycle or motion cycle normalization processing on the amplitude of the padded signal to generate a standardized multi-source hemiplegia rehabilitation data set.

[0007] Optionally, the plurality of source hemiplegia rehabilitation original data includes inertial measurement unit joint angle sequence, inertial measurement unit angular velocity sequence, surface electromyography signal sequence, foot pressure matrix sequence, gait or upper limb task time sequence event marker sequence, joint torque estimation sequence, and posture compensation marker sequence.

[0008] Optionally, the construction of the rehabilitation heterogeneous graph to be self-supervised training comprises: Map each type of data channel in the standardized multi-source hemiplegia rehabilitation data set to a joint node set, a muscle group node set, a time sequence event node set, and a mechanical virtual node set according to a preset node type mapping rule to form a rehabilitation heterogeneous graph node set; Based on the rehabilitation heterogeneous graph node set, respectively establish a structural connection edge set, a muscle group function coordination edge set, a time sequence dependence edge set, and a mechanical coupling edge set, and form a rehabilitation heterogeneous graph edge set; Construct a basic attribute vector for each node in the rehabilitation heterogeneous graph, and all node basic attribute vectors together form a node attribute set; An edge correlation weight is calculated for each edge in the rehabilitation heterogeneous graph, and all edge correlation weights collectively form an edge weight set; The node attribute set and the edge weight set are stored uniformly to form an initial attribute set of the rehabilitation heterogeneous graph; The node set of the rehabilitation heterogeneous graph, the edge set of the rehabilitation heterogeneous graph, and the initial attribute set of the rehabilitation heterogeneous graph are uniformly encapsulated to obtain a rehabilitation heterogeneous graph structure to be self-supervised training.

[0009] Optionally, the rehabilitation heterogeneous graph to be self-supervised training is set with a multi-granularity mask strategy, and a mask coding process is performed, comprising: The multi-granularity mask strategy comprises a node feature block random mask strategy, an edge existence disturbance mask strategy, and a continuous time slice clipping mask strategy; In the node feature block random mask strategy, a group of continuous feature dimensions is selected from the basic attribute vector according to the type of each node in the rehabilitation heterogeneous graph, and the selected feature dimensions are masked to generate a node feature mask matrix; In the edge existence disturbance mask strategy, each edge in the rehabilitation heterogeneous graph is deleted or replaced according to a set edge disturbance probability to generate a disturbed edge set, and an edge disturbance mask matrix is generated based on the difference between the disturbed edge set and the original edge set; In the continuous time slice clipping mask strategy, a continuous time period with a length equal to the product of the clipping ratio and the total number of time slices is randomly selected in the attribute dimension of the node feature of the rehabilitation heterogeneous graph changing over time, and the selected continuous time period is masked to generate a time dimension mask matrix; The node feature mask matrix, the edge disturbance mask matrix, and the time dimension mask matrix are combined to form a unified mask configuration set; The rehabilitation heterogeneous graph structure to be self-supervised training and the unified mask configuration set are input into the encoder module of the improved decoding enhanced mask self-supervised graph learner to finally generate a masked rehabilitation heterogeneous graph representation.

[0010] Optionally, the masked rehabilitation heterogeneous graph representation is generated, comprising: Each node in the rehabilitation heterogeneous graph structure to be self-supervised training is masked according to the node feature mask matrix to form a set of masked node features; Each edge in the rehabilitation heterogeneous graph structure to be self-supervised training is reserved or deleted according to the reservation state of the edge disturbance mask matrix to form a disturbed edge set, and the edge connection relationship of the rehabilitation heterogeneous graph structure to be self-supervised training is updated based on the disturbed edge set to obtain a disturbed rehabilitation heterogeneous graph structure; In the disturbed rehabilitation heterogeneous graph structure, the attribute dimension changing over time of the node feature set is masked in a continuous time period according to the time dimension mask matrix to form a time dimension masked node feature set; inputting the time dimension masking node feature set and the perturbed rehabilitation heterogeneous graph structure into an enhanced graph attention encoder module of the improved decoding enhanced mask self-supervised graph learner to obtain fused node features; inputting the fused node features into a multi-layer residual structure decoder module of the improved decoding enhanced mask self-supervised graph learner to generate a node reconstruction feature set and an edge structure reconstruction set; differences between the node reconstruction feature set and the original node attribute vector and structural differences between the edge structure reconstruction set and the original edge set are calculated as node reconstruction loss and edge reconstruction loss, respectively, and are weighted and fused to form a joint reconstruction loss, and the improved decoding enhanced mask self-supervised graph learner is trained through the joint reconstruction loss to obtain a masked rehabilitation heterogeneous graph representation with masking recovery capability and structure reconstruction capability.

[0011] Optionally, the improved decoding enhanced mask self-supervised graph learner comprises: a cross-modal multi-head self-attention unit, a node adaptive neighborhood aggregation unit and a time sequence perception residual update unit, the cross-modal multi-head self-attention unit is used to calculate cross-modal attention weights between joint nodes, muscle group nodes, time sequence event nodes and mechanical virtual nodes, and to perform adaptive estimation of feature importance between cross-modal nodes, the node adaptive neighborhood aggregation unit performs adaptive weighted fusion of node features and adjacent node features according to the cross-modal attention weights and the edge association weight set in the perturbed rehabilitation heterogeneous graph structure, and performs cross-modal neighborhood information aggregation of node features, and the time sequence perception residual update unit adds original time sequence position embedding information of node features on the basis of cross-modal neighborhood information aggregation of node features, and obtains fused node features through time sequence residual connection and layer normalization processing.

[0012] Optionally, the improved decoding enhanced mask self-supervised graph learner comprises: inputting the masked rehabilitation heterogeneous graph representation into a multi-level cross-layer attention decoding module, and the multi-level cross-layer attention decoding module comprises a multi-layer structure decoding unit and a cross-layer feature fusion unit; performing reverse mapping processing on the fused node features of each node of the masked rehabilitation heterogeneous graph representation in the structure decoding unit, recovering the feature dimensions of the masked nodes, outputting a node reconstruction feature set, and performing edge connection relationship prediction based on the node feature representation and the perturbed edge set, and outputting an edge structure reconstruction set; calculating reconstruction errors between the node reconstruction feature set and the original node feature set, defining a node reconstruction uncertainty set, calculating structural differences between the edge structure reconstruction set and the original edge weight set, and defining an edge structure reconstruction uncertainty set; The node reconstruction feature set and the edge structure reconstruction set are spliced and fused to form an initial latent representation set, and the latent representation associated with the corresponding node is archived to form an associated latent representation set with uncertainty annotation; The time smoothing regular processing and the inter-session alignment processing are performed on the associated latent representation set to obtain a time-consistent latent representation set. A small amount of clinical score annotation sample set is extracted, and is matched with the time-consistent latent representation set to obtain a labeled training sample. The labeled training sample is input into a supervised mapping module, and the supervised mapping module adopts a full connection neural network structure, and each latent representation vector is input into a group of full connection layers to output an initial rehabilitation score mapping model.

