CN116597940A

CN116597940A - Modeling method of movement disorder symptom quantitative evaluation model

Info

Publication number: CN116597940A
Application number: CN202310479465.9A
Authority: CN
Inventors: 王哲龙; 林芳; 刘芮辰; 石鑫
Original assignee: Dalian University of Technology
Current assignee: Dalian University of Technology
Priority date: 2023-04-10
Filing date: 2023-04-28
Publication date: 2023-08-15

Abstract

The invention relates to the technical field of movement disorder symptom evaluation, and provides a modeling method of a movement disorder symptom quantitative evaluation model, which comprises the following steps: step 100: collecting human body action data and establishing a human body action recognition model; step 200: extracting tremor characteristics and bradykinesia characteristics of a patient, and carrying out pearson correlation analysis; step 300: acquiring gait parameters of a patient; step 400: obtaining a dyskinesia feature set; step 500: based on the fully connected neural network, establishing a movement disorder symptom quantitative evaluation network; training the movement disorder symptom quantitative evaluation network by using the movement disorder feature training set, and testing the movement disorder symptom quantitative evaluation network by using the movement disorder feature testing set to finally obtain a movement disorder symptom quantitative evaluation model. The quantitative evaluation model provided by the invention can be used for accurately carrying out grading evaluation on the movement disorder symptoms of the patient.

Description

Modeling method of movement disorder symptom quantitative evaluation model

Technical Field

The invention relates to the technical field of movement disorder symptom evaluation, in particular to a modeling method of a movement disorder symptom quantitative evaluation model.

Background

Dyskinesia such as Parkinson's disease and cerebral apoplexy clinically can appear dyskinesia symptoms (bradykinesia, resting tremor, myotonia, gesture balance disorder, etc.), the dyskinesia symptoms appear first in patients with dyskinesia such as Parkinson's disease at the early stage, and dyskinesia is mainly represented in the middle and later stages of cerebral apoplexy, hemiplegia, etc. The evaluation of dyskinesia symptoms has important significance for diagnosis of dyskinesia diseases such as Parkinson and rehabilitation evaluation of diseases such as cerebral apoplexy.

Currently, clinical assessments for dyskinesia symptoms mostly rely on medical scales, for example, the most common dyskinesia disease, parkinson's disease, is assessed according to the unified parkinsonism rating scale of the institute of dyskinesia (Movement Disorder Society-Unified Parkinson's Disease Rating Scale, MDS-UPDRS). This evaluation method is widely used clinically, but has certain drawbacks: consuming a lot of time and having a certain degree of subjectivity in analysis, results in a difference between the evaluation results of different doctors. Meanwhile, different evaluation results have influence on the treatment and rehabilitation of diseases, and accordingly, certain difficulties exist in evaluating the treatment effect and rehabilitation effect.

In clinical assessment of dyskinesia symptoms, it is necessary to develop an intelligent, objective quantitative analysis of dyskinesia symptoms. At present, quantitative evaluation of dyskinesia symptoms is attempted by using inertial sensors, myoelectric sensors or video modes in the research fields of medical engineering and the like, but in clinic, the disease development degree of patients is different, all practical problems are difficult to solve by using a single sensor mode, and under practical conditions, the actions of the patients are random and disordered compared with the actions of normal people, so that the difficulty of action recognition and quantitative feature extraction is increased; when a plurality of researchers conduct quantitative research on dyskinesia symptoms, inertial data or video data of normal people and dyskinesia patients such as parkinsons are classified and identified through machine learning or deep learning, dyskinesia diseases such as normal people and parkinsons are effectively distinguished, quantitative and graded evaluation on dyskinesia symptoms is omitted, and in addition, the result of evaluation is inaccurate due to incomplete information acquired through a single sensor. The existing quantitative analysis technology of dyskinesia symptoms based on multi-modal information is complex in process, complex in operation and mostly requires laboratory environment, is not suitable for clinical application, and besides, extracted features are not fully extracted, and the degree of accuracy of quantitative classification is low.

The invention combines the multisource information to greatly improve the recognition performance of the clinical test action of the dyskinesia symptom, the multidimensional and efficient dyskinesia characteristics extracted by utilizing the multisystem information have very important value for the quantitative and graded evaluation of the high definition of the dyskinesia symptom, and the convenient, intelligent and objective quantitative analysis of the diseases has great significance for the treatment and rehabilitation evaluation of the diseases.

