Spatiotemporal feature quantification assessment method for parkinsonian motor symptoms

Abstract

The present application relates to the technical field of medical data analysis, in particular to a spatiotemporal feature quantitative evaluation method for Parkinson's disease motor symptoms, comprising: deploying an inertial measurement unit to collect three-dimensional acceleration angular velocity magnetic field data, constructing a human body connection structure to generate a connection neural network topology, calculating a trajectory direction angle and angular velocity change adaptive weighting to judge stability, aggregating adjacent features to form a multi-round convolution propagation to form a spatiotemporal fusion motion feature set, identifying tremor gait amplitude according to a time sequence multi-head attention to extract a time sequence feature mode, and calculating tremor gait coordination to generate a Parkinson's disease motor symptom quantitative evaluation result, wherein, in the present application, a sensor network motion data topology connection is constructed, multi-dimensional part information is fused to capture limb coordination, adaptive weight decay and connection convolution deep aggregation features are used to maintain time sequence stability and capture long-range associations, and multi-head attention is combined to distinguish tremor frequency and gait changes, thereby improving the consistency of Parkinson's symptom recognition and evaluation.

Description

Technical Field

[0001] This invention relates to the field of medical data analysis technology, and in particular to a method for quantitative assessment of the spatiotemporal characteristics of motor symptoms in Parkinson's disease. Background Technology

[0002] The field of medical data analysis technology involves comprehensive research on the extraction of disease-related features, parameter calculation, and pattern recognition using multi-source medical information. Its core aspects include physiological signal acquisition, motion behavior recognition, time series analysis, spatial trajectory reconstruction, and multi-dimensional feature fusion. This field provides objective evidence for disease diagnosis, progression monitoring, and efficacy evaluation by quantitatively analyzing multimodal data generated by patients in natural or task-specific states, and is widely used in research on neurological diseases, movement disorders, and rehabilitation medicine.

[0003] Traditional quantitative assessment of motor symptoms in Parkinson's disease involves clinicians manually scoring patients' motor performance using scales or relying on signal parameters from a single sensor to estimate the degree of motor impairment. This type of assessment describes motor signals using time-series statistical analysis and trajectory point calculation, reflecting motor function by extracting parameters such as gait cycle changes, hand tremor amplitude, and rhythmic characteristics. However, this traditional method primarily relies on one-dimensional or low-dimensional time-domain features for quantitative analysis, lacking a comprehensive utilization of the correlation between temporal and spatial characteristics, and thus failing to fully characterize the abnormal motor patterns in Parkinson's disease patients.

[0004] Existing technologies largely rely on single sensors to acquire motion signals, making it difficult to capture the full picture of motion patterns from multiple angles. This is especially true for complex compound movements or postural changes, where single-point signals are easily affected by external interference, leading to feature distortion. Temporal analysis methods have significant limitations, failing to simultaneously reveal spatial relationships between different body parts and unable to accurately capture the global trend of rhythmic changes over time. Assessment methods based on manual scales lack objectivity and are easily influenced by subjective factors, resulting in inconsistent and unstable assessment results, and failing to provide accurate and quantitative data on disease progression for clinical practice. Summary of the Invention

[0005] To address the limitations of existing technologies that rely heavily on single sensors to acquire motion signals, making it difficult to capture the full picture of motion patterns from multiple angles, especially in complex compound movements or postural changes, where single-point signals are easily affected by external interference, leading to feature distortion, and to improve the effectiveness of existing technologies, which often rely on single sensors to acquire motion signals, this invention provides a method for the spatiotemporal feature quantification assessment of motor symptoms in Parkinson's disease.

[0006] To achieve the above objectives, the present invention employs the following technical solution: a method for quantitatively assessing the spatiotemporal characteristics of motor symptoms in Parkinson's disease, comprising the following steps:

[0007] S1: By deploying inertial measurement unit sensors in the joints of the patient's limbs and trunk to acquire three-dimensional acceleration data, angular velocity data, and magnetic field strength data, a connection structure with sensors as nodes and human limb connections as edges is constructed to generate a connection neural network topology.

[0008] S2: Call the coordinate parameters of the node trajectory points in the connected neural network topology, calculate the change angle and angular velocity of the direction change of adjacent trajectory segments, use the adaptive weight decay algorithm to judge the stability of the motion features, and generate the weight parameters of the connected convolutional network.

[0009] S3: The node adjacent motion feature parameters are weighted and aggregated according to the weight parameters of the connected convolutional network. Based on the propagation mechanism of the connected neural network, the motion association information between nodes across topological levels is processed through multiple rounds of connected convolution to form a spatiotemporal fusion motion feature set.

[0010] S4: Based on the spatiotemporal fusion motion feature set, temporal feature data is obtained by arranging it in time sequence. A multi-head attention mechanism is used to identify temporal patterns of tremor frequency peak, gait cycle changes, and motion amplitude fluctuations, and to extract temporal feature patterns of Parkinson's disease symptoms.

[0011] As a further embodiment of the present invention, the connected neural network topology includes node spatial distribution parameters, node adjacency matrix weight parameters, and human body part topological relationship parameters; the connected convolutional network weight parameters include node feature weight coefficients, edge connection weight coefficients, and stability decay weights; the spatiotemporal fusion motion feature set includes local spatiotemporal coupling features, global motion correlation features, and multi-layer propagation fusion features; and the Parkinson's disease symptom temporal feature patterns include tremor frequency patterns, gait rhythm patterns, and motion amplitude fluctuation patterns.

[0012] As a further aspect of the present invention, the specific steps of S1 are as follows:

[0013] S101: By deploying inertial measurement unit sensors in the joints of the patient's limbs and trunk, the system monitors acceleration signals, angular velocity signals, and magnetic field strength signals, performs inter-frame alignment of timestamps and performs amplitude normalization calculations, and generates a sensor node index set based on sensor identifiers and spatial location information.

[0014] S102: Based on the sensor node index set, obtain the three-dimensional coordinate data between nodes, calculate the Euclidean distance vector and orientation angle of adjacent nodes, compare the distance vector with the limb connection judgment threshold, record the node number pairs whose distance vector is lower than the threshold and encode and integrate them to generate a limb connection edge set.

[0015] The threshold for determining limb connection is determined based on the average limb length statistics from the clinical human movement database.