[0013] Optionally, the dynamic neighborhood fireworks algorithm comprises: A parameter optimization set is set, and the parameter optimization set is randomly initialized based on the initial rehabilitation score mapping model to generate an initial parameter individual set; Based on the multi-objective fitness function, the initial parameter individual set is used as the basis to initialize a population by using the dynamic neighborhood fireworks algorithm to obtain an initial spark population set, and each spark individual in the initial spark population set corresponds to a specific parameter combination of the parameter optimization set. For each spark individual in the initial spark population set, the fitness is calculated based on the initial rehabilitation score mapping model, the time-consistent latent representation set, the node reconstruction uncertainty set and the edge structure reconstruction uncertainty set to obtain a multi-objective fitness value of each spark individual. The fitness distribution characteristics are calculated according to the multi-objective fitness values of all spark individuals in the initial spark population set. Based on the node centrality feature of the rehabilitation heterogeneous graph, the node centrality feature is used to measure the importance of the node in the rehabilitation heterogeneous graph. According to the fitness distribution characteristics and the node centrality feature, the explosion radius parameter and the spark number allocation parameter corresponding to each spark individual are adaptively adjusted, the explosion radius parameter determines the neighborhood size searched by the spark individual in the parameter space, and the spark number allocation parameter determines the number of candidate spark individuals generated by each spark individual. According to the dynamically adjusted explosion radius parameter and the spark number allocation parameter, the explosion generation operation is performed on the initial spark population set, and the candidate spark individual is randomly generated within the corresponding explosion radius parameter range with the parameter optimization set of each spark individual as the center. The randomly displaced perturbation processing is performed on the explosion-generated candidate spark individual, and the boundary correction processing is performed on the candidate spark individual after the displacement perturbation processing to obtain a candidate spark parameter set meeting the constraint condition.

[0014] The beneficial effects of the present application are: (1) The application constructs a multi-layer rehabilitation heterogeneous graph containing joint nodes, muscle group nodes, time sequence event nodes and mechanical virtual nodes, realizes multi-granularity mask reconstruction self-supervision through a decoding enhanced mask self-supervised graph learner, significantly improves the accurate modeling of cross-joint, muscle group coordination and time sequence coupling characteristics in high-dimensional, heterogeneous spatiotemporal data, and can automatically complete feature repair and robust embedding representation under the condition of noise, missing and occasional sensor packet loss, output comprehensive representation which is highly sensitive to medium and long term rehabilitation trend changes and insensitive to occasional abnormalities, and the stability is obviously improved under real clinical multi-modal data.

[0015] (2) The application proposes a dynamic neighborhood fireworks multi-objective optimization strategy, integrates clinical related indicators into a unified fitness function, dynamically adjusts the search neighborhood and the number of sparks combined with the graph structure information, ensures that the parameter optimization process considers global exploration and local refinement, adjusts the parameters in real time according to the patient stage and data distribution through the self-adjustment of fitness distribution, effectively solves the problems of parameter rigidity, tuning relying on artificial and difficulty in dynamically adapting to individual differences of traditional methods, and the dynamic neighborhood fireworks optimization can find better solutions in complex multi-objective conflict scenarios, and improve the applicability and generalization ability of the model in different patients and stages. BRIEF DESCRIPTION OF DRAWINGS

[0016] The accompanying drawings are included to provide a further understanding of the application, and constitute a part of the specification, which together with the embodiments of the application, is used to explain the application, and does not constitute a limitation of the application. In the drawings: Fig. 1 A flowchart of a hemiplegia rehabilitation intelligent evaluation method based on dynamic neighborhood fireworks optimization is provided for the application; Fig. 2 A dynamic search neighborhood and spark generation mechanism diagram in the parameter optimization process of a hemiplegia rehabilitation intelligent evaluation method based on dynamic neighborhood fireworks optimization is provided for the application; DETAILED DESCRIPTION EMBODIMENT

[0017] REFERENCE Figs. 1-2 A hemiplegia rehabilitation intelligent evaluation method based on dynamic neighborhood fireworks optimization, comprising: Collecting a multi-source hemiplegia rehabilitation original data set and performing preprocessing to generate a standardized multi-source hemiplegia rehabilitation data set; In this embodiment, the preprocessing operation comprises: Collecting a multi-source hemiplegia rehabilitation original data set, the multi-source hemiplegia rehabilitation original data comprising inertial measurement unit joint angle sequences, inertial measurement unit angular velocity sequences, surface electromyogram signal sequences, foot pressure matrix sequences, gait or upper limb task time sequence event marker sequences, joint torque estimation sequences and posture compensation marker sequences; The initial synchronization calibration is performed on a multi-source hemiplegia rehabilitation original data set, a unified timestamp sequence is established, and resampling alignment is performed on different sampling frequency signals to obtain a time-aligned multi-source hemiplegia rehabilitation data set; Noise suppression processing, artifact rejection processing, abnormal peak detection and truncation processing, and missing segment marking processing are performed on the time-aligned multi-source hemiplegia rehabilitation data set to form a marked and purified multi-source hemiplegia rehabilitation data set. Interpolation reconstruction and local time slice completion processing are performed on the marked and purified multi-source hemiplegia rehabilitation data set, and normalization processing and gait cycle or action cycle normalization processing are performed on the amplitude of the completed signal to generate a standardized multi-source hemiplegia rehabilitation data set.

[0018] A self-supervised training rehabilitation heterogeneous graph is constructed based on the standardized multi-source hemiplegia rehabilitation data set. In the embodiment, the construction of the self-supervised training rehabilitation heterogeneous graph includes: Each type of data channel in the standardized multi-source hemiplegia rehabilitation data set is mapped into a joint node set, a muscle group node set, a time sequence event node set, and a mechanical virtual node set according to a preset node type mapping rule, to form a rehabilitation heterogeneous graph node set. Each joint node represents a joint of the subject, and the attributes corresponding to the joint node include a standardized joint angle sequence and a standardized joint angular velocity sequence. Each muscle group node represents a muscle group, and the attributes corresponding to the muscle group node include a standardized surface electromyogram signal feature. Each time sequence event node represents a specific action event or gait phase, and the attributes corresponding to the time sequence event node are standardized action event time stamps. Each mechanical virtual node represents mechanical properties constituted by plantar pressure distribution or power estimation, and the attributes corresponding to the mechanical virtual node are standardized plantar pressure or power parameters.