Disclosure of Invention

The invention mainly solves the technical problem that dyskinesia diseases such as Parkinson and the like and dyskinesia symptoms of cerebral apoplexy are difficult to comprehensively and objectively evaluate in the prior art, and provides a modeling method of a dyskinesia symptom quantitative evaluation model, which can effectively overcome the defect that clinical symptoms of patients are manually and subjectively analyzed.

The invention provides a modeling method of a quantitative evaluation model of dyskinesia symptoms, which comprises the following steps:

step 100: collecting human body action data and establishing a human body action recognition model;

step 200: extracting tremor characteristics and bradykinesia characteristics of a patient, and carrying out pearson correlation analysis;

step 300: acquiring gait parameters of a patient;

step 400: obtaining a dyskinesia feature set by using the tremor feature obtained in the step 200, the first bradykinesia feature, the second bradykinesia feature and the gait parameters obtained in the step 300, and taking the medical scale scores of neurology specialists as labels; randomly dividing the dyskinesia feature set into a dyskinesia feature training set and a dyskinesia feature test set according to a proportion, and randomly selecting a part of the dyskinesia feature training set as a dyskinesia feature verification set;

step 500: based on the fully connected neural network, establishing a movement disorder symptom quantitative evaluation network; training the movement disorder symptom quantitative evaluation network by using the movement disorder feature training set, and testing the movement disorder symptom quantitative evaluation network by using the movement disorder feature testing set to finally obtain a movement disorder symptom quantitative evaluation model.

Further, step 100 includes the following steps 101-103:

step 101: collecting human body action data by using N inertial sensor nodes respectively fixed at different parts of a human body; acquiring human body action videos by using video acquisition equipment;

step 102: window segmentation is carried out on the human body action data acquired by each wearable sensor node by utilizing a sliding window segmentation technology, so as to obtain a plurality of action data fragments of each wearable sensor node;

step 103: forming an action data set by utilizing a plurality of action data fragments of each inertial sensor node, dividing the action data set into an action data training set and an action data test set according to proportion, and randomly selecting a part of the action data training set as an action data verification set;

step 104: establishing an action recognition network;

step 105: training the action recognition network by utilizing the action data training set to obtain an action recognition model; verifying the action recognition model by using the action data verification set; testing the motion recognition model by using the motion recognition model test set to obtain a human motion recognition model with optimal performance; wherein, the action category identified by the human action identification model comprises: tremor, bradykinesia, or gait.

Further, the action recognition network includes: an FCN module and an LSTM module; the number of input variables is v; the variable is used as input in parallel, and the input length of each time sequence is l frames;

the FCN module comprises K convolution layers;

the time step of the LSTM module is set to be m; setting an attention mechanism before the LSTM module and setting a dropout layer after the LSTM module;

the mechanism of attention in the LSTM module is shown below, focusing the LSTM model on variables of different dimensions:

wherein w is _i Is v _i Is the sum of the weights of (1), each v _i The weight sigma of (2) _i Defined as sigma _i ＝f(ν _i )⊙ν _i ', as indicated by Hadamard product, f (. Cndot.) is a fully attached layer;

the cost function includes L ₂ Regularization and unbalance-like loss function two parts:

wherein, when the category is i, y _i =1, otherwise y _i =0; k is the number of categories, p is the output of the model; beta _i Calculated according to the following formula:

wherein n is _i The number of samples in class j is represented, and δ is a hyper-parameter.

Further, step 200 includes the following steps 201-204:

step 201: extracting tremor characteristics of the tremor action fragments of tremor actions; the tremor feature includes: root mean square, sample entropy, mean value, variance, covariance, skewness and kurtosis of the differences of acceleration, angular velocity and surface electromyographic signals;

step 202: for an action segment of bradykinesia action, extracted first bradykinesia features; extracting a second action delay characteristic from the corresponding human action video segment;

the first bradykinesia feature comprises: peak-to-peak value, rectified mean value, variance, standard deviation, skewness, kurtosis, root mean square, peak factor, margin factor, fuzzy entropy of acceleration and angular velocity; mutual information entropy, cross correlation coefficient, KL divergence of sEMG signals, fuzzy entropy and permutation entropy of sEMG; roll angle and pitch angle of the wrist; combining the acceleration, and the square root of the calculated number of the square sum of the triaxial angular velocities;

the second bradykinesia feature comprises: human body joint point information and joint angles;

step 203: and (3) carrying out pearson correlation analysis on the tremor characteristics obtained in the step 201, the first bradykinesia characteristics and the second bradykinesia characteristics obtained in the step 202 and medical scale scores assessed by neurology specialists, obtaining pearson correlation coefficients, and determining the effectiveness of the extracted characteristics according to the pearson correlation coefficients.