[0016] S103: Based on the limb connection edge set and sensor node index set, perform aggregation calculation and matrix encoding on the node data vector, calculate the correspondence between node attributes and edge connection weights based on the encoding results, and obtain the topology of the connected neural network.

[0017] As a further aspect of the present invention, the specific steps of S2 are as follows:

[0018] S201: Obtain the coordinate parameters of the node trajectory points in the connected neural network topology, perform a difference operation on the coordinate vectors of adjacent nodes, extract the trajectory segment direction change amplitude sequence based on the comparison result of the direction vector difference and the direction change reference value, and generate a direction change angle sequence set.

[0019] The reference value for directional change is dynamically determined based on the statistical mean or standard deviation of the directional change of the trajectory segment in normal motion state in clinical samples.

[0020] S202: Based on the set of direction change angle sequences, perform time difference on the angle difference between trajectory segments, calculate the angle change gradient, and then determine the angular velocity change based on the ratio of the change gradient to the sampling time interval of the trajectory points, and establish a set of angular velocity change parameters;

[0021] S203: Call the set of angular velocity change parameters and pair them with the sequence of trajectory segment direction change amplitude. Perform adaptive weight decay processing on the fluctuation amplitude of the paired data. Determine the stability of motion features based on the convergence characteristics of the decay coefficient and generate weight parameters for the connected convolutional network.

[0022] As a further aspect of the present invention, the specific steps of S3 are as follows:

[0023] S301: Based on the weight parameters of the connected convolutional network, perform weighted calculation on the motion feature parameters in the node adjacency relationship, aggregate features according to the node connection weight ratio, reconstruct the motion vector structure of the local neighborhood of the node, and generate a weighted motion representation set of the node;

[0024] S302: Based on the node weighted motion representation set, using the neighbor node index relationship of the connected neural network propagation mechanism, perform multi-round propagation aggregation on the node representation vector, calculate feature similarity, and filter node vectors whose similarity values ​​reach the similarity threshold to perform feature fusion, thereby generating a global propagation motion feature set;

[0025] The similarity threshold is determined based on the upper limit of the 95% confidence interval of the training sample distribution or the model's adaptive clustering segmentation point;

[0026] S303: For the global propagation motion feature set, based on the time index identifier and spatial node coordinate parameters, perform correlation calculation on the time component and spatial component and analyze the spatiotemporal coupling relationship to generate a spatiotemporal fusion motion feature set.

[0027] As a further aspect of the present invention, the specific steps of S4 are as follows:

[0028] S401: Obtain the peak value of tremor frequency, gait cycle change and motion amplitude fluctuation data frames from the spatiotemporal fusion motion feature set, perform linear difference calculation on the tremor frequency difference between adjacent time segments, and filter out the mutation indexes with difference values ​​greater than the frequency fluctuation threshold to generate a tremor frequency time series sequence.

[0029] S402: Based on the tremor frequency time series, call the gait cycle change and motion amplitude fluctuation series, perform normalization fusion on the feature difference under the same time index, arrange them on a unified time axis to form a continuous sequence, and generate a multi-dimensional time series feature sequence set;

[0030] S403: Based on the multidimensional temporal feature sequence set, perform multi-head attention weight calculation on the time window, perform weighted aggregation and temporal matching based on the correlation between channels, and generate a temporal feature pattern of Parkinson's disease symptoms.

[0031] As a further aspect of the present invention, the frequency fluctuation threshold is determined based on clinical tremor frequency statistics and the individual physiological fluctuation range of the patient.

[0032] The multi-head attention weight calculation for the time window refers to extracting continuous feature segments using a fixed-length sliding time window in a multi-dimensional temporal feature sequence set, calculating the inter-channel feature similarity matrix for each time window through a multi-head attention mechanism, aggregating the multi-head output results based on the channel weighting coefficients, and completing feature matching through temporal position encoding to form a temporal feature pattern of Parkinson's disease symptoms.

[0033] As a further aspect of the present invention, the method further includes step S5:

[0034] S5: Call the tremor severity parameter, gait abnormality parameter, and limb coordination parameter in the Parkinson's disease symptom time sequence feature pattern, calculate the symptom type quantitative score value, and generate the Parkinson's disease motor symptom quantitative assessment result;

[0035] The quantitative assessment results of motor symptoms of Parkinson's disease include tremor severity score, gait abnormality score, and limb coordination score.

[0036] As a further aspect of the present invention, the specific steps of S5 are as follows:

[0037] S501: Based on the tremor severity parameters, gait abnormality parameters, and limb coordination parameters extracted from the Parkinson's disease symptom time-series feature pattern, normalize and differ the parameter time series, and aggregate the calculation results according to the time index to generate a time-series feature deviation dataset.

[0038] S502: Call the multi-dimensional parameters in the time-series feature deviation dataset, perform a weighted comparison between the tremor severity parameter and the gait abnormality parameter, filter out the parameter group with the difference exceeding the limit based on the motion abnormality benchmark value, and perform integral mapping to generate a motion abnormality feature vector set;

[0039] S503: Based on the parameter amplitudes in the set of abnormal movement feature vectors, perform linear mapping on the limb coordination parameters, and aggregate them with weighted sums of tremor severity and gait abnormality, extract the mean result, and generate a quantitative assessment result of Parkinson's disease motor symptoms.

[0040] As a further aspect of the present invention, the motion abnormality baseline value is determined based on the original statistical characteristics of the tremor severity parameter and gait abnormality parameter formed during multiple continuous monitoring of Parkinson's disease patients.

[0041] The linear mapping of the parameter amplitudes in the motion abnormality feature vector set to the limb coordination parameters refers to adjusting the limb coordination parameters linearly according to the amplitude change ratio of the parameters in the motion abnormality feature vector set, so that the mapping result is consistent with the dynamic change characteristics of the parameters.

[0042] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0043] In this invention, a topological connection structure for motion data is constructed through a sensor network, enabling the integration of site-specific motion information across multiple dimensions, capturing dynamic coordination between limbs, and thus achieving global feature extraction of tremor and gait cycles. The introduced adaptive weight decay and connected convolutional networks deeply aggregate motion features, maintaining the stability of time-series features and capturing motion correlations between remote nodes through multi-round convolutional propagation, achieving a comprehensive representation of motion abnormalities. The temporal feature extraction combined with a multi-head attention mechanism allows for more detailed identification of tremor frequency and gait changes in temporal patterns, effectively improving the ability to identify Parkinson's disease symptoms and forming a more objective and consistent assessment system. Attached Figure Description

[0044] To more clearly illustrate the technical solutions in the embodiments of the present invention, the attached connections used in the embodiments will be briefly introduced below. Obviously, the attached connections described below are only some embodiments of the present invention. For those skilled in the art, other attached connections can be obtained based on the attached connections without creative effort.