[0019] Based on the rehabilitation heterogeneous graph node set, a structural connection edge set, a muscle group function coordination edge set, a time sequence dependence edge set, and a mechanical coupling edge set are respectively established, and a rehabilitation heterogeneous graph edge set is formed. The structural connection edge connects two joint nodes having an anatomical connection relationship. The muscle group function coordination edge connects a joint node and a muscle group node. The time sequence dependence edge connects a joint node or a muscle group node and a time sequence event node. The mechanical coupling edge connects a joint node and a mechanical virtual node.

[0020] A basic attribute vector is constructed for each node in the rehabilitation heterogeneous graph, and all node basic attribute vectors collectively form a node attribute set. The basic attribute vector of the joint node includes a normalized joint angle sequence, a normalized joint angular velocity sequence, a normalized joint torque estimation sequence, and a normalized joint power estimation sequence, the basic attribute vector of the muscle group node includes an electromyogram root mean square value, an electromyogram zero-crossing rate, and an electromyogram waveform length, the basic attribute vector of the time sequence event node includes a motion event phase position and a phase interval, and the basic attribute vector of the mechanical virtual node includes a two-dimensional position of a center of pressure trajectory of a foot bottom at a node time, a total pressure value, and a pressure change amplitude in adjacent gait cycles.

[0021] An edge correlation weight is calculated for each edge in the rehabilitation heterogeneous graph, and all edge correlation weights collectively form an edge weight set; A fixed weight is assigned to a structural connection edge to represent a stable anatomical connection, a muscle group functional synergy edge weight is obtained by calculating mutual information between a joint angular velocity sequence and an electromyogram signal feature, a time sequence dependent edge weight is obtained by calculating an absolute distance between motion event phase positions and applying exponential decay, and a mechanical coupling edge weight is obtained by calculating a dynamic time warping distance between a joint angle sequence and a center of pressure trajectory of a foot bottom and performing normalization.

[0022] The node attribute set and the edge weight set are uniformly stored to form an initial attribute set of the rehabilitation heterogeneous graph; The node set of the rehabilitation heterogeneous graph, the edge set of the rehabilitation heterogeneous graph, and the initial attribute set of the rehabilitation heterogeneous graph are uniformly encapsulated to obtain a rehabilitation heterogeneous graph structure to be self-supervised training.

[0023] A multi-granularity mask strategy is set for the rehabilitation heterogeneous graph to be self-supervised training to obtain a mask configuration set, and an improved decoding enhanced mask self-supervised graph learner is used to perform mask encoding processing on the rehabilitation heterogeneous graph to be self-supervised training and the mask configuration set to obtain a masked rehabilitation heterogeneous graph representation. In this embodiment, a multi-granularity mask strategy is set for the rehabilitation heterogeneous graph to be self-supervised training, and mask encoding processing is performed, including: The multi-granularity mask strategy includes a node feature block random mask strategy, an edge existence disturbance mask strategy, and a continuous time slice clipping mask strategy. In the node feature block random mask strategy, a group of continuous feature dimensions are selected from the basic attribute vector according to the type of each node in the rehabilitation heterogeneous graph, and the selected feature dimensions are masked to generate a node feature mask matrix. The masked positions in the node feature mask matrix are represented by 0, and the unmasked positions are represented by 1. The node feature mask matrix records the masking state of each feature dimension of each node.

[0024] In the edge disturbance mask strategy, each edge in the rehabilitation heterogeneous graph is deleted or replaced according to the set edge disturbance probability, to generate a disturbed edge set, and an edge disturbance mask matrix is generated based on the difference between the disturbed edge set and the original edge set. The disturbed positions in the edge disturbance mask matrix are represented by 0, and the undisturbed positions are represented by 1. The edge disturbance mask matrix records the retention state of each edge.

[0025] In the continuous time slice clipping mask strategy, a continuous time period with a length equal to the product of the clipping ratio and the total number of time slices is randomly selected in the attribute dimension of the node feature of the rehabilitation heterogeneous graph changing over time, and the selected continuous time period is masked to generate a time dimension mask matrix. The masked positions in the time dimension mask matrix are represented by 0, and the unmasked positions are represented by 1. The time dimension mask matrix records the masking state of each time slice.

[0026] The node feature mask matrix, the edge disturbance mask matrix and the time dimension mask matrix are combined to form a unified mask configuration set. The rehabilitation heterogeneous graph structure to be self-supervised trained and the unified mask configuration set are input into the encoder module of the improved decoding enhanced mask self-supervised graph learner, and finally a masked rehabilitation heterogeneous graph representation is generated.

[0027] In the embodiment, generating the masked rehabilitation heterogeneous graph representation comprises: Each node in the rehabilitation heterogeneous graph structure to be self-supervised trained is masked according to the node feature mask matrix to form a set of masked node features. The values of the node features after the masking processing are set to zero at the masked positions, and the original values are kept at the unmasked positions.

[0028] Each edge in the rehabilitation heterogeneous graph structure to be self-supervised trained is subjected to retention or deletion processing according to the retention state of the edge disturbance mask matrix, to form a disturbed edge set, and the edge connection relationship of the rehabilitation heterogeneous graph structure to be self-supervised trained is updated based on the disturbed edge set, to obtain a disturbed rehabilitation heterogeneous graph structure. In the disturbed rehabilitation heterogeneous graph structure, the attribute dimension of the node feature set changing over time is subjected to continuous time period masking processing according to the time dimension mask matrix, to form a time dimension masked node feature set. The node features after the continuous time period masking processing are set to zero in the corresponding continuous time period, and the original values are kept in the unmasked time period.

[0029] The time dimension masked node feature set and the disturbed rehabilitation heterogeneous graph structure are input into the enhanced graph attention encoder module of the improved decoding enhanced mask self-supervised graph learner, to obtain fused node features. The fused node features are input into a multi-layer residual structure decoder module of the improved decoding enhanced mask self-supervised graph learner to generate a node reconstruction feature set and an edge structure reconstruction set. In embodiment 1, the fused node features are input into a multi-layer residual structure decoder module of the improved decoding enhanced mask self-supervised graph learner, and the multi-layer residual structure decoder module includes multiple decoding layers. Each decoding layer sequentially performs feature reverse mapping and structure reconstruction processing on the fused node feature representation. In the feature reverse mapping process, the fused node feature representation is mapped back to each attribute dimension of the node original feature space, and the masked node features are restored and reconstructed. In the structure reconstruction processing, the original connection relationship of the edges is inferred based on the fused node feature representation and the disturbed edge set, and the disturbed edge connection state is restored and reconstructed. The intermediate reconstruction results output by each decoding layer are fused through a residual connection unit layer by layer, and a node reconstruction feature set and an edge structure reconstruction set are output. The node reconstruction feature set is used to restore the masked feature dimensions in the node basic attribute vector, and the edge structure reconstruction set is used to restore the disturbed edge connection state in the disturbed edge set.