Further, step 300 includes the following steps 301-303:

step 301: calculating gait characteristics required by gait analysis through the acceleration data and the angular velocity data;

step 302: converting acceleration under a sensor coordinate system into a navigation coordinate system for double integration to obtain a space position under the navigation coordinate system; and correcting the time-space parameter drift;

step 303: the TCN is used for extracting the characteristics of the inertial signals, the TCN is used for realizing the information perception of gait time sequence, and the cavity convolution is used for expanding the receptive field of signals in the Parkinson gait analysis; according to the multi-layer residual error connection structure, gradual calculation of inertia signals from local to global is realized, and gait parameters capable of representing dyskinesia diseases such as Parkinson and the like and cerebral apoplexy are obtained.

According to the modeling method of the movement disorder symptom quantitative evaluation model, the movement disorder symptom quantitative evaluation model is built by mainly utilizing FCN, LSTM, attention mechanism and class unbalance loss function in the action recognition layer, and the model can effectively capture important variables in input variables and improve the ability of the model to learn class unbalance problems, so that the algorithm performance is improved; in the aspect of feature extraction, a CEEMDAN algorithm is utilized, parameters are set according to disease tremor characteristics, the frequency and the amplitude of acceleration signals in tremor action segments are extracted, the bradykinesia features of bradykinesia test actions are extracted based on myoelectricity information, inertia information and video information, the second bradykinesia features are extracted based on video, and gait time parameters, space parameters and joint parameter features are extracted by utilizing an extended Kalman filtering method based on quaternion by the inertia information and TCN; and at the quantitative evaluation level, a dyskinesia characteristic set is established, and the dyskinesia characteristics are classified by utilizing a fully-connected neural network design classifier, so that quantitative and hierarchical evaluation of dyskinesia symptoms is realized. The invention can effectively overcome the defects of dyskinesia diseases such as Parkinson and the like and dyskinesia symptoms of cerebral apoplexy caused by manual subjective evaluation, and the provided intelligent quantitative evaluation method is used for clinical evaluation and can effectively improve the objectivity and the accuracy of the evaluation result. The recognition result obtained by using the recognition model provided by the invention is obviously superior to the conventional recognition method.

Drawings

FIG. 1 is a flow chart of an implementation of a modeling method of a quantitative evaluation model for dyskinesia symptoms provided by the invention;

FIG. 2 is a process schematic diagram of a modeling method of a quantitative assessment model for dyskinesia symptoms provided by the invention;

FIG. 3 is a network block diagram of an action recognition network provided by the present invention;

fig. 4 is a schematic diagram of a process for acquiring gait parameters of patients suffering from dyskinesia such as parkinson's disease and cerebral apoplexy according to the present invention.

Detailed Description

In order to make the technical problems solved by the invention, the technical scheme adopted and the technical effects achieved clearer, the invention is further described in detail below with reference to the accompanying drawings and the embodiments. It is to be understood that the specific embodiments described herein are merely illustrative of the invention and are not limiting thereof. It should be further noted that, for convenience of description, only some, but not all of the matters related to the present invention are shown in the accompanying drawings.

1-2, the modeling method of the quantitative evaluation model for dyskinesia symptoms provided by the embodiment of the invention comprises the following processes:

step 100: collecting human body action data, and establishing a human body action recognition model, wherein the human body action recognition model comprises the following steps of 101-103;

step 101: collecting human body action data by using N inertial sensor nodes respectively fixed at different parts of a human body; and acquiring human motion videos by using video acquisition equipment.

Firstly, respectively fixing N inertial sensor nodes on N parts of the upper limb of a subject, wherein each inertial sensor node comprises an accelerometer, a magnetometer, a gyroscope and a surface electromyographic signal sensor; wherein, the inertial sensor can adopt a wearable inertial sensor.