[0045] Figure 1 The following is a schematic diagram of the steps of this invention. Detailed Implementation

[0046] The technical solution of the present invention will now be described in conjunction with the attached link.

[0047] In embodiments of the present invention, words such as "exemplarily," "for example," etc., are used to indicate that something is an example, illustration, or description. Any embodiment or design described as "exemplary" in the present invention should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of the word "exemplary" is intended to present the concept in a concrete manner. Furthermore, in embodiments of the present invention, the meaning expressed by "and / or" can be both, or either one.

[0048] In the embodiments of the present invention, "connecting image" and "connecting piece" may sometimes be used interchangeably. It should be noted that when their distinction is not emphasized, their intended meanings are consistent. "of", "corresponding (relevant)" and "corresponding" may sometimes be used interchangeably. It should be noted that when their distinction is not emphasized, their intended meanings are consistent.

[0049] In this embodiment of the invention, sometimes a subscript such as W1 may be written in a non-subscript form such as W1. When the difference is not emphasized, the meaning they express is the same.

[0050] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the attached links and specific embodiments.

[0051] Please see Figure 1 This invention provides a method for quantitatively assessing the spatiotemporal characteristics of motor symptoms in Parkinson's disease, comprising the following steps:

[0052] S1: By deploying inertial measurement unit sensors in the joints of the patient's limbs and trunk to acquire three-dimensional acceleration data, angular velocity data, and magnetic field strength data, a connection structure with sensors as nodes and human limb connections as edges is constructed to generate a connection neural network topology.

[0053] S2: Call the coordinate parameters of the node trajectory points in the connected neural network topology, calculate the change angle and angular velocity of the direction change of adjacent trajectory segments, use the adaptive weight decay algorithm to judge the stability of motion features, and generate the weight parameters of the connected convolutional network.

[0054] S3: The motion feature parameters of adjacent nodes are weighted and aggregated according to the weight parameters of the connected convolutional network. Based on the propagation mechanism of the connected neural network, the motion association information between nodes across topological levels is processed through multiple rounds of connected convolution to form a spatiotemporal fusion motion feature set.

[0055] S4: Based on the spatiotemporal fusion motion feature set, temporal feature data is obtained by arranging it in time series. A multi-head attention mechanism is used to identify temporal patterns of tremor frequency peak, gait cycle changes, and motion amplitude fluctuations, and extract temporal feature patterns of Parkinson's disease symptoms.

[0056] S5: Call the parameters of tremor severity, gait abnormality, and limb coordination in the Parkinson's disease symptom time-series feature pattern, calculate the quantitative score of symptom type, and generate the quantitative assessment results of Parkinson's disease motor symptoms.

[0057] The connection neural network topology includes node spatial distribution parameters, node adjacency matrix weight parameters, and human body part topological relationship parameters. The connection convolutional network weight parameters include node feature weight coefficients, edge connection weight coefficients, and stability decay weights. The spatiotemporal fusion motion feature set includes local spatiotemporal coupling features, global motion correlation features, and multi-layer propagation fusion features. The Parkinson's disease symptom temporal feature patterns include tremor frequency patterns, gait rhythm patterns, and motion amplitude fluctuation patterns. The quantitative assessment results of Parkinson's disease motor symptoms include tremor severity scores, gait abnormality scores, and limb coordination scores.

[0058] The specific steps of S1 are as follows:

[0059] S101: By deploying inertial measurement unit sensors in the joints of the patient's limbs and trunk, the system monitors acceleration signals, angular velocity signals, and magnetic field strength signals, performs inter-frame alignment of timestamps and performs amplitude normalization calculations, and generates a sensor node index set based on sensor identifiers and spatial location information.

[0060] By examining the patient's left wrist, left elbow, left shoulder, right wrist, right elbow, right shoulder, sternum, left hip, left knee, left ankle, right hip, right knee, right ankle, pelvis, and head, a total of Inertial measurement unit sensors deployed in various locations monitor and acquire their triaxial acceleration signals, triaxial angular velocity signals, and triaxial magnetic field strength signals, for example, in At a sampling rate of Hz, the results are shown in Table 1. Partial raw data near the time point.

[0061] Table 1: Sensor Raw Signal Sampling Table

[0062]

[0063] As shown in Table 1, the timestamps of the collected data frames exist. ms, ms, In cases of misalignment such as milliseconds, first establish a standard time grid, for example, using... ms is the interval ( ms), the raw timestamp of each sensor ms, ms, ms and ms and The milliseconds (ms) perform a difference comparison and select the standard time point with the smallest absolute difference as the aligned timestamp. For example... ms distance ms is ms, distance ms is ms, therefore ms and their corresponding data frames are classified into ms time point, MS is similarly categorized. ms, MS is also included. After completing inter-frame alignment, the next step is to perform amplitude normalization calculation, which retrieves the maximum physical range value of the pre-set signal channel. and minimum value For example, accelerometer ( ), ( ), angular velocity ( ), ( After alignment At time ms, the acceleration X value of IMU-001 is Its normalized calculation is The angular velocity Y of IMU-001 is Its normalized calculation is ,right All of the sensors All axis data undergo this normalization calculation. Finally, an integer index is built based on the sensor's factory-unique identifier (e.g., 'IMU-001') and its spatial location information determined in the deployment plan (e.g., 'left wrist'). The mapping relationship with sensors (identifiers, locations), for example Corresponding to ('IMU-001', 'left wrist'), Corresponding to ('IMU-002', 'left elbow'), ..., Corresponding to ('IMU-015', 'header'), generate a sensor node index set.

[0064] S102: Based on the sensor node index set, obtain the three-dimensional coordinate data between nodes, calculate the Euclidean distance vector and orientation angle of adjacent nodes, compare the distance vector with the limb connection judgment threshold, record the node number pairs whose distance vector is lower than the threshold and encode and integrate them to generate the limb connection edge set.

[0065] The threshold for determining limb connectivity is determined based on the average limb length statistics from the clinical human movement database.

[0066] Based on the sensor node index set I={1: left wrist, 2: left elbow, 3: left shoulder, ...}, each index node is first obtained through attitude calculation and integral localization using aligned and normalized inertial signal data. Aligning timestamps Three-dimensional spatial coordinates of time For example, in At that moment, node 1 (left wrist) was acquired. Node 2 (left elbow) Node 3 (left shoulder) The coordinate units are all meters. Next, for the predefined human skeletal connections in anatomical terms, such as the left forearm (connected by node 1 and node 2) and the left upper arm (connected by node 2 and node 3), the Euclidean distance between adjacent nodes is calculated. , The calculation process is as follows:

[0067] ;

[0068] The calculation process is as follows:

[0069] ;

[0070] Simultaneously calculate the direction angle, for example Compared to The azimuth and elevation angles are then used to calculate the distance. Threshold for determining limb connection Perform a comparison, the threshold The determination was based on the Clinical Human Motion Database, which statistically analyzes the average limb length of humans. and standard deviation For example, the average length of the left forearm meters, standard deviation Meters, average length of left upper arm meters, standard deviation Meters, threshold Set as ,Right now rice, Meters, performing comparisons ( (meters) and ( rice), If the condition is true, record the node number pair. Perform comparison ( (meters) and ( rice), If the condition is true, record the node number pair. If the calculated distance of a pair of nodes (e.g., node 1 and node 3) Meters, and its corresponding threshold (if it exists). for Rice, then Not less than Do not record this number pair. Node number pairs recorded at any time The encoding is integrated to generate a set of limb connection edges.

[0071] S103: Based on the limb connection edge set and sensor node index set, perform aggregation calculation and matrix encoding on the node data vector, calculate the correspondence between node attributes and edge connection weights based on the encoding results, and obtain the topology of the connected neural network.

[0072] Based on the generated limb connection edge set With the generated sensor node index set (in First, an aggregation calculation is performed on the node data vectors. This refers to the retrieved data corresponding to the index. of, in Momentary, normalized Axis signal data, for example Vector of time node 1 This is 3D vectors, aggregation calculation refers to... Each node The dimensional vectors are aggregated, and then matrix encoding is performed to convert them into vectors. indivual dimensional vector Stacked in order of node index, forming a (Right now ) node attribute matrix ,in The A row is a node eigenvectors At the same time, using edge sets Build a (Right now The adjacency matrix of ) , elements Set as (If the side) Existence (Chinese) or (If the edge does not exist), then, according to the matrix and These two encoding results calculate node attributes and edge connection weights. The correspondence between them The calculation is based on the obtained Euclidean distance between nodes. ,set up The calculation process is as follows (when (time), for example Rice, then , Rice, then Use the calculated weights Replace adjacency matrix The corresponding The weighted adjacency matrix is ​​obtained. ,in , , , , If the weights of the remaining positions are China but China is also This kind of The settings reflect the nodes Attributes (coordinates) and nodes The spatial relationship between attributes (coordinates), and and Combinations (e.g., through matrix multiplication) This describes in detail how node attributes are passed and associated through edge connection weights to obtain the topology of the connected neural network.

[0073] The specific steps of S2 are as follows:

[0074] S201: Obtain the coordinate parameters of the trajectory points of the nodes in the connected neural network topology, perform difference operations on the coordinate vectors of adjacent nodes, extract the trajectory segment direction change amplitude sequence based on the comparison result of the direction vector difference and the direction change reference value, and generate a set of direction change angle sequence.

[0075] The baseline value for directional change is dynamically determined based on the statistical mean or standard deviation of the directional changes of the trajectory segments in normal motion states in clinical samples;

[0076] Obtain the generated connected neural network topology, specifically by retrieving node 1 (left wrist) at consecutive timestamps. The three-dimensional coordinate parameters of the trajectory points, for example. , , , The time interval is First, a difference operation is performed on the coordinate vectors of adjacent nodes to obtain the direction vector between two adjacent timestamps. For example, arrive Direction vector Calculated as , arrive Direction vector Calculated as , arrive Direction vector Calculated as Then calculate the angle between consecutive direction vectors, i.e., the magnitude of the direction change, for example... and The angle between The calculation begins with the vector dot product:

[0077] ,

[0078] Then calculate the vector magnitude. , Then the included angle Obtained by the inverse cosine function Similarly, calculate and The angle between Next, the calculated direction vector difference (i.e., the included angle) is compared with the reference value for direction change. The setting refers to a set containing A clinical sample database of healthy subjects performing the "finger-nose" test was used to analyze their wrist trajectory segments in each... The angle of change of direction within a ms interval is used to obtain the statistical mean. The standard deviation is , base value Set as The calculated included angle ( )and ( ) for comparison, The value was determined to be a significant change. Extract and store the sequence of changes in the direction of the trajectory segment, and set the included angle. ( )and ( ) for comparison, If the change is deemed insignificant, the value is discarded or marked as normal fluctuation. This extraction process is repeated for the entire movement of the node to generate a set of direction change angle sequences.

[0079] S202: Based on the set of direction change angle sequences, perform time difference on the angle difference between trajectory segments, calculate the angle change gradient, and then determine the angular velocity change based on the ratio of the change gradient to the sampling time interval of trajectory points, and establish a set of angular velocity change parameters;

[0080] Based on the generated set of orientation change angle sequences, for example, a sequence segment extracted from node 1 (left wrist) at multiple consecutive timestamps is... The timestamp corresponding to the data point is First, time difference is performed on the angle difference between the trajectory segments to calculate the gradient of the angle change. For example, in Gradient of angle change at time for ,exist Gradient of angle change at time for ,exist Gradient of angle change at time for The gradient value reflects the rate of change in the direction of motion. The change in angular velocity is then determined based on the ratio of this gradient to the sampling time interval of the trajectory points. In this embodiment, ms, i.e. s, change in angular velocity The calculation process is as follows Therefore, in Change in angular velocity at time t ,exist Change in angular velocity at time t ,exist Change in angular velocity at time t The change in angular velocity is essentially the angular acceleration of the change in trajectory direction, as shown in Table 2.

[0081] Table 2: Calculation Table of Nodal Angular Velocity Change

[0082]

[0083] Table 2 lists some of the calculation data for a single node, summarizing the angular velocity changes calculated from the timestamps. For all Each node performs the same calculation process to establish a set of angular velocity change parameters.

[0084] S203: Call the angular velocity change parameter set and pair it with the trajectory segment direction change amplitude sequence. Perform adaptive weight decay processing on the fluctuation amplitude of the paired data. Determine the stability of motion features based on the convergence characteristics of the decay coefficient and generate the weight parameters of the connected convolutional network.

[0085] Call the established set of angular velocity change parameters, for example, node 1 at... Angular velocity change sequence at time t. And compared with the obtained, unfiltered original trajectory segment direction change amplitude sequence, i.e. At the corresponding timestamp The values ​​are paired and mapped to form a time series tuple, for example:

[0086] , , ;

[0087] Next, adaptive weight decay processing is performed on the fluctuation amplitude of the paired data. First, an initial weight is set. And set an attenuation coefficient. This attenuation coefficient is set based on empirical data, for the tremor characteristics of Parkinson's disease (in... The capture rate (Hz) is higher than average, indicating a sensitivity of over 100%. Calculate the fluctuation amplitude at each time point. Assuming for ,but Update weights ,set up ,but ,exist Calculate the fluctuation amplitude at each time point. Update weights:

[0088] ,

[0089] This iterative calculation continues, based on the decayed weights. To determine the stability of motion characteristics based on convergence properties, a convergence judgment window is set, for example... time steps ( (ms), and set a convergence threshold. ,exist At time 1, calculate the change in weights within the calculation window. ,Will and If a comparison is made, If the weights converge, it indicates that the motion pattern has entered a stable state. If this pattern persists, it indicates an unstable movement pattern with continuous and significant fluctuations. Finally, the final convergent weight value calculated over a period of time (e.g., a complete movement cycle), or the statistical characteristics of the weight sequence in the non-convergent case (e.g., the mean), will be used. This serves as a quantitative indicator for the node during this motion phase, generating the weight parameters for the connected convolutional network.

[0090] The specific steps for S3 are as follows:

[0091] S301: Based on the weight parameters of the connected convolutional network, the motion feature parameters in the node adjacency relationship are weighted and calculated. The features are aggregated according to the node connection weight ratio to reconstruct the motion vector structure of the local neighborhood of the node and generate a weighted motion representation set of the node.

[0092] Based on the generated connected convolutional network weight parameters, for example, the stability weights of node 2 (left elbow) during a specific motion phase are obtained. Simultaneously, the original motion feature parameters are retrieved, specifically referring to the adjacent nodes of node 2, i.e., node 1 (left wrist) and node 3 (left shoulder). Moment 3D normalized signal vector, and First, a weighted calculation is performed on the motion characteristic parameters in the adjacency relationship. This calculation includes the stability weight of node 2 itself. Multiply by the eigenvectors of the adjacent nodes, i.e.:

[0093] ,

[0094] as well as ,

[0095] Next, feature aggregation is performed based on the node connection weight ratio. Here, the connection weights are derived from the calculated weighted adjacency matrix. ,in , First, calculate the sum of the adjacency weights. Then calculate the weight ratio of the connections. , The feature aggregation process involves multiplying the weighted feature vectors of adjacent nodes by the proportion of their connection weights, and then summing the results. Specifically... Example of calculating the first dimension of a vector: Apply this calculation Each dimension yields a new aggregated feature vector. The aggregation process also needs to add the feature vector of node 2 itself. Through such calculations, node 2 was reconstructed. The structure of the local neighborhood motion vector at time t, this new vector It no longer represents only the movement of the left elbow, but integrates the movement information of the left wrist and left shoulder based on the physical connection tightness, for Each node performs this process on a timestamp to generate a weighted motion representation set for the nodes.

[0096] S302: Based on the weighted motion representation set of nodes, the neighbor node index relationship of the connection neural network propagation mechanism is used to perform multi-round propagation aggregation on the node representation vectors, calculate feature similarity, and select node vectors whose similarity values ​​reach the similarity threshold to perform feature fusion and generate a global propagation motion feature set;

[0097] The similarity threshold is determined based on the upper limit of the 95% confidence interval of the training sample distribution or the model's adaptive clustering split point;

[0098] Based on the generated node weighted motion representation set, for example, the representation vector of node 1 (left wrist) obtained after the first round of propagation is: The representation vector of node 8 (right hip) is Using the adjacency index relationship of the neural network propagation mechanism, a second round of propagation aggregation is performed on the node representation vector. Taking node 1 as an example, its adjacency node is node 2 (left elbow). The second round of calculation will retrieve the representation vector obtained by node 2 after the first round of propagation. And based on the connection weight right Perform transformations and aggregations to generate the second-round representation vector for node 1. After being preset After round-robin aggregation, the final node representation vector is obtained, for example, the final vector of node 1. The final vector of node 7 (sternum) Next, we calculate the feature similarity between these two nodes. Here, we use cosine similarity for calculation, which involves calculating the vectors... and The dot product, divided by the product of the magnitudes of the two vectors, for example, the dot product... Vector magnitude , The similarity value is Then, select node vectors whose similarity values ​​reach a similarity threshold. The determination is based on a set of... A training sample database of motor data from Parkinson's disease patients was used. The feature similarity of node pairs during tremor episodes was calculated to form a distribution. The upper limit of the 95% confidence interval of this distribution was then taken, and the value was obtained through statistical calculation. Therefore, set The calculated similarity value and Comparison, The motion characteristics of node 1 and node 7 are determined to be highly similar, therefore the vectors of these two nodes are... and Perform feature fusion, which involves averaging the corresponding elements of the two vectors. For example, the first dimension of the fused vector might be... The fused vector is used as the global feature representation shared by the two nodes. This process is repeated for the node pair to generate a global propagation motion feature set.

[0099] S303: For the global propagation motion feature set, based on the time index identifier and spatial node coordinate parameters, perform correlation calculations on the time components and spatial components and analyze the spatiotemporal coupling relationship to generate a spatiotemporal fusion motion feature set;

[0100] For the generated global propagation motion feature set, where each node corresponds to a fused high-dimensional feature vector at each time point, for example, in Under a series of time index identifiers, the feature vector sequence shared by node 1 (left wrist) and node 7 (sternum) is as follows: Meanwhile, based on the spatial node coordinate parameters, for example, the feature vector sequence of node 10 (left ankle) itself is... Next, the temporal and spatial components of the feature vectors are correlated and calculated, as shown in Table 3.

[0101] Table 3: Example Table of Spatiotemporal Feature Correlation Calculation

[0102]

[0103] As shown in Table 3, the correlation calculation here specifically involves calculating the cross-correlation function between the time series of the "left wrist-sternum" fusion feature and the time series of the "left ankle" feature to analyze the spatiotemporal coupling relationship between the two. The calculation process involves time-shifting the "left ankle" sequence relative to the "left wrist-sternum" sequence, for example, shifting... time steps (i.e.) (ms), then calculate the sum of the products of corresponding elements in the two translated sequences, normalize this sum to obtain the cross-correlation coefficient, such as in The cross-correlation coefficient calculated at ms is This value is defined as Between, when the absolute value of the cross-correlation coefficient is higher than At that time, it was determined to be a strong coupling, between and When the coupling is between these values, it is considered moderate coupling, below the specified value. When the coupling is weak or non-coupling, the calculation result is determined. This is within a strongly coupled region, indicating that the patient's left ankle movement pattern lags temporally behind the movement patterns of its upper limb core region (wrist and sternum). The correlation between the two motions is very high. The correlation values ​​of the node pairs under different time lags are concatenated or stacked with the original global feature vector to form a feature matrix with higher dimension and richer information. This matrix contains the fine motion features of each body part and also encodes the transmission and delay relationship of motion patterns between different body parts, generating a spatiotemporal fusion motion feature set.

[0104] The specific steps of S4 are as follows:

[0105] S401: Acquire data frames of tremor frequency peak, gait cycle change and motion amplitude fluctuation from the spatiotemporal fusion motion feature set, perform linear difference calculation on the tremor frequency difference between adjacent time segments, and filter out abrupt change indices with difference values ​​greater than the frequency fluctuation threshold to generate a tremor frequency time series sequence.

[0106] From the generated spatiotemporal fusion motion feature set, extract the feature associated with the left wrist node (node ​​1) with a time window length of [missing information]. seconds, sliding step size For each second of data frame, a Fourier transform is performed on the acceleration signal within that frame, and the peak frequency in the energy spectrum is extracted as the peak flutter frequency for that time segment. For example, in a continuous... Time segments (index) The peak values ​​of the tremor frequencies obtained from the data are as follows: Hz, Hz, Hz, Hz, firstly, perform linear difference calculation on the flutter frequency difference between adjacent time segments, that is, calculate Hz, Hz, Hz, then filter the mutation indexes whose difference value is greater than the frequency fluctuation threshold, which is the frequency fluctuation threshold. The settings refer to a set containing A database of movement data from healthy adults during normal walking and upper limb swinging tasks was used to statistically analyze the movement of limb nodes in any continuous sequence. The change in the dominant frequency within a window of seconds, the mean of the change is obtained. Hz, standard deviation is Hz, threshold Set as Hz, then the calculated difference value is compared with this threshold, the first difference value Hz and Hz comparison, The second difference value is determined to be a normal fluctuation. Hz and Hz comparison, If it is determined to be a frequency mutation, its corresponding time segment index is recorded. The third difference value Hz and Hz comparison, If the fluctuation is deemed normal, the original frequency data will be processed through this process. The corresponding mutation information is integrated to generate a tremor frequency time series.

[0107] S402: Based on the tremor frequency time series, call the gait cycle change and motion amplitude fluctuation series, perform normalization fusion on the feature difference under the same time index, arrange them on a unified time axis to form a continuous sequence, and generate a multi-dimensional time series feature sequence set;

[0108] Based on the generated tremor frequency time series, for example in arrive Time period sequence Simultaneously, it calls the gait cycle change sequence extracted under the same time segment index, for example... And the motion amplitude fluctuation sequence (measured by the standard deviation of wrist joint range of motion), for example First, normalization is performed on the feature values ​​of the three sequences at the same time index. This process requires determining the observation period for each feature (e.g., ...). The maximum and minimum values ​​within (seconds), assuming the maximum value of the tremor frequency is Hz, minimum value Hz, the maximum value of the gait period is s, minimum value s, the maximum value of the motion amplitude is m, minimum value m, with Taking the data at any given time as an example, the tremor frequency The normalized calculation of Hz is as follows gait cycle The normalized calculation of s is as follows Motion amplitude The normalized calculation of m is as follows The normalized values ​​are then fused; here, fusion refers to combining them into a multi-dimensional vector, for example, in... At time t, the resulting feature vector is As shown in Table 4.

[0109] Table 4: Normalized Representation of Multidimensional Temporal Features

[0110]

[0111] Table 4 lists the multidimensional feature vectors under continuous-time indexing, arranging the vectors on a unified time axis. The elements are arranged in sequence to form a continuous time series. Each element in the sequence is a three-dimensional feature vector, generating a multi-dimensional time series feature sequence set.

[0112] S403: Based on the multi-dimensional temporal feature sequence set, perform multi-head attention weight calculation on the time window, perform weighted aggregation and temporal matching according to the correlation between channels, and generate a temporal feature pattern of Parkinson's disease symptoms.

[0113] Based on the generated multidimensional temporal feature sequence set, for example, extracting a sequence containing... time steps ( to Using a time window as the processing unit, multi-head attention weights are first calculated for the data within that window, with time steps... eigenvectors For example, a query vector is generated by mapping the data through three independent transformation matrices (query, key, and value matrices, respectively). Key vector Sum value vector Next, calculate Key vectors of time steps within the window ( to The dot product of ) yields a set A raw attention score, for example, calculated , , The fractions are scaled and processed by a normalized exponential function to convert them into a set of sums. Attention weights, for example, right The weight is The weight of itself is ,right The weight is Then, multiply the weights by the value vector corresponding to each time step. And sum them up to get The new representation under this attention head is repeated in parallel in another independent attention head, and the results of the two heads are concatenated. Next, a weighted aggregation is performed based on the correlation between channels, where channels refer to the three dimensions of tremor frequency, gait cycle, and motion amplitude. The Pearson correlation coefficients between each pair of these three feature sequences within the current time window are calculated to obtain, for example, the (frequency-amplitude) correlation. (High correlation), (frequency-gait) correlation is (Moderate negative correlation), (amplitude-gait) correlation is (Moderate negative correlation), set an aggregation weight that is proportional to the absolute value of the correlation. For example, the weight assigned to the frequency channel is... The weight of the amplitude channel is The weights of the gait channels are The vectors processed by the attention mechanism are weighted and summed to obtain a final aggregated feature vector. Finally, temporal matching is performed. The aggregated feature vector sequence generated by the continuous time window is dynamically time-warped and the distance is calculated with the standard Parkinson's disease symptom patterns (such as the "end-of-dose phenomenon" pattern and the "dyskinesia" pattern) stored in the pattern library in advance. The pattern with the smallest distance is selected as the matching result of the current time period, and the temporal feature pattern of Parkinson's disease symptoms is generated.

[0114] Please refer to Connection 6. The specific steps for S5 are as follows:

[0115] S501: Based on the tremor severity parameters, gait abnormality parameters, and limb coordination parameters extracted from the temporal feature pattern of Parkinson's disease symptoms, normalization and differencing operations are performed on the parameter time series, and the calculation results are aggregated according to the time index to generate a temporal feature deviation dataset.

[0116] Based on the generated temporal feature pattern of Parkinson's disease symptoms, firstly in each Within a time window of milliseconds (ms), tremor severity parameters are extracted from the matched symptom patterns. For example, by analyzing the peak tremor frequency and fluctuations in motion amplitude, and combining this with clinical scoring criteria, a tremor severity parameter is quantified. arrive The values ​​between, where This indicates the absence of tremor. This represents profound tremor, and parameters related to the degree of gait abnormality are extracted. For example, based on the analysis of gait cycle changes, stride length variability, and swing phase duration, a value between [missing information] and [missing information] is also quantified. arrive The values ​​between, where This represents a normal gait. This represents an extremely abnormal gait and extracts limb coordination parameters, such as quantifying them by calculating the synchronicity and symmetry of the left and right limb movement trajectories. arrive The values ​​between, where Representing complete coordination, Representing extreme disharmony, in continuous Time window Inside, assuming the parameter sequence for the severity of tremor in the left wrist is as follows: The gait anomaly severity parameter sequence is as follows: The sequence of limb coordination parameters is Next, normalization is performed on the parameter time series. This normalization is a linear scaling of the maximum and minimum values ​​of each parameter throughout the patient's entire monitoring period. For example, the maximum value for tremor severity is... The minimum value is Then in The normalized value of the moment's tremor ,exist Time for Perform the same normalization on the parameter sequence to obtain the normalized sequence:

[0117] ;

[0118] ;

[0119] ;

[0120] The normalized parameter sequence is then subjected to differencing to capture the rate of change of the parameters. For example, the differencing sequence for tremor severity is:

[0121] ;

[0122] Similarly, calculate the difference sequences of gait and coordination:

[0123] ;

[0124] ;

[0125] Finally, the results are aggregated according to the time index to form a multidimensional time series, with each time index... At this point, the aggregation result is a set of three normalized values ​​and three difference values. dimensional vectors, for example, in At the time index, the aggregation result is Generate a time-series feature deviation dataset.

[0126] S502: Call the multi-dimensional parameters in the time series feature deviation dataset, perform a weighted comparison between the tremor severity parameter and the gait abnormality parameter, filter out the parameter group with the difference exceeding the limit based on the motion abnormality benchmark value, and perform integral mapping to generate a motion abnormality feature vector set;

[0127] Call the multidimensional parameters from the generated time-series feature bias dataset to... Vector at time index For example, among which It is a normalized tremor severity parameter , It is a normalized gait anomaly degree parameter First, a weighted comparison is performed between the normalized tremor severity parameter and the normalized gait abnormality parameter, and weights are set accordingly. and These two weights are set based on clinicians' experience regarding the relative importance of tremor and gait in the diagnosis of Parkinson's disease symptoms. Tremor is often more indicative in the diagnosis of Parkinson's disease in the resting phase, while gait is more critical in the later progression. Therefore, tremor is given a slightly higher weight. The weighted comparison value is calculated as follows:

[0128] ,

[0129] Next, based on the motion anomaly baseline value, parameter groups with excessive differences are selected. Set as This value is based on The statistical mean of the differences between tremor and gait calculated using similar methods was added to the values ​​of ... The weighted comparison value, obtained by multiplying the standard deviation, is used to distinguish between normal fluctuations and potential pathological abnormalities. and Comparison, If the motion difference at that time point is determined to be within the normal range, then at another time point... Weighted comparison value ,same It is still considered normal, but if The sudden increase in the intensity of the tremor caused , ,but ,at this time If the motion difference at that time point is determined to be excessive, the weighted comparison value of the parameter group that exceeds the limit is determined. and their corresponding differential flutter values Sum of difference gait values An integral mapping is performed, which accumulates the out-of-limit values ​​over a continuous time window to reflect the persistence and cumulative effect of the anomaly. For example, if continuous... All time windows have exceeded the limit, so this Each window The values ​​are summed to obtain a cumulative outlier score, which is then combined with the difference parameter, for example... Together, they form a motion anomaly feature vector, generating a motion anomaly feature vector set.

[0130] S503: Based on the parameter amplitude in the set of abnormal movement feature vectors, perform linear mapping on the limb coordination parameters, and aggregate them with weighted averages of tremor severity and gait abnormality, extract the mean results, and generate quantitative assessment results of Parkinson's disease motor symptoms.

[0131] Based on the parameter magnitudes in the generated motion anomaly feature vector set, for example in Weighted comparison value in the motion anomaly feature vector identified at each time step As a comprehensive quantitative amplitude of tremor severity and gait abnormality, a linear mapping is performed on the extracted limb coordination parameters, for example, Normalized limb coordination parameters at different times Define a linear mapping function The slope and intercept It was obtained through regression analysis of clinical coordination scores and normalized data, aiming to amplify coordination parameters to an assessment range equivalent to tremor and gait. The mapped limb coordination parameters are: The mapped limb coordination parameter is then weighted and aggregated with the severity of tremor and the degree of gait abnormality. Here, the severity of tremor and the degree of gait abnormality are directly calculated using the obtained weighted comparison values. Set three weights (Used for weighted comparison values) (For the mapped limb coordination parameters), weights are assigned based on the clinical importance of the symptoms' impact on daily life. Considering the balanced role of tremor, gait, and coordination in Parkinson's disease assessment, relatively balanced weights are used here to calculate the aggregate value:

[0132] ,

[0133] Repeat this weighted aggregation process for the time window and associated nodes, and extract the mean of the aggregation results, for example, in During the monitoring period of minutes, a total of Each aggregate value (each (ms one), sum the aggregated values ​​and then divide by . For example, the final mean result is This generates quantitative assessment results of motor symptoms in Parkinson's disease.

[0134] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for quantitatively assessing the spatiotemporal characteristics of motor symptoms in Parkinson's disease, characterized in that, Includes the following steps: S1: By deploying inertial measurement unit sensors in the patient's limb joints and torso to acquire three-dimensional acceleration data, angular velocity data, and magnetic field strength data, a connection structure is constructed with sensors as nodes and human limb connections as edges, generating a connected neural network topology; the specific steps of S1 are as follows: S101: By deploying inertial measurement unit sensors in the joints of the patient's limbs and trunk, the system monitors acceleration signals, angular velocity signals, and magnetic field strength signals, performs inter-frame alignment of timestamps and performs amplitude normalization calculations, and generates a sensor node index set based on sensor identifiers and spatial location information. S102: Based on the sensor node index set, obtain the three-dimensional coordinate data between nodes, calculate the Euclidean distance vector and orientation angle of adjacent nodes, compare the distance vector with the limb connection judgment threshold, record the node number pairs whose distance vector is lower than the threshold and encode and integrate them to generate a limb connection edge set. S103: Based on the limb connection edge set and sensor node index set, perform aggregation calculation and matrix encoding on the node data vector, calculate the correspondence between node attributes and edge connection weights based on the encoding results, and obtain the topology of the connected neural network. S2: Call the coordinate parameters of the node trajectory points in the connected neural network topology, calculate the change angle and angular velocity of adjacent trajectory segments, use an adaptive weight decay algorithm to determine the stability of the motion features, and generate the weight parameters of the connected convolutional network; the specific steps of S2 are as follows: S201: Obtain the coordinate parameters of the node trajectory points in the connected neural network topology, perform a difference operation on the coordinate vectors of adjacent nodes, extract the trajectory segment direction change amplitude sequence based on the comparison result of the direction vector difference and the direction change reference value, and generate a direction change angle sequence set. S202: Based on the set of direction change angle sequences, perform time difference on the angle difference between trajectory segments, calculate the angle change gradient, and then determine the angular velocity change based on the ratio of the change gradient to the sampling time interval of the trajectory points, and establish a set of angular velocity change parameters; S203: Call the set of angular velocity change parameters and pair them with the sequence of trajectory segment direction change amplitude. Perform adaptive weight decay processing on the fluctuation amplitude of the paired data. Determine the stability of motion features based on the convergence characteristics of the decay coefficient. Generate weight parameters for the connected convolutional network. S3: Weighted aggregation calculation of node adjacency motion feature parameters is performed based on the weight parameters of the connected convolutional network. Motion association information between nodes across topological levels is processed through multiple rounds of connected convolutions according to the connected neural network propagation mechanism, forming a spatiotemporal fusion motion feature set. The specific steps of S3 are as follows: S301: Based on the weight parameters of the connected convolutional network, perform weighted calculation on the motion feature parameters in the node adjacency relationship, aggregate features according to the node connection weight ratio, reconstruct the motion vector structure of the local neighborhood of the node, and generate a weighted motion representation set of the node; S302: Based on the node weighted motion representation set, using the neighbor node index relationship of the connected neural network propagation mechanism, perform multi-round propagation aggregation on the node representation vector, calculate feature similarity, and filter node vectors whose similarity values ​​reach the similarity threshold to perform feature fusion, thereby generating a global propagation motion feature set; S303: For the global propagation motion feature set, based on the time index identifier and spatial node coordinate parameters, perform correlation calculation on the time component and spatial component and analyze the spatiotemporal coupling relationship to generate a spatiotemporal fusion motion feature set; S4: Based on the spatiotemporal fusion motion feature set, temporal feature data is obtained by arranging it in time sequence. A multi-head attention mechanism is used to identify temporal patterns of tremor frequency peak, gait cycle changes, and motion amplitude fluctuations, and to extract temporal feature patterns of Parkinson's disease symptoms.

2. The method for quantitative assessment of spatiotemporal characteristics of motor symptoms in Parkinson's disease according to claim 1, characterized in that, The connected neural network topology includes node spatial distribution parameters, node adjacency matrix weight parameters, and human body part topological relationship parameters. The connected convolutional network weight parameters include node feature weight coefficients, edge connection weight coefficients, and stability decay weights. The spatiotemporal fusion motion feature set includes local spatiotemporal coupling features, global motion correlation features, and multi-layer propagation fusion features. The Parkinson's disease symptom temporal feature patterns include tremor frequency patterns, gait rhythm patterns, and motion amplitude fluctuation patterns.

3. The method for quantitative assessment of spatiotemporal characteristics of motor symptoms in Parkinson's disease according to claim 1, characterized in that, The specific steps of S4 are as follows: S401: Obtain the peak value of tremor frequency, gait cycle change and motion amplitude fluctuation data frames from the spatiotemporal fusion motion feature set, perform linear difference calculation on the tremor frequency difference between adjacent time segments, and filter out the mutation indexes with difference values ​​greater than the frequency fluctuation threshold to generate a tremor frequency time series sequence. S402: Based on the tremor frequency time series, call the gait cycle change and motion amplitude fluctuation series, perform normalization fusion on the feature difference under the same time index, arrange them on a unified time axis to form a continuous sequence, and generate a multi-dimensional time series feature sequence set; S403: Based on the multidimensional temporal feature sequence set, perform multi-head attention weight calculation on the time window, perform weighted aggregation and temporal matching based on the correlation between channels, and generate a temporal feature pattern of Parkinson's disease symptoms.

4. The method for quantitative assessment of spatiotemporal characteristics of motor symptoms in Parkinson's disease according to claim 3, characterized in that, The frequency fluctuation threshold is determined based on clinical tremor frequency statistics and the individual physiological fluctuation range of the patient. The multi-head attention weight calculation for the time window refers to extracting continuous feature segments using a fixed-length sliding time window in a multi-dimensional temporal feature sequence set, calculating the inter-channel feature similarity matrix for each time window through a multi-head attention mechanism, aggregating the multi-head output results based on the channel weighting coefficients, and completing feature matching through temporal position encoding to form a temporal feature pattern of Parkinson's disease symptoms.

Images