[0030] The differences between the node reconstruction feature set and the node original attribute vector, and the structural differences between the edge structure reconstruction set and the original edge set are calculated as node reconstruction loss and edge reconstruction loss, respectively, and are weighted and fused to form a joint reconstruction loss. The joint reconstruction loss is used for training to obtain a masked rehabilitation heterogeneous graph representation with masking recovery ability and structure reconstruction ability.

[0031] The improved decoding enhanced mask self-supervised graph learner includes: The cross-modal multi-head self-attention unit is used to calculate the cross-modal attention weights between the joint nodes, muscle group nodes, time event nodes, and mechanical virtual nodes, and to perform feature importance adaptive estimation between cross-modal nodes. The node adaptive neighborhood aggregation unit performs adaptive weighted fusion of node features and adjacent node features based on the cross-modal attention weights and the edge association weight set in the disturbed rehabilitation heterogeneous graph structure, and performs cross-modal neighborhood information aggregation of node features. The time-aware residual update unit adds original time sequence position embedding information to the node features based on the cross-modal neighborhood information aggregation of the node features, and obtains fused node features through time sequence residual connection and layer normalization processing.

[0032] In the embodiment 1, the cross-modal multi-head self-attention unit calculates the cross-modal attention weight between each node and the node of different modal type, the cross-modal attention weight measures the importance contribution degree of the feature between different modal nodes, the node adaptive neighborhood aggregation unit adaptively fuses the node feature and the feature of the adjacent node according to the cross-modal attention weight and the edge association weight set, obtains the cross-modal structural feature representation in the local neighborhood of the node, and the time-aware residual update unit fuses the original time sequence position embedding information of the node feature based on the cross-modal structural feature representation in the local neighborhood of the node, and obtains the fused node feature representation considering the local structure and global time sequence information through the time sequence residual connection and layer normalization processing.

[0033] The multi-level cross-layer attention decoding module is used for reconstructing and decoding the masked rehabilitation heterogeneous graph representation, and outputs a node reconstruction feature set, an edge structure reconstruction set and a reconstruction uncertainty set. The initial latent representation set is generated based on the node reconstruction feature set and the edge structure reconstruction set, and the reconstruction uncertainty set and the initial latent representation set are associated and archived to obtain an associated latent representation set. The time smoothing regularization processing and the inter-session alignment processing are performed on the associated latent representation set to obtain a time-consistent latent representation set. A small amount of clinical score labeled samples are extracted and mapped and calibrated with the time-consistent latent representation set to train the supervised mapping module to obtain an initial rehabilitation score mapping model. In the embodiment, generating the initial latent representation set and performing calibration modeling includes: The masked rehabilitation heterogeneous graph representation is input into the multi-level cross-layer attention decoding module, and the multi-level cross-layer attention decoding module includes a multi-layer structure decoding unit and a cross-layer feature fusion unit. The structure decoding unit restores the node feature and predicts the edge connection of the masked rehabilitation heterogeneous graph representation, and the cross-layer feature fusion unit is used for weighting and fusing the results output by each decoding layer to obtain a layer-by-layer optimized reconstruction output.

[0034] In the structure decoding unit, the fused node feature of each node of the masked rehabilitation heterogeneous graph representation is reversely mapped to restore the feature dimension of the node, and the node reconstruction feature set is output. The edge connection relationship prediction is performed based on the node feature representation and the perturbed edge set, and the edge structure reconstruction set is output. Each node reconstruction feature vector in the node reconstruction feature set includes the reconstruction result of all basic attribute dimensions of the node, and each reconstruction edge weight in the edge structure reconstruction set represents the structure relationship recovery result of the corresponding edge.

[0035] The reconstruction error between the node reconstruction feature set and the original node feature set is calculated to define the node reconstruction uncertainty set, and the structure difference between the edge structure reconstruction set and the original edge weight set is calculated to define the edge structure reconstruction uncertainty set. The node reconstruction uncertainty measures the credibility of the reconstruction of each node feature. The node reconstruction uncertainty is obtained by calculating the sum of squares of node feature errors and normalizing them. The edge structure reconstruction uncertainty measures the credibility of the reconstruction of each edge structure relationship. The edge structure reconstruction uncertainty is obtained by calculating the edge weight difference and applying exponential smoothing.

[0036] The node reconstruction feature set and the edge structure reconstruction set are spliced ​​and fused to form the initial potential representation set, and the potential representation associated with the corresponding node is archived to form an associated potential representation set with uncertainty annotation; Each latent representation vector in the initial latent representation set includes both the semantic feature information of the node and the contextual feature information of the node in the rehabilitation heterogeneous graph structure. Each element in the associated latent representation set includes the latent representation vector, the node reconstruction uncertainty, and the edge structure reconstruction uncertainty of the node-related edges.

[0037] Perform temporal smoothing regularization and inter-session alignment on the associated latent representation set to obtain a temporally consistent latent representation set; In Example 1, the temporal smoothing regularization process is to smoothly update the potential representations of adjacent time points of each node in chronological order. The temporal smoothing regularization goal is to constrain the degree of change of the potential representation of each node within a continuous time period. The temporal smoothing regularization strength is measured by the sum of the squares of the differences between the potential representations of all consecutive time points. The weight of the smoothing target is controlled by the smoothing coefficient parameter.

[0038] In Example 1, the inter-session alignment process uses a global mean-centered weighted strategy to uniformly align the latent representations in different rehabilitation sessions. The cross-session alignment goal is to minimize the weighted difference between the latent representation of each session and the global average latent representation of all sessions to obtain a temporally consistent set of latent representations.

[0039] Extract a small set of clinical score annotated samples and perform one-to-one matching with the temporally consistent latent representation set to obtain labeled training samples; Each sample in the clinical score annotated sample set includes a node identifier and a clinical score label corresponding to the node.

[0040] The labeled training samples are input into the supervised mapping module, which uses a fully connected neural network structure to input each potential representation vector into a set of fully connected layers and output an initial rehabilitation score mapping model; In Example 1, a nonlinear activation function is used to realize the mapping from the input latent representation to the score label space. During the training process, the mean square error between the score label and the model prediction value is used as the loss function to establish a supervised mapping relationship between the latent representation space and the score label space, and output the initial rehabilitation score mapping model.