Acquiring the whole body movement disorder test action of the subject by using an inertial sensor node to obtain human body action data; such as forearm swing, bimanual lifting, etc.; the human motion data includes: acceleration data, angular velocity data, magnetometer data, and surface myoelectricity data for each wearable node. Specifically, for the ith wearable node, the acquired human motion data comprise acceleration data of x axis, y axis and z axis and angular velocity data of x axis, y axis and z axis, magnetometer data of x axis, y axis and z axis, and surface myoelectricity data of channel ch1 and channel ch2, and then for the ith wearable node, the original motion data matrix has 11 columns. The acceleration data, the angular velocity data and the magnetometer data belong to inertial information, and the surface myoelectricity data belong to surface myoelectricity information.

And the video acquisition equipment is fixed at a fixed distance in front of the subject, and human body action videos of the subject are acquired.

And uploading the human body action video acquired by the video acquisition equipment and the human body action data acquired by the N wearable sensor nodes to an upper computer data processing platform by utilizing the receiving node.

Step 102: and window segmentation is carried out on the human body action data acquired by each wearable sensor node by utilizing a sliding window segmentation technology, so as to obtain a plurality of action data fragments of each wearable sensor node.

After the human motion data acquired in step 101 is acquired, window segmentation is performed on the human motion data. The embodiment mainly adopts a sliding window segmentation technology to carry out window division on the action data: first, a window size of a fixed length is selected, and then the moving window divides the motion data.

In particular, for any piece of motion sequence data, let the size of the dividing window be L, if the length of the motion data matrix is L _A The motion data matrix may be partitioned intoAnd a plurality of data windows, each window having a split data matrix size of l x 11 dimensions.

Step 103: and forming an action data set by utilizing a plurality of action data fragments of each inertial sensor node, dividing the action data set into an action data training set and an action data test set according to proportion, and randomly selecting a part of the action data training set as an action data verification set.

Step 104: and establishing an action recognition network.

As shown in fig. 3, the action recognition network includes: an FCN module and an LSTM module; the number of input variables is v; the variables are input in parallel, and each time series has an input length of l frames.

The FCN (Fully Convolutional Networks, full convolutional network) module includes K convolutional layers. The filter size of the convolution layer is set to n. A pooling layer is arranged after the nth convolution layer.

The time step of the LSTM (Long Short-Term Memory) module is set to m. An attention mechanism is placed before the LSTM module and a dropout layer is placed after the LSTM to prevent overfitting. Utilizing L in cost function ₂ Regularization prevents model overfitting, and the loss function is an imbalance-like loss function.

wherein w is _i Is v _i Is the sum of the weights of (1), each v _i The weight sigma of (2) _i Defined as sigma _i ＝f(ν _i )vν _i As indicated herein, the product Hadamard was added, and f (. Cndot.) was a fully attached layer.

wherein, when the category is i, y _i =1, otherwise y _i =0; k is the number of categories and p is the output of the model. Beta _i Calculated according to the following formula:

Step 200: the tremor and bradykinesia characteristics of the patient are extracted and pearson correlation analysis is performed, comprising steps 201-204 as follows:

the patient of the present invention comprises: patients with dyskinesia such as Parkinson's disease and cerebral apoplexy.

Step 201: and extracting tremor characteristics of the tremor action fragments of tremor actions.

The tremor feature includes: root mean square, sample entropy, mean, variance, covariance, skewness, kurtosis of the differences in acceleration, angular velocity and surface electromyographic signals. Wherein, the correlation coefficient and the coherence coefficient of the two-channel surface electromyographic signals.

And decomposing the tremor mode from the acceleration and the angular velocity in the inertia information by a complete noise auxiliary aggregation empirical mode decomposition method, and obtaining the amplitude and the frequency of the tremor mode by frequency domain analysis.

The algorithm of the complete noise auxiliary aggregation empirical mode decomposition method comprises the following steps:

(1) Adding Gaussian white noise toTo the signal x (n) to be decomposed to obtain a new signal x (n) +epsilon ₀ ω ⁱ (n) EMD decomposing the new signal until a first new intrinsic mode function is obtained

(2) Calculating a first residual error

(3) Decompose r ₁ (n)+ε ₁ E ₁ ω ⁱ (N), i=1,..n, until the first modality is found, defined as a second intrinsic mode function:

(4) For k=2..k, the minimum residual is calculated for K

(5) Decomposition and implementation r _k (n)+ε _k E _k ω ⁱ (N), i=1,..n, until the first modality is found and defined as

(6) Returning to step (4).

Steps (4) to (6) are looped until a residual is obtained that cannot be decomposed under the established standard.

Step 202: for an action segment of bradykinesia action, extracted first bradykinesia features; and extracting a second motion retardation feature from the corresponding human motion video segment.