[0041] The multi-objective fitness function is formed by a weighted sum of a rehabilitation score prediction accuracy index, a stage discrimination sensitivity index and a structural explanation sparsity index; The rehabilitation score prediction accuracy index in embodiment 1 is the error between the rehabilitation score output by the initial rehabilitation score mapping model and the clinical score labeled sample set, which is obtained by calculating the average of the mean square error between the score prediction value and the actual label value of all labeled nodes; the stage discrimination sensitivity index is the ability of the initial rehabilitation score mapping model to distinguish the stage category at different rehabilitation stages, which is obtained by calculating the score mean value difference between different stage label samples and taking the ratio of the absolute value of the score mean value difference to the sample variance after normalization as the sensitivity evaluation result; the structural explanation sparsity index is the sparsity degree of the initial rehabilitation score mapping model in the process of score output to the structural explanation of the rehabilitation heterogeneous graph, which is obtained by calculating the sparsity of the score attribution weight distribution, and the sparsity is measured by L1 norm or Gini coefficient.

[0042] An initial parameter individual set is generated based on the initial rehabilitation score mapping model, and population initialization and global exploration are performed based on the multi-objective fitness function and the initial parameter individual set using the dynamic neighborhood fireworks algorithm to obtain an initial fireworks population set; the fitness distribution characteristics of each fireworks individual in the initial fireworks population set and the node centrality characteristics of the rehabilitation heterogeneous graph are calculated, and the explosion radius parameter and the fireworks number allocation parameter are adaptively set based on the above two characteristics to form a dynamic neighborhood configuration; the initial fireworks population set is exploded, displaced and boundary corrected according to the dynamic neighborhood configuration to generate a candidate fireworks parameter set; In this embodiment, the dynamic neighborhood fireworks algorithm comprises: An initial parameter individual set is generated based on the initial rehabilitation score mapping model, and population initialization and global exploration are performed based on the multi-objective fitness function and the initial parameter individual set using the dynamic neighborhood fireworks algorithm to obtain an initial fireworks population set; the fitness distribution characteristics of each fireworks individual in the initial fireworks population set and the node centrality characteristics of the rehabilitation heterogeneous graph are calculated, and the explosion radius parameter and the fireworks number allocation parameter are adaptively set based on the above two characteristics to form a dynamic neighborhood configuration; the initial fireworks population set is exploded, displaced and boundary corrected according to the dynamic neighborhood configuration to generate a candidate fireworks parameter set; The parameter optimization set in embodiment 1 includes node reconstruction feature weight coefficients, edge structure reconstruction weight coefficients, latent representation smoothing coefficients, cross-session alignment weight coefficients, score mapping loss weight coefficients, and the number of hidden layer neurons in the fully connected neural network structure in the supervised mapping module, and other parameters that need to be optimized can also be selected; An initial parameter individual set is generated based on the initial rehabilitation score mapping model, and population initialization and global exploration are performed based on the multi-objective fitness function and the initial parameter individual set using the dynamic neighborhood fireworks algorithm to obtain an initial fireworks population set; the fitness distribution characteristics of each fireworks individual in the initial fireworks population set and the node centrality characteristics of the rehabilitation heterogeneous graph are calculated, and the explosion radius parameter and the fireworks number allocation parameter are adaptively set based on the above two characteristics to form a dynamic neighborhood configuration; the initial fireworks population set is exploded, displaced and boundary corrected according to the dynamic neighborhood configuration to generate a candidate fireworks parameter set; For each fireworks individual in the initial fireworks population set, the fitness is calculated based on the initial rehabilitation score mapping model, the time-consistent latent representation set, the node reconstruction uncertainty set and the edge structure reconstruction uncertainty set to obtain the multi-objective fitness value of each fireworks individual; The fitness distribution feature is calculated according to multi-objective fitness values of all spark individuals in the initial spark population set. The fitness distribution feature includes a fitness mean value, a fitness variance value and a fitness extreme difference value, and is used to measure a concentration or dispersion degree of the current population in the parameter space.

[0043] The node centrality feature is based on a node centrality feature of the rehabilitation heterogeneous graph, and is used to measure an importance degree of the node in the rehabilitation heterogeneous graph. The node centrality feature is obtained by calculating a weighted sum of connection edge weights of each node in the rehabilitation heterogeneous graph structure, and a node with a higher node centrality feature value has a greater influence on parameter optimization.

[0044] The explosion radius parameter and the spark number allocation parameter corresponding to each spark individual are adaptively adjusted according to the fitness distribution feature and the node centrality feature, the explosion radius parameter determines a neighborhood size searched by the spark individual in the parameter space, and the spark number allocation parameter determines a number of candidate spark individuals generated by each spark individual. In embodiment 1, when the fitness distribution feature shows that the fitness values of the spark individuals are concentrated, the explosion radius parameter is dynamically reduced and the spark number allocation parameter is increased to strengthen local fine search of the parameters; when the fitness values of the spark individuals are dispersed, the explosion radius parameter is dynamically increased and the spark number allocation parameter is reduced to strengthen global exploration of the parameter space, and the concentration of the fitness values of the spark individuals and the dispersion of the fitness values of the spark individuals are defined artificially.

[0045] The higher the node centrality feature value is, the greater the spark number allocation parameter corresponding to the spark individual is, so as to ensure fine adjustment of parameters related to key nodes in the optimization process.

[0046] The initial spark population set is subjected to an explosion generation operation according to the dynamically adjusted explosion radius parameter and the spark number allocation parameter, and candidate spark individuals are randomly generated within a range of the explosion radius parameter and centered on a parameter optimization set of each spark individual; The candidate spark individuals generated by the explosion are subjected to random displacement disturbance processing, and the candidate spark individuals subjected to the displacement disturbance processing are subjected to boundary correction processing to obtain a candidate spark parameter set satisfying a constraint condition. The displacement disturbance processing is to apply random disturbance to parameter values in the candidate spark individual to improve diversity of the candidate spark individual, and a disturbance amplitude of the random disturbance is controlled by a unified disturbance amplitude parameter, and the boundary correction processing is to ensure that the parameter values in each candidate spark individual are located in a search range of the set parameter optimization set, and a parameter value exceeding the search range is corrected to a closest boundary value of the corresponding search range.