The first bradykinesia feature comprises: peak-to-peak value, rectified mean value, variance, standard deviation, skewness, kurtosis, root mean square, peak factor, margin factor, fuzzy entropy of acceleration and angular velocity; mutual information entropy, cross correlation coefficient, KL divergence of sEMG signals, fuzzy entropy and permutation entropy of sEMG; roll angle and pitch angle of the wrist; and (3) synthesizing acceleration, and calculating square root of the sum of the square of the triaxial angular velocities.

The second bradykinesia feature comprises: human body joint point information and joint angles.

And estimating the posture of the RGB human body image in the human body action video acquired by the video acquisition equipment by adopting an OpenPose estimation model, and extracting human body joint information. To protect the privacy of the user, all human action videos will be processed into a skeleton sequence. The kth human body node information is defined as P _k (x, y) obtaining joint angles based on human body joint point information

Wherein P is _k-1 ，P _k ，P _k+1 Is the coordinates of three adjacent human body joints.

Step 300: the gait parameters of the patient are acquired, as shown in fig. 4, comprising the following steps 301-303:

step 301: gait characteristics required for gait analysis are calculated from the acceleration data and the angular velocity data.

And (3) recursively calculating on the basis of the previous state of the system by using an extended Kalman filtering method based on quaternion and taking the minimum mean square error of data as a principle, performing Taylor series expansion on a system state differential equation for a nonlinear system, and reserving derivatives of corresponding orders according to requirements so as to convert the nonlinear system into a linear system for analysis. The system differential equation is:

wherein w is _x 、w _y 、w _x Angular velocity measured for a gyroscope about X, Y, Z axis. The corresponding expression is:

where Δt is the sensor sampling period. After a differential equation of state transition in Kalman filtering is obtained, an observation equation is next obtained, and gravitational acceleration is projected to a navigation coordinate system to obtain the following components:

expression z of the formula corresponding to Kalman filtering _k ＝Hx _k +v _k . Wherein H is the jacobian matrix of acceleration vectors with respect to carrier state quaternions:

after the state transition differential equation and the observation equation are obtained, recursive calculation is carried out according to a Kalman filtering formula. For a discrete linear system, its state can be described by the following equation:

x _k ＝Ax _k-1 +Bu _k +w _k

wherein x is _k-1 For the state of the system at the last moment, A is a state transition matrix, namely a linear relation matrix of the state of the system at the current moment and the state at the last moment, u _k For the control quantity of the system at the moment k, B is a state input matrix, namely the linear relation between the state of the system and the control quantity of the system, and w _k To satisfy the gaussian distributed system noise, its covariance matrix is Q. Meanwhile, for this system, there is the observation equation as follows:

z _k ＝Hx _k +v _k

wherein z is _k X is the observed value of the system k moment _k For the current real state of the system, H is the system observation matrix, i.e. the linear relation between the observed quantity and the system state, v _k To satisfy the gaussian distributed white noise error, the covariance matrix is R. The Kalman filtering process can be divided into two links of state prediction and state correction, wherein the state prediction is to obtain the current state estimation value according to the optimal estimation of the previous system and the state transition differential equation.

x _k|k-1 ＝Ax _k-1|k-1 +Bu _k

Wherein x is _k|k-1 For the current system state predicted value obtained by the system according to the optimal estimation at the last moment, x _k-1|k-1 The optimal estimation is carried out for the last moment. After state prediction, the covariance of the system is updated, according to the multiplication of the covarianceThe product formula can obtain the covariance update formula of the system as follows:

P _k|k-1 ＝AP _k-1|k-1 A ^T +Q

wherein P is _k|k-1 Represents x _k|k-1 Corresponding covariance, P _k-1|k-1 Represents x _k-1|k-1 The corresponding covariance, Q, is the noise covariance matrix of the state transfer equation, and the specific value depends on the accuracy or error dispersion of the sensor in the practical engineering application. The two steps of state prediction are the two steps, the state updating link is followed, and the current Kalman gain kg can be obtained by combining the result of state prediction _k ：

kg _k ＝P _k|k-1 H ^T /(HP _k|k-1 H ^T +R)

Wherein R is the Gaussian noise covariance matrix of the observation equation, and the specific value of the Gaussian noise covariance matrix is the same as Q and also depends on the sensor precision error index in actual engineering application. Obtaining the Kalman gain, namely obtaining the optimal estimation:

x _k|k ＝x _k|k-1 +kg _k (z _k -Hx _k|k-1 )

it should be noted that in an actual system, if the state variable or the observation variable is plural, the Kalman gain kg _k Is an n×m matrix, where N is the number of state variables and M is the number of observation variables. Finally, the optimal estimate x is updated according to equation (4) _k|k To calculate P next recursively _k|k ＝(I-kg _k H)P _k|k-1

Wherein I is an identity matrix. After Kalman filtering fusion, the sensor quaternion gesture x at any sampling time can be obtained _k|k 。

In gait analysis, the heel strike point is typically selected as the start of the gait cycle. The gait cycle can be divided into a stationary phase and a swing phase according to whether the single-sided foot is in contact with the ground or not, and the gait cycle is divided into a support phase from the start of the single-sided heel strike to the lift-off of the single-sided toe, wherein the single-sided foot is always in contact with the ground in the process. The swing phase is from the single-side toe off to the next heel strike. Gait time parameters are calculated from the stationary phase and the swing phase.

Step 302: converting acceleration under a sensor coordinate system into a navigation coordinate system for double integration to obtain a space position under the navigation coordinate system; and corrects the time-space parameter drift.

Specifically, the calculation of the space parameters is mainly completed on a navigation coordinate system, the acceleration under the sensor coordinate system is converted into the navigation coordinate system according to the gesture information to perform double integration, the space position under the navigation coordinate system is obtained, and the space-time parameter drift is corrected based on the assumption of zero speed of the foot in the relative rest stage contacting with the ground.

Extracting each step position according to gait phase division, wherein the X-axis coordinate value of a kth step horizontal coordinate system is p _X (k) The Y-axis coordinate value is p _Y (k) The Z-axis coordinate value is p _Z (k) Then the formula for the stride length is expressed as follows:

clearance refers to the highest ground clearance height of the foot in the swing phase, i.e., the ground clearance height of the foot at mid-swing phase, i.e., the maximum step height. Clearance can reflect the conditions of human lower limb muscle strength, knee joint, hip joint coordination, flexibility and the like.

Height＝max(p _Z (k))

Traversing according to the posture information, and setting joint vectorsThe joint angle information is calculated by the joint vector, and the calculation formula is as follows:

wherein the method comprises the steps ofAnd->Respectively adjacent joint vectors.

Gait time, space and joint angle gait parameters are obtained through the calculation.

Step 400: obtaining a dyskinesia feature set by using the tremor feature obtained in the step 200, the first bradykinesia feature, the second bradykinesia feature and the gait parameters obtained in the step 300, and taking the medical scale scores of neurology specialists as labels; and dividing the dyskinesia feature set into a dyskinesia feature training set and a dyskinesia feature test set according to proportion, and randomly selecting a part of the dyskinesia feature training set as a dyskinesia feature verification set.

In this step, the dyskinesia symptom quantitative evaluation network is a classifier algorithm based on a fully connected neural network.

The application method of the quantitative evaluation model for the dyskinesia symptoms comprises the following steps:

collecting human body motion data and human body motion video of a patient, classifying the human body motion data of the patient by utilizing a motion recognition model, and outputting motion types of motion fragments; and carrying out quantitative evaluation on the movement disorder symptoms of the patient by using a quantitative evaluation model of the movement disorder symptoms.

The invention is further illustrated by way of example:

for example, human hand motion data is acquired by two wearable sensor nodes, each comprising a three-axis accelerometer, a three-axis magnetometer, a three-axis gyroscope and a two-channel surface electromyographic signal sensor, with a sampling frequency of 1024Hz. Human gait data are acquired through seven wearable sensor nodes, and wearing positions are respectively a waist, left and right thighs, left and right shanks and left and right ankles. The subject needs to rest for one minute at the beginning of the experiment for sensor initialization preparation. The motion designed by the experiment is a ten-meter gait test, namely, under the condition that an experimenter is not assisted by people, the ten-meter gait test is completed according to the comfortable speed and the maximum speed of the experimenter. The subjects were 40 total persons, between 41-95 years of age. Two sensor nodes are respectively placed on the right hand wrist and the left hand wrist of the experimental object. In addition, the actions designed for this experiment include: (a) two hands are placed on the chair armrest; (b) double hand lifting; (c) forearm swing; (d) finger tapping; (e) palm movement; (f) a finger nose; (g) a ten meter walk test. And a camera is adopted to collect video motion data in the motion data collection process. In addition, the collected original data is processed in the Python, and a final recognition result is obtained by combining a written recognition algorithm. The present example has collected a total of 560 motion sequences (40 person x 2 sensor x 7 motion) and 40 video bands.