[0047] The candidate spark parameter set is loaded into the initial rehabilitation score mapping model respectively, and multi-objective evaluation is performed in combination with the time consistent latent representation set and the reconstruction uncertainty set, to obtain a candidate spark fitness result set. Based on the candidate spark fitness result set, the best spark is screened, the inferior spark is eliminated, and the memory archive is updated, to generate an iteratively updated spark population set and record a parameter evolution trajectory set. It is judged whether the iteratively updated spark population set meets the termination condition. If not, the explosion generation and evaluation step is returned to be executed until the termination condition is met and the optimal parameter configuration set is output. In Example 1, each candidate spark individual in the candidate spark parameter set is loaded into the initial rehabilitation score mapping model, the corresponding model structure and parameter setting are updated, and a candidate parameter mapping model is constructed. The candidate parameter mapping model is inputted in combination with the time consistent latent representation set, the node reconstruction uncertainty set, and the edge structure reconstruction uncertainty set. Multi-objective fitness function evaluation is performed on each candidate spark individual in the candidate spark parameter set, and a candidate spark fitness result set is obtained. The candidate spark fitness result set is sorted and screened, and the best spark screening operation is performed according to the multi-objective Pareto optimal criterion. The candidate spark individuals with non-dominated frontiers in fitness performance are selected into the next generation population to form a best spark set. Based on the candidate spark fitness result set, the inferior spark elimination operation is performed, and the candidate spark individuals with fitness values significantly worse than the multi-objective threshold are removed from the population space to form an optimized iteratively updated spark population set. The best spark individuals selected in each round and their corresponding parameter configurations are recorded into a memory archive set, and the best spark individuals in all rounds are formed into an optimal solution trajectory set to form a parameter evolution trajectory set. It is judged whether the iteratively updated spark population set meets the termination condition. The termination conditions include reaching the maximum number of iterations or the optimal spark individual fitness continuously improving for multiple rounds without obvious improvement. If not, the current iteratively updated spark population set is returned to perform the explosion generation, disturbance, and boundary correction steps for a new round of iteration. If the termination condition is met, the spark individual with the optimal fitness in the current parameter evolution trajectory set is output as the final optimal parameter configuration set.

[0048] The optimal parameter configuration set is applied to the initial rehabilitation score mapping model to obtain an optimized rehabilitation score model. The optimized rehabilitation score model is used to score and infer the time consistent latent representation set, and outputs the comprehensive rehabilitation score, the sub-dimension rehabilitation score set, the rehabilitation trend time series, and the node and edge explanation attribution set.

[0049] The optimal parameter configuration set is loaded to the initial rehabilitation score mapping model in embodiment 1, each parameter in the initial rehabilitation score mapping model is synchronously replaced and valid, and an optimized rehabilitation score model is obtained; the time sequence consistent latent representation set is input into the optimized rehabilitation score model, and score inference is performed in time sequence to obtain a predicted score output sequence corresponding to each node and time point; According to the evaluation index weight coefficient in the optimal parameter configuration set, the score output sequence is weighted and fused to form a comprehensive rehabilitation score, the sub-score components corresponding to each evaluation index in the score output sequence are extracted respectively to form a sub-dimension rehabilitation score set, and the comprehensive rehabilitation score is directly arranged according to the time index to obtain a rehabilitation trend time sequence; Based on the attention weight, parameter weight distribution, node reconstruction uncertainty and edge structure reconstruction uncertainty generated by the optimized rehabilitation score model in the inference process, the key node contribution and the key edge relationship contribution are extracted, and a node and edge explanation attribution set is generated: Comprehensive rehabilitation score: refers to a single score result obtained by weighting and fusing all evaluation indexes according to evaluation index weight coefficients; Sub-dimension rehabilitation score set: refers to a score result set obtained by separately calculating each evaluation index; Rehabilitation trend time sequence: refers to a sequence of comprehensive rehabilitation scores arranged in time sequence, used to represent the time variation trend of rehabilitation process; Node and edge explanation attribution set: refers to an explanation information set composed of key node contribution sorting results and key edge relationship contribution sorting results, used to show the dependence degree of score results on nodes and edges; The comprehensive rehabilitation score, the sub-dimension rehabilitation score set, the rehabilitation trend time sequence and the node and edge explanation attribution set are output, and the evaluation result generation of the optimized rehabilitation score model is completed. Embodiment

[0050] In a lower limb rehabilitation evaluation project of a hemiplegic patient, the implementation team wore an inertial measurement unit, a surface electromyography sensor and a plantar pressure distribution pad for a certain patient during a standard walking test, and collected continuous 15-minute lower limb movement data. During the test, the system recorded that the angle data of the right ankle IMU node lost at the 8th minute of walking, lasting for 2.3 seconds, and then the electromyography signal also drifted obviously due to electrode slip, and the plantar pressure data had breakpoints missing due to hardware jitter from the 12th minute to the 12th minute and 15 seconds.

[0051] After the data is uploaded to the analysis system, it is automatically identified that there are a total of 7 signals missing or affected by noise in the gait sequence. Specifically, there is a single continuous 14-frame 0-value interval in the ankle joint angle sequence, abnormal peak values (maximum value up to 2.5 times the baseline) in the root mean square value of the electromyographic signal appear in several gait cycles, and the foot pressure center trajectory loses frames, resulting in the interruption of the pressure distribution curve. The system completes time synchronization and preliminary normalization of all signal streams in the data preprocessing stage, and marks all abnormal segments in the feature matrix.

[0052] Under the traditional method, the data analysis system first uses linear interpolation to complete the 0-value interval, replaces the abnormal peak value of the electromyographic signal with the local mean, and fills the median value for the foot pressure interruption segment. Subsequently, the system extracts features such as step symmetry, maximum activity of the ankle joint, and activation rate of the lower limb muscle group, outputs a rehabilitation score based on a static weighted regression model, and the patient's score is 68 points in the 5th to 7th minute, but drops to 63 points in the 8th minute due to interpolation error, and drops to 60 points in the 12th minute due to distortion of pressure data completion.

[0053] When using the method of the present application, a joint-muscle-event-mechanics heterogeneous graph is constructed using a multi-source data tensor. The graph model automatically masks the feature interval corresponding to the right ankle IMU node at the 8th minute, and randomly masks part of the time slices of the electromyographic node and the foot pressure node. The missing data is recovered using a self-supervised reconstruction mechanism. In this process, the reconstruction uncertainty down-regulates the weight of the masked area. After the completion of the mask self-supervision, the system sends the node and edge latent representation of the patient to a lightweight fully connected neural network, combines the previous stage of artificial scoring samples (the patient's 4th, 7th, and 10th minutes have doctor evaluation labels, which are 67, 70, and 73 respectively), trains the rehabilitation score mapping model, enters the multi-objective optimization stage, and automatically adjusts the sub-score weight and sliding window parameters using the dynamic neighborhood fireworks algorithm. Due to the large impact of the loss of the pressure node on the patient in the 12th minute, the algorithm adaptively relaxes the related parameters, and the optimized model outputs 66 points in the 8th and 12th minutes, without mutation. The rehabilitation trend score curve shows a smooth upward trend over time.

[0054] Data comparison shows that the mean square error of the output of the mask self-supervised and dynamic neighborhood optimization model in all abnormal intervals is 2.4 points, while the mean square error of the traditional method in the same interval is 5.9 points. During the 4th to 13th minutes of this patient, the average deviation of the comprehensive rehabilitation score at each time point from the doctor's artificial annotation is only 1.8 points. The trend curves of sub-dimension scores such as ankle joint symmetry and muscle decoupling degree are highly consistent with manual annotation. The model's automatic attribution results show that the ankle joint node and the gastrocnemius muscle group node have the top two attention weights in the overall score, which has a high reference value for the doctor's secondary review feedback.