Then, window segmentation is performed on the acquired action sequence. For example, for the ith action sequence, the size of the segmentation window is taken to be 3072, i.e., one data window every 3072 sampling points.

If the length of the motion sequence matrix is L _A The motion data matrix may be partitioned intoAnd a number of data windows, each window having a size of 3072×11 dimension of the divided data matrix. Each feature vector is considered as a data sample, all of which are made into a data set, which is identified and classified.

The present example uses a simple cross-validation method to evaluate the recognition algorithm. Firstly, randomly taking data of 30 test objects as training data for model training, wherein in the training process, the data of 6 test objects are randomly selected by a model without participating in training, and are used as a verification set for verification of the model, and the data of the rest 10 test objects are used as test data. Model parameter setting: the variables are input in parallel, and each time series has an input length of 3072 frames. The filter sizes of the convolution layers in the model are set to 128, 128, 256, 256. The time step of LSTM is set to 32. There is a notice mechanism before LSTM and a dropout layer with a drop rate of 0.5 after LSTM to prevent overfitting. Delta and lambda in the cost function of the model are set to 0.01 and 0.2, respectively.

Table 1 shows the classification accuracy obtained under different recognition methods. Wherein the present example gives the recognition results obtained when the test sample is recognized using different variables; in addition, the method also provides recognition results obtained by other similar methods. Nearest neighbor classifier (KNN), support Vector Machine (SVM). From the results, the method provided by the invention can obtain the highest recognition accuracy.

TABLE 1 accuracy of action classification obtained under different recognition methods

Method

SVM

KNN

MSCNN

Tapnet

MLSTMFCN

The invention is that

Recognition rate

60.16％

62.50％

55.20％

78.13％

91.40％

92.20％

The identified data segment features are extracted, using the right hand dataset as an example, and as previously described, the 25 features extracted for tremor action include: the y-axis acceleration, the y-axis angular velocity, the root mean square of the two-channel sEMG difference, the sample entropy, the mean, the variance, the covariance, the skewness and the kurtosis. Correlation coefficient and coherence coefficient of two-channel sEMG. And decomposing the y-axis angular velocity by using CEEMDAN algorithm to obtain the amplitude and frequency of the tremor component. The 74 features extracted for bradykinesia action include: peak-to-peak value, rectified mean, variance, standard deviation, skewness, kurtosis, root mean square, peak factor, margin factor, fuzzy entropy of 3-axis acceleration and 3-axis angular velocity; mutual information entropy, cross correlation coefficient, KL divergence of two channels of sEMG signals, fuzzy entropy of two channels of sEMG signals and difference value of the two channels of sEMG signals, and permutation entropy; roll angle and pitch angle of the wrist; combining the acceleration, and the square root of the calculated number of the square sum of the triaxial angular velocities; based on the right elbow joint angle of the human joint.

The extracted tremor characteristics and MDS-UPDRS tremor scores are subjected to pearson correlation analysis, and the results are as follows: the magnitude and frequency of the tremor component, the sample entropy of the two-channel sEMG difference, the coherence coefficient of the two-channel sEMG, the root mean square of the y-axis acceleration, the root mean square of the y-axis angular velocity and the sample entropy are significantly correlated with the tremor score (p < 0.01). The extracted bradykinesia features were analyzed for pearson correlation with MDS-UPDRS bradykinesia scores as a result of: 68 of the extracted bradykinesia features were very significantly correlated with the MDS-UPDRS bradykinesia score (p < 0.001), where the root mean square of the y-and z-axis angular velocities, the peak factor of the x-axis acceleration, the fuzzy entropy of the x-and y-axis accelerations, and the fuzzy entropy of the sEMG signal of channel 1 were not statistically significantly different from the MDS-UPDRS bradykinesia score (p > 0.05).