[0055] In another application, the analysis system continuously evaluated the morning exercise gait data of 5 hemiplegic patients, and the system found that one of the patients had a temporary and sharp decrease in left knee joint range of motion at the 3rd training, and the right calf muscle electromyogram increased synchronously. The system detected the change in real time and automatically marked the abnormality in the graph structure. The dynamic neighborhood fireworks algorithm adaptively enhanced the left knee node parameter search through node centrality, and finally output the joint motion symmetry index in the sub-dimension score from 75 to 69 in this period. After recovery, it rose again. Compared with the traditional evaluation period, the average processing was not captured, and the score fluctuation was not captured. The subsequent clinical follow-up confirmed that the patient did indeed have subjective discomfort and abnormal force during the period.

[0056] In the whole data set statistics, the overall trend of the score sequence of all patients using the method of the application is consistent with the doctor's grading curve, the abnormal section score fluctuation is reduced by 38%, the sub-dimension score and the node / edge attribution report are explained by the doctor's blind evaluation, and the average score is increased from 2.7 to 4.3 (full score 5) by the traditional method; the average absolute error between the score output of the model in nearly 12000 training samples and the doctor's score is less than 2.5 points, which is significantly better than the traditional linear model.

[0057] The above describes only the preferred specific embodiments of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art can make equivalent replacements or changes to the technical solutions and inventive concepts of the present application within the technical range disclosed by the present application, which should be covered within the protection scope of the present application.

Claims

1. An intelligent assessment method for hemiplegia rehabilitation based on dynamic neighborhood fireworks optimization, characterized by: include: Collecting multi-source hemiplegia rehabilitation raw data sets, and pre-processing them to generate standardized multi-source hemiplegia rehabilitation data sets; Construct a heterogeneous rehabilitation graph for self-supervised training based on a standardized multi-source hemiplegia rehabilitation data set; An improved decoding-enhanced masked self-supervised graph learner is used to perform mask encoding on the self-supervised training rehabilitation heterogeneous graph and the mask configuration set to obtain the masked rehabilitation heterogeneous graph representation; The masked rehabilitation heterogeneous graph representation is reconstructed and decoded through a multi-level cross-layer attention decoding module to obtain an initial rehabilitation score mapping model. The multi-objective fitness function is formed by weighted summation of the rehabilitation score prediction accuracy index, the stage discrimination sensitivity index and the structural interpretation sparsity index. Set the parameter optimization set, use the dynamic neighborhood fireworks algorithm to perform population initialization and wide-area exploration based on the multi-objective fitness function and the initial parameter individual set, and perform explosion generation, displacement perturbation and boundary correction to generate a candidate spark parameter set; The candidate spark parameter sets are loaded into the initial rehabilitation score mapping model and multi-objective evaluation is performed to screen the winning sparks and update the population. The optimization is iterated until the conditions are met and the optimal parameter configuration set is output; The optimal parameter configuration set is applied to the initial rehabilitation score mapping model to obtain an optimized rehabilitation score model, which is then used to perform score inference on the temporally consistent latent representation set.

2. The intelligent assessment method for hemiplegia rehabilitation based on dynamic neighborhood fireworks optimization according to claim 1, characterized in that: The pre-processing operation includes: Collect multi-source hemiplegia rehabilitation raw data sets; Perform initial synchronization calibration on the multi-source hemiplegia rehabilitation raw data set, establish a unified timestamp sequence, and perform resampling alignment on signals with different sampling frequencies to obtain a time-aligned multi-source hemiplegia rehabilitation data set; Perform noise suppression, artifact removal, abnormal peak detection and truncation, and missing segment marking on the time-aligned multi-source hemiplegia rehabilitation data set to form a labeled and purified multi-source hemiplegia rehabilitation data set; The labeled and purified multi-source hemiplegia rehabilitation data set is interpolated and reconstructed and the local time slice is padded. The padded signal amplitude is normalized and the gait cycle or movement cycle is normalized to generate a standardized multi-source hemiplegia rehabilitation data set.

3. The intelligent assessment method for hemiplegia rehabilitation based on dynamic neighborhood fireworks optimization according to claim 2, characterized in that: The multi-source hemiplegia rehabilitation raw data includes an inertial measurement unit joint angle sequence, an inertial measurement unit angular velocity sequence, a surface electromyography signal sequence, a plantar pressure matrix sequence, a gait or upper limb task timing event marker sequence, a joint torque estimation sequence, and a posture compensation marker sequence.

4. The intelligent assessment method for hemiplegia rehabilitation based on dynamic neighborhood fireworks optimization according to claim 2, characterized in that: The construction of the rehabilitation heterogeneous graph to be self-supervised training includes: According to the preset node type mapping rules, various data channels in the standardized multi-source hemiplegia rehabilitation data set are mapped into joint node sets, muscle group node sets, time series event node sets and mechanical virtual node sets to form a rehabilitation heterogeneous graph node set; Based on the rehabilitation heterogeneous graph node set, the structural connection edge set, muscle group function synergy edge set, temporal dependency edge set and mechanical coupling edge set are established respectively, and form the rehabilitation heterogeneous graph edge set; Construct a basic attribute vector for each node in the rehabilitation heterogeneous graph, and all node basic attribute vectors together form a node attribute set; The edge association weight is calculated for each edge in the rehabilitation heterogeneous graph, and all edge association weights together form the edge weight set; The node attribute set and the edge weight set are stored uniformly to form the initial attribute set of the rehabilitation heterogeneous graph; The rehabilitation heterogeneous graph node set, the rehabilitation heterogeneous graph edge set and the rehabilitation heterogeneous graph initial attribute set are uniformly encapsulated to obtain the rehabilitation heterogeneous graph structure to be trained through self-supervision.

5. The intelligent assessment method for hemiplegia rehabilitation based on dynamic neighborhood fireworks optimization according to claim 4, characterized in that: The method sets a multi-granularity mask strategy for the self-supervised training rehabilitation heterogeneous graph and performs mask encoding processing, including: Set up multi-granularity masking strategies, including node feature block random masking strategy, edge existence perturbation masking strategy and continuous time slice clipping masking strategy; In the node feature block random masking strategy, a set of continuous feature dimensions is selected from the basic attribute vector according to the type of each node in the rehabilitation heterogeneous graph, and the selected feature dimensions are masked to generate a node feature mask matrix; In the edge perturbation mask strategy, each edge in the rehabilitation heterogeneous graph is deleted or replaced according to the set edge perturbation probability to generate a perturbed edge set. The edge perturbation mask matrix is ​​generated based on the difference between the perturbated edge set and the original edge set. In the continuous time slice cropping mask strategy, a continuous time segment with a length equal to the product of the cropping ratio and the total number of time slices is randomly selected from the attribute dimension of the node features of the rehabilitation heterogeneous graph that changes with time. The selected continuous time segment is masked to generate a time dimension mask matrix. The node feature mask matrix, the edge perturbation mask matrix and the time dimension mask matrix are combined to form a unified mask configuration set; The heterogeneous graph structure to be self-supervised trained and the unified mask configuration set are input into the encoder module of the improved decoding enhanced masked self-supervised graph learner, and finally the masked rehabilitation heterogeneous graph representation is generated.

6. The intelligent assessment method for hemiplegia rehabilitation based on dynamic neighborhood fireworks optimization according to claim 5, characterized in that: The method of generating a masked rehabilitation heterogeneous graph representation comprises: Each node in the self-supervised training rehabilitation heterogeneous graph structure is masked according to the node feature mask matrix to form a masked node feature set; For each edge in the self-supervised training rehabilitation heterogeneous graph structure, the original edge set is retained or deleted according to the retention state of the edge perturbation mask matrix to form a perturbed edge set, and the edge connection relationship of the self-supervised training rehabilitation heterogeneous graph structure is updated based on the perturbed edge set to obtain the perturbed rehabilitation heterogeneous graph structure; In the perturbation rehabilitation heterogeneous graph structure, the attribute dimension of the node feature set that changes with time is masked for a continuous time period according to the time dimension mask matrix to form a time dimension masked node feature set; The time dimension masked node feature set and the perturbation recovery heterogeneous graph structure are input into the enhanced graph attention encoder module of the improved decoding enhanced masked self-supervised graph learner to obtain the fused node features; The fused node features are input into the multi-layer residual structure decoder module of the improved decoding enhanced mask self-supervised graph learner to generate a node reconstruction feature set and an edge structure reconstruction set; The difference between the node reconstructed feature set and the node original attribute vector, and the structural difference between the edge structure reconstructed set and the original edge set are calculated as the node reconstruction loss and edge reconstruction loss, respectively, and weighted fusion is formed into a joint reconstruction loss. Through training through the joint reconstruction loss, a masked recovery heterogeneous graph representation with masking recovery and structure reconstruction capabilities is obtained.

7. The intelligent assessment method for hemiplegia rehabilitation based on dynamic neighborhood fireworks optimization according to claim 5, characterized in that: The improved decoding-enhanced mask self-supervised graph learner comprises: A cross-modal multi-head self-attention unit, a node adaptive neighborhood aggregation unit and a timing-aware residual update unit. The cross-modal multi-head self-attention unit is used to calculate the cross-modal attention weights between joint nodes, muscle group nodes, timing event nodes, and mechanical virtual nodes, and perform adaptive estimation of the feature importance between cross-modal nodes. The node adaptive neighborhood aggregation unit performs adaptive weighted fusion of the node itself and adjacent node features based on the cross-modal attention weights and the edge association weight set in the perturbation rehabilitation heterogeneous graph structure, and performs cross-modal neighborhood information aggregation of node features. The timing-aware residual update unit adds the original temporal position embedding information of the node features on the basis of the cross-modal neighborhood information aggregation of the node features, and obtains the fused node features through temporal residual connection and layer normalization.

8. The intelligent assessment method for hemiplegia rehabilitation based on dynamic neighborhood fireworks optimization according to claim 7, characterized in that: Generating an initial potential representation set and performing calibration modeling includes: The masked rehabilitation heterogeneous graph representation is input into a multi-level cross-layer attention decoding module, which includes a multi-layer structure decoding unit and a cross-layer feature fusion unit; In the structure decoding unit, the fused node features of each node in the masked rehabilitation heterogeneous graph representation are reverse mapped to restore the masked feature dimensions of the node, output the node reconstruction feature set, and perform edge connection relationship prediction based on the node feature representation and the perturbed edge set, and output the edge structure reconstruction set; The reconstruction error between the node reconstruction feature set and the original node feature set is calculated to define the node reconstruction uncertainty set. The structural difference between the edge structure reconstruction set and the original edge weight set is calculated to define the edge structure reconstruction uncertainty set. The node reconstruction feature set and the edge structure reconstruction set are spliced ​​and fused to form the initial potential representation set, and the potential representation associated with the corresponding node is archived to form an associated potential representation set with uncertainty annotation; Perform temporal smoothing regularization and inter-session alignment on the associated latent representation set to obtain a temporally consistent latent representation set; Extract a small set of clinical score annotated samples and perform one-to-one matching with the temporally consistent latent representation set to obtain labeled training samples; The labeled training samples are input into the supervised mapping module, which adopts a fully connected neural network structure, inputs each potential representation vector into a set of fully connected layers, and outputs the initial rehabilitation score mapping model.

9. The intelligent assessment method for hemiplegia rehabilitation based on dynamic neighborhood fireworks optimization according to claim 8, wherein the dynamic neighborhood fireworks algorithm comprises: Setting a parameter optimization set, and randomly initializing the parameter optimization set based on the initial rehabilitation score mapping model to generate an initial parameter individual set; Based on the multi-objective fitness function, the dynamic neighborhood fireworks algorithm is used to initialize the population based on the initial parameter individual set to obtain the initial spark population set. Each spark individual in the initial spark population set corresponds to a specific parameter combination of the parameter optimization set. For each spark individual in the initial spark population set, the fitness is calculated based on the initial rehabilitation score mapping model, the temporal consistent potential representation set, the node reconstruction uncertainty set, and the edge structure reconstruction uncertainty set to obtain the multi-objective fitness value of each spark individual; Calculate the fitness distribution characteristics based on the multi-objective fitness values ​​of all spark individuals in the initial spark population set; Based on the node centrality feature of the rehabilitation heterogeneous graph, the node centrality feature is used to measure the importance of the node in the rehabilitation heterogeneous graph; Adaptively adjust the explosion radius parameter and spark quantity allocation parameter corresponding to each spark individual according to the fitness distribution characteristics and node centrality characteristics. The explosion radius parameter determines the size of the neighborhood that the spark individual searches in the parameter space, and the spark quantity allocation parameter determines the number of candidate spark individuals generated by each spark individual. The initial spark population set is subjected to explosion generation operation according to the dynamically adjusted explosion radius parameter and spark quantity distribution parameter, and candidate spark individuals are randomly generated within the corresponding explosion radius parameter range with the parameter optimization set of each spark individual as the center; The candidate spark individuals generated by the explosion are subjected to random displacement perturbation processing, and the boundary correction processing is performed on the candidate spark individuals after the displacement perturbation processing to obtain a set of candidate spark parameters that meet the constraint conditions.

Images