Gait raw data acquired by an inertial sensor are processed in Matlab, and gait time parameters, space parameters and joint parameter characteristics are extracted based on an extended Kalman filtering method of quaternion. The method is characterized in that the inertial data are fused based on acceleration and angular velocity information, and the sensor attitude is calculated through information complementation. And calculating the intensity of the inertia signal at each moment, judging whether the gait cycle is in a stationary phase or not through the inertia signal worn on the foot, and dividing each gait cycle according to the phase information. And converting acceleration signals of the sensors into a navigation coordinate system according to the gesture information, and calculating the position of the foot of the user in space. The displacement of the acceleration in the navigation coordinate system is calculated using double integration, and the position drift is corrected using zero-velocity update. And calculating the posture of each joint according to a traversing method by using a human skeleton model and a forward kinematics method, reconstructing human motions and calculating joint angle changes. And dividing the space-time parameter and the joint parameter information by using the phase information to obtain discrete gait parameter characteristics. The TCN is used for extracting the characteristics of the inertial signals, the TCN is used for realizing the information perception of gait time sequence, and the cavity convolution is used for expanding the receptive field of signals in the Parkinson gait analysis. Gradually calculating inertia signals from local to global according to the multi-layer residual error connection structure, obtaining gait parameters capable of representing the Parkinson disease, and effectively representing gait characteristics of the Parkinson patient by the extracted gait characteristics.

The present example uses a simple cross-validation method to evaluate the evaluation algorithm. Establishing a dyskinesia feature set based on gait features, tremor features and bradykinesia features, classifying patient disease grade into three categories of primary, intermediate and advanced, and classifying the data set according to 3:1 is divided into a training set and a testing set, the classification model provided by the invention is used for quantitative evaluation, and the recognition result obtained by multiple tests on the testing set is more than 90%.

According to the identification method for the motor disorder diseases such as Parkinson and the like and the clinical motor disorder test actions of the cerebral apoplexy patients based on the inertial and myoelectric information, the motor disorder diseases such as Parkinson and the like and the clinical test action identification method for the cerebral apoplexy patients based on bimodal data fusion is provided for the data of the motor disorder diseases such as Parkinson and the like and the cerebral apoplexy patients different from normal people and with higher complexity, an identification model is established based on FCN and LSTM, an attention mechanism is added, and a cost function is improved, so that the capability of capturing important characteristics of the model and solving unbalanced class problems is improved, and algorithm accuracy is improved. The invention also provides a method for extracting the dyskinesia characteristics of patients suffering from parkinsonism and cerebral apoplexy, which can extract quantitative characteristics of gait parameters, tremors and bradykinesia characteristics, and verify the effectiveness of the extracted characteristics by adopting a pearson correlation analysis method. In addition, the quantitative and graded evaluation method for the dyskinesia diseases such as Parkinson and the like and the cerebral apoplexy dyskinesia symptoms can objectively and quantitatively evaluate the dyskinesia symptoms of patients in a graded manner, and effectively overcomes the defects of manual evaluation.

Finally, it should be noted that: the above embodiments are only for illustrating the technical solution of the present invention, and not for limiting the same; although the invention has been described in detail with reference to the foregoing embodiments, it will be understood by those of ordinary skill in the art that: the technical scheme described in the foregoing embodiments is modified or some or all of the technical features are replaced equivalently, so that the essence of the corresponding technical scheme does not deviate from the scope of the technical scheme of the embodiments of the present invention.

Claims

1. A modeling method of a quantitative evaluation model for movement disorder symptoms, which is characterized by comprising the following steps:

step 300: acquiring gait parameters of a patient;

2. The modeling method of a quantitative assessment model for movement disorder symptoms according to claim 1, wherein step 100 includes steps 101 to 103 as follows:

step 104: establishing an action recognition network;

3. A method of modeling a quantitative assessment model of movement disorder symptoms according to claim 2, wherein the action recognition network comprises: an FCN module and an LSTM module; the number of input variables is v; the variable is used as input in parallel, and the input length of each time sequence is l frames;

the FCN module comprises K convolution layers;

wherein w is _i Is v _i Is the sum of the weights of (1), each v _i The weight sigma of (2) _i Defined as sigma _i ＝f(ν _i )vν _i ', as indicated by Hadamard product, f (. Cndot.) is a fully attached layer;

wherein, when the category is i, y _i =1, otherwise y _i =0; k is the number of categories, p is the output of the model;β _i calculated according to the following formula:

4. The method of modeling a quantitative assessment model for a symptom of dyskinesia according to claim 1, wherein step 200 includes steps 201-204 as follows:

5. The modeling method of a quantitative assessment model for movement disorder symptoms according to claim 1, wherein step 300 includes steps 301 to 303 as follows: