A modeling method of a quantitative evaluation model of dyskinesia symptoms

Abstract

The application 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 action data, and establishing a human action recognition model; step 200: extracting tremor characteristics and movement retardation characteristics of a patient, and performing Pearson correlation analysis; step 300: obtaining gait parameters of the patient; step 400: obtaining a movement disorder characteristic set; step 500: based on a full-connection neural network, a movement disorder symptom quantitative evaluation network is established; the movement disorder symptom quantitative evaluation network is trained by using a movement disorder characteristic training set, the movement disorder symptom quantitative evaluation network is tested by using a movement disorder characteristic test set, and finally, a movement disorder symptom quantitative evaluation model is obtained. The quantitative evaluation model can be used for grading evaluation of the movement disorder symptoms of the patient.

Description

Technical Field

[0001] This invention relates to the field of movement disorder symptom assessment technology, and in particular to a modeling method for a quantitative assessment model of movement disorder symptoms. Background Technology

[0002] Motor disorders such as Parkinson's disease and stroke clinically present with motor symptoms (bradykinesia, resting tremor, rigidity, postural instability, etc.). Patients in the early stages of Parkinson's disease and other motor disorders initially experience motor symptoms, while in the middle and later stages of stroke, hemiplegia, and other diseases, motor dysfunction becomes the primary manifestation. The assessment of motor symptoms is of great significance for the diagnosis of Parkinson's disease and other motor disorders, as well as for the rehabilitation assessment of diseases such as stroke.

[0003] Currently, clinical assessment of movement disorder symptoms largely relies on medical scales. For example, Parkinson's disease, the most common movement disorder, is assessed using the Movement Disorder Society-Unified Parkinson's Disease Rating Scale (MDS-UPDRS). This assessment method is widely used clinically, but it has certain drawbacks: it is time-consuming and its analysis is somewhat subjective, leading to discrepancies in assessment results among different physicians. Furthermore, these differing assessment results can impact treatment and rehabilitation, making the evaluation of treatment and rehabilitation outcomes more challenging.

[0004] In the clinical assessment of movement disorder symptoms, it is necessary to develop an intelligent and objective quantitative analysis method for movement disorder symptoms. Currently, research fields such as biomedical engineering are attempting to quantify movement disorder symptoms using inertial sensors, electromyography (EMG) sensors, or video. However, in clinical practice, patients' disease progression varies, and a single sensor approach cannot solve all practical problems. Under real-world conditions, patients' movements are more spontaneous and disordered than those of healthy individuals, increasing the difficulty of movement recognition and quantitative feature extraction. Many researchers, in their quantitative studies of movement disorder symptoms, use machine learning or deep learning to classify and identify inertial or video data from healthy individuals and patients with movement disorders such as Parkinson's disease. While this effectively distinguishes between healthy individuals and those with Parkinson's, it neglects the quantitative grading assessment of movement disorder symptoms. Furthermore, the information obtained from a single sensor is incomplete, leading to inaccurate assessment results. Existing quantitative analysis techniques for movement disorder symptoms based on multimodal information are cumbersome, complex to operate, and mostly require a laboratory environment, making them unsuitable for clinical applications. Additionally, the extracted features are not comprehensive enough, resulting in low precision in quantification and grading.

[0005] This invention combines multi-source information to greatly improve the recognition performance of clinical test movements for movement disorder symptoms. The multi-dimensional and efficient movement disorder features extracted using multimodal information are of great value for high-precision quantitative and graded assessment of movement disorder symptoms. Convenient, intelligent, and objective disease quantitative analysis is of great significance for the treatment and rehabilitation assessment of diseases. Summary of the Invention

[0006] This invention primarily addresses the technical problem of existing technologies' inability to comprehensively and objectively quantify the symptoms of movement disorders such as Parkinson's disease and stroke. It proposes a modeling method for a quantitative assessment model of movement disorder symptoms, which can effectively overcome the drawbacks of subjective human analysis of patients' clinical symptoms. The feature extraction method proposed in this invention extracts various movement disorder features that are significantly correlated with the scale scores of neurologists. Furthermore, the quantitative assessment model proposed in this invention can accurately grade and assess the patients' movement disorder symptoms.

[0007] This invention provides a modeling method for a quantitative assessment model of movement disorder symptoms, comprising the following steps:

[0008] Step 100: Collect human motion data and establish a human motion recognition model;

[0009] Step 200: Extract the patient's tremor and bradykinesia characteristics and perform Pearson correlation analysis;

[0010] Step 300: Obtain the patient's gait parameters;

[0011] Step 400: Using the tremor features, first bradykinesia features, second bradykinesia features obtained in Step 200 and the gait parameters obtained in Step 300, obtain a movement disorder feature set, and use the medical scale scores of neurology experts as labels; and randomly divide the movement disorder feature set into a movement disorder feature training set and a movement disorder feature test set according to the proportion, and randomly select a part of the movement disorder feature training set as the movement disorder feature validation set.

[0012] Step 500: Based on a fully connected neural network, establish a quantitative assessment network for movement disorder symptoms; train the quantitative assessment network for movement disorder symptoms using a movement disorder feature training set, and test the network using a movement disorder feature test set to finally obtain a quantitative assessment model for movement disorder symptoms.

[0013] Furthermore, step 100 includes the following steps 101-103:

[0014] Step 101: Collect human motion data using N inertial sensor nodes fixed to different parts of the human body; collect human motion video using video acquisition equipment;

[0015] Step 102: Using the sliding window segmentation technique, the human motion data collected by each wearable sensor node is segmented into windows to obtain multiple motion data segments for each wearable sensor node.

[0016] Step 103: Using multiple motion data segments from each inertial sensor node, form a motion dataset. Randomly divide the motion dataset into a motion data training set and a motion data test set according to the proportion. Randomly select a portion of the motion data training set as the motion data validation set.

[0017] Step 104: Establish an action recognition network;

[0018] Step 105: Train the action recognition network using the action data training set to obtain the action recognition model; validate the action recognition model using the action data validation set; test the action recognition model using the action recognition model test set to obtain the best-performing human action recognition model; wherein, the action categories recognized by the human action recognition model include: tremor, bradykinesia, or gait.

[0019] Furthermore, the action recognition network includes: an FCN module and an LSTM module; the number of input variables is v; the variables are used as input in parallel, and the input length of each time series is l frames;

[0020] The FCN module includes K convolutional layers;

[0021] The time step of the LSTM module is set to m; an attention mechanism is set before the LSTM module and a dropout layer is set after the LSTM;

[0022] The attention mechanism in the LSTM module is shown in the following equation, which focuses the attention of the LSTM model on variables in different dimensions:

[0023]

[0024] Among them, w i For ν i The sum of weights for each ν i weight σ i Defined as σ i =f(ν i )⊙ν i ', ⊙ represents the Hadamard product, and f(·) represents a fully connected layer;

[0025] The cost function consists of two parts: L2 regularization and an imbalance-like loss function.

[0026]

[0027] Where, when category i, y i =1, otherwise y i =0; k is the number of categories, p is the model output; β i Calculate according to the following formula:

[0028]

[0029] Where, n i δ represents the number of samples in class j, and is a hyperparameter.

[0030] Furthermore, step 200 includes the following steps 201-204:

[0031] Step 201: For the tremor movement segment, extract its tremor features; the tremor features include: root mean square of the difference between acceleration, angular velocity and surface electromyography signal, sample entropy, mean, variance, covariance, skewness and kurtosis;

[0032] Step 202: For motion segments with slowed motion, extract the first motion slowness feature; and extract the second motion slowness feature for the corresponding human motion video segment;

[0033] The first motion sluggishness features include: peak-to-peak value, rectified average value, variance, standard deviation, skewness, kurtosis, root mean square, peak factor, margin factor, and fuzzy entropy of acceleration and angular velocity; mutual information entropy, cross-correlation coefficient, KL divergence of sEMG signals, and fuzzy entropy and permutation entropy of sEMG; roll and pitch angles of the wrist; resultant acceleration and the arithmetic square root of the sum of squares of the three-axis angular velocities;

[0034] The second characteristic of bradykinesia includes: information on human joint points and joint angles;

[0035] Step 203: Perform Pearson correlation analysis on the tremor features obtained in Step 201, the first bradykinesia feature and the second bradykinesia feature obtained in Step 202, and the medical scale scores assessed by neurology experts to obtain the Pearson correlation coefficient, and determine the validity of the proposed features based on the Pearson correlation coefficient.

[0036] Furthermore, step 300 includes the following steps 301-303:

[0037] Step 301: Calculate the gait characteristics required for gait analysis using acceleration and angular velocity data;

[0038] Step 302: Transform the acceleration in the sensor coordinate system to the navigation coordinate system and perform double integration to obtain the spatial position in the navigation coordinate system; and correct for the spatiotemporal parameter drift.

[0039] Step 303: Use TCN to extract features of inertial signals, realize the perception of gait time series information through TCN, and use dilated convolution to expand the receptive field of signals in Parkinson's gait analysis; according to the multi-layer residual connection structure, realize the stepwise calculation of inertial signals from local to global, and obtain gait parameters that can characterize Parkinson's disease and other movement disorders and stroke.

[0040] This invention provides a modeling method for a quantitative assessment model of movement disorder symptoms. At the action recognition level, it primarily utilizes FCN, LSTM, attention mechanisms, and a class imbalance loss function to construct the quantitative assessment model. This model effectively captures important variables in the input variables and enhances the model's ability to learn class imbalance problems, thereby improving algorithm performance. At the feature extraction level, it uses the CEEMDAN algorithm, combined with parameters set based on the characteristics of tremor, to extract the frequency and amplitude of acceleration signals in tremor action segments. Based on electromyographic, inertial, and video information, it extracts bradykinesic features of the bradykinesic test action, extracts a second bradykinesic feature based on the video, and uses inertial information based on the quaternion-based extended Kalman filter and TCN to extract gait time parameters, spatial parameters, and joint parameter features. At the quantitative assessment level, it establishes a movement disorder feature set and uses a fully connected neural network to design a classifier to classify the movement disorder features, achieving a quantitative and graded assessment of movement disorder symptoms. This invention effectively overcomes the drawbacks of subjective human assessment of movement disorder symptoms in Parkinson's disease and stroke. The proposed intelligent quantitative assessment method can effectively improve the objectivity and accuracy of assessment results in clinical evaluation. The recognition results obtained using the recognition model proposed in this invention are significantly better than those obtained using traditional recognition methods. Attached Figure Description

[0041] Figure 1 This is a flowchart illustrating the implementation of the modeling method for the quantitative assessment model of movement disorder symptoms provided by this invention.

[0042] Figure 2 This is a schematic diagram illustrating the modeling process of the quantitative assessment model for movement disorder symptoms provided by the present invention;

[0043] Figure 3 This is a network structure diagram of the action recognition network provided by the present invention;

[0044] Figure 4 This is a schematic diagram of the process for obtaining gait parameters of patients with Parkinson's disease and stroke, as provided by the present invention. Detailed Implementation

[0045] To make the technical problems solved by this invention, the technical solutions adopted, and the technical effects achieved clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and not intended to limit it. Furthermore, it should be noted that, for ease of description, only the parts relevant to the invention are shown in the accompanying drawings, not all of them.

[0046] like Figure 1-2 As shown in the figure, the modeling method for a quantitative assessment model of movement disorder symptoms provided by an embodiment of the present invention includes the following process:

[0047] Step 100: Collect human motion data and establish a human motion recognition model, including the following steps 101-103;

[0048] Step 101: Collect human motion data using N inertial sensor nodes fixed to different parts of the human body; collect human motion video using video acquisition equipment.

[0049] First, N inertial sensor nodes are fixed to N locations on the subject's upper limb. Each inertial sensor node includes an accelerometer, a magnetometer, a gyroscope, and a surface electromyography (EMG) sensor. Wearable inertial sensors can be used.

[0050] Inertial sensor nodes are used to collect the subject's whole-body motion impairment test movements to obtain human motion data; for example, forearm rotation, and arms extended horizontally. The human motion data includes: acceleration data, angular velocity data, magnetometer data, and surface electromyography (EMG) data for each wearable node. Specifically, for the i-th wearable node, the collected human motion data includes acceleration data and angular velocity data along the x, y, and z axes, magnetometer data along the x, y, and z axes, and EMG data from channels ch1 and ch2. Therefore, the original motion data matrix for the i-th wearable node has 11 columns. Among these, acceleration data, angular velocity data, and magnetometer data belong to inertial information, while surface EMG data belongs to surface EMG information.

[0051] The video capture device is fixed at a fixed distance directly in front of the subject to capture video of the subject's human movements.

[0052] The receiving node uploads the human motion video captured by the video acquisition device and the human motion data collected by N wearable sensor nodes to the host computer data processing platform.

[0053] Step 102: Using the sliding window segmentation technique, the human motion data collected by each wearable sensor node is segmented into windows to obtain multiple motion data segments for each wearable sensor node.

[0054] After acquiring the human motion data collected in step 101, the human motion data is segmented into windows. This embodiment mainly uses the sliding window segmentation technique to divide the motion data into windows: first, a fixed-length window size is selected, and then the window is moved to segment the motion data.

[0055] Specifically, for any action sequence data, let the size of the segmentation window be l, and the length of the motion data matrix be L. A Then the motion data matrix can be divided into There are 1 data window, and the size of the segmented data matrix in each window is l×11 dimensional.

[0056] Step 103: Using multiple motion data segments from each inertial sensor node, form a motion dataset. Randomly divide the motion dataset into a motion data training set and a motion data test set according to the proportion. Randomly select a portion of the motion data training set as the motion data validation set.

[0057] Step 104: Establish an action recognition network.

[0058] like Figure 3 As shown, the action recognition network includes: an FCN module and an LSTM module; the number of input variables is v; the variables are used as input in parallel, and the input length of each time series is l frames.

[0059] The FCN (Fully Convolutional Networks) module consists of K convolutional layers. The filter size of each convolutional layer is set to n. A pooling layer is placed after the Nth convolutional layer.

[0060] The time step of the LSTM (Long Short-Term Memory) module is set to m. An attention mechanism is set before the LSTM module, and a dropout layer is set after the LSTM to prevent overfitting. L2 regularization is used in the cost function to prevent model overfitting, and the loss function is the class-imbalanced loss function.

[0061] The attention mechanism in the LSTM module is shown in the following equation, which focuses the attention of the LSTM model on variables in different dimensions:

[0062]

[0063] Among them, w i For νi The sum of weights for each ν i weight σ i Defined as σ i =f(ν i )vν i ', ⊙ represents the Hadamard product, and f(·) is a fully connected layer.

[0064] The cost function consists of two parts: L2 regularization and an imbalance-like loss function.

[0065]

[0066] Where, when category i, y i =1, otherwise y i =0; k is the number of categories, p is the model output. β i Calculate according to the following formula:

[0067]

[0068] Where, n i δ represents the number of samples in class j, and is a hyperparameter.

[0069] Step 105: Train the action recognition network using the action data training set to obtain the action recognition model; validate the action recognition model using the action data validation set; test the action recognition model using the action recognition model test set to obtain the best-performing human action recognition model; wherein, the action categories recognized by the human action recognition model include: tremor, bradykinesia, or gait.

[0070] Step 200: Extract the patient's tremor and bradykinesia characteristics and perform Pearson correlation analysis, including the following steps 201-204:

[0071] The patients described in this invention include patients with movement disorders such as Parkinson's disease and stroke patients who have symptoms of movement disorders.

[0072] Step 201: Extract the tremor features from the action segments of the tremor action.

[0073] The tremor characteristics include: root mean square, sample entropy, mean, variance, covariance, skewness, and kurtosis of the differences between acceleration, angular velocity, and surface electromyography (EMG) signals. Among these, the correlation coefficient and coherence coefficient of the dual-channel EMG signals are also considered.

[0074] By using a fully noise-assisted aggregated empirical mode decomposition method, flutter modes are decomposed from the acceleration and angular velocity in inertial information, and the amplitude and frequency of the flutter modes are obtained through frequency domain analysis.

[0075] The algorithm steps of the fully noise-assisted aggregated empirical mode decomposition method are as follows:

[0076] (1) Adding Gaussian white noise to the signal to be decomposed x(n) yields a new signal x(n) + ε0ω i (n), perform EMD decomposition on the new signal until the first new intrinsic mode function is obtained.

[0077] (2) Calculate the first residual

[0078] (3) Decompose r1(n)+ε1E1ω i (n), i = 1, ..., N, until the first mode is found and defined as the second intrinsic mode function:

[0079] (4) For K = 2, ..., K, calculate the minimum residual.

[0080] (5) Decompose and implement r k (n)+ε k E k ω i (n), i = 1, ..., N until the first mode is found and defined as

[0081] (6) Return to step (4).

[0082] Repeat steps (4) to (6) until a residual that cannot be decomposed under the given criteria is obtained.

[0083] Step 202: For motion segments with slowed motion, extract the first motion slowed feature; and extract the second motion slowed feature for the corresponding human motion video segment.

[0084] The first motion sluggishness features include: peak-to-peak value, rectified average value, variance, standard deviation, skewness, kurtosis, root mean square, peak factor, margin factor, and fuzzy entropy of acceleration and angular velocity; mutual information entropy, cross-correlation coefficient, KL divergence of sEMG signals, and fuzzy entropy and permutation entropy of sEMG; roll and pitch angles of the wrist; resultant acceleration, and the arithmetic square root of the sum of squares of the three-axis angular velocities.

[0085] The second characteristic of bradykinesia includes: information on human joint points and joint angles.

[0086] The OpenPose estimation model is used to estimate the pose of RGB human images in human motion videos captured by video acquisition devices, and to extract human joint information. To protect user privacy, all human motion videos will be processed into skeleton sequences. The k-th human joint information is defined as P. k (x,y), joint angles obtained based on human joint information.

[0087]

[0088] In the formula, P k-1 P k P k+1 These are the coordinates of three adjacent human body joints.

[0089] Step 203: Perform Pearson correlation analysis on the tremor features obtained in Step 201, the first bradykinesia feature and the second bradykinesia feature obtained in Step 202, and the medical scale scores assessed by neurology experts to obtain the Pearson correlation coefficient, and determine the validity of the proposed features based on the Pearson correlation coefficient.

[0090] Step 300: Obtain the patient's gait parameters, such as... Figure 4 As shown, it includes the following steps 301-303:

[0091] Step 301: Calculate the gait characteristics required for gait analysis using acceleration and angular velocity data.

[0092] Using a quaternion-based extended Kalman filter method, and adhering to the principle of minimizing the mean square error of the data, the algorithm recursively calculates the system state based on the previous state. For nonlinear systems, a Taylor series expansion of the system state differential equation is performed, retaining derivatives of the corresponding orders as needed to transform the nonlinear system into a linear system for analysis. The system differential equation is:

[0093]

[0094] Where w x w y w x This refers to the angular velocity of rotation about the X, Y, and Z axes, measured by a gyroscope. The corresponding expressions are:

[0095]

[0096] Where Δt is the sensor sampling period. After obtaining the differential equation for the state transition in the Kalman filter, the observation equation is then derived by projecting the gravitational acceleration onto the navigation coordinate system:

[0097]

[0098] The formula corresponds to the Kalman filter expression z k =Hx k +v k Where H is the Jacobian matrix of the acceleration vector with respect to the quaternion of the carrier state:

[0099]

[0100] After obtaining the state transition differential equation and the observation equation, recursive calculations are performed according to the Kalman filter formula. For a discrete linear system, its state can be described by the following equation:

[0101]

[0102] Where Δt is the sensor sampling period. After obtaining the differential equation for the state transition in the Kalman filter, the observation equation is then derived by projecting the gravitational acceleration onto the navigation coordinate system:

[0103]

[0104] The formula corresponds to the Kalman filter expression z k =Hx k +v k Where H is the Jacobian matrix of the acceleration vector with respect to the quaternion of the carrier state:

[0105]

[0106] After obtaining the state transition differential equation and the observation equation, recursive calculations are performed according to the Kalman filter formula. For a discrete linear system, its state can be described by the following equation:

[0107] x k =Ax k-1 +Bu k +w k

[0108] Where x k-1 Let A be the state of the system at the previous time step, and let A be the state transition matrix, which is the linear relationship matrix between the current state and the previous state. k Let w be the control input to the system at time k, B be the state input matrix, i.e., the linear relationship between the system state and the system control input, and w be the control input. k For system noise to satisfy a Gaussian distribution, its covariance matrix is ​​Q. Meanwhile, for this system, the observation equation is as follows:

[0109] z k =Hx k +v k

[0110] Among them, zk Let x be the observed value of the system at time k. k Let H be the current true state of the system, and let H be the system observation matrix, i.e., the linear relationship between the observations and the system state. k To satisfy the Gaussian distribution of white noise error, its covariance matrix is ​​R. The Kalman filtering process can be divided into two stages: state prediction and state correction. State prediction is to obtain the current state estimate based on the previous optimal estimate of the system and the state transition differential equation.

[0111] x k|k-1 =Ax k-1|k-1 +Bu k

[0112] Where x k|k-1 Let x be the predicted value of the current system state obtained by the system based on the optimal estimate at the previous moment. k-1|k-1 This is the optimal estimate from the previous time step. After state prediction, the system's covariance is also updated. Based on the product formula for covariance, the system's covariance update formula is:

[0113] P k|k-1 =AP k-1|k-1 A T +Q

[0114] Among them, P k|k-1 Represents x k|k-1 The 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 transition equation. Its specific value depends on the sensor's accuracy or error distribution in practical engineering applications. These are the two steps of state prediction. The next step is the state update stage. Combining the state prediction results, the current Kalman gain kg can be obtained. k :

[0115] kg k =P k|k-1 H T / (HP k|k-1 H T +R)

[0116] Where R is the Gaussian noise covariance matrix of the observation equation, and its specific value, like Q, depends on the sensor accuracy error index in practical engineering applications. By obtaining the Kalman gain, the optimal estimate can be obtained.

[0117] x k|k =x k|k-1 +kg k (z k -Hx k|k-1 )

[0118] It is important to note that in practical systems, if there are multiple state variables or observation variables, the Kalman gain (kg) will be higher. k It 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 needs to be updated according to formula (4). k|k The covariance is used to calculate P in the next recursive calculation. k|k = (I-kg) k H)P k|k-1

[0119] Where I is the identity matrix. After fusion using Kalman filtering, the sensor quaternion attitude x at any sampling time can be obtained. k|k .

[0120] In gait analysis, the heel strike is typically chosen as the starting point of the gait cycle. Based on whether one foot is in contact with the ground, the gait cycle can be divided into a stationary phase and a swing phase. The stance phase begins with the heel striking the ground and continues until the toes leave the ground, during which the foot remains in contact with the ground. The swing phase begins with the toes leaving the ground and continues until the next heel strike. Gait time parameters are calculated based on the stationary and swing phases.

[0121] Step 302: Transform the acceleration in the sensor coordinate system to the navigation coordinate system and perform double integration to obtain the spatial position in the navigation coordinate system; and correct the spatiotemporal parameter drift.

[0122] Specifically, the calculation of spatial parameters is mainly completed in the navigation coordinate system. The acceleration in the sensor coordinate system is transformed into the navigation coordinate system based on the attitude information and double integrated to obtain the spatial position in the navigation coordinate system. Based on the assumption of zero velocity of the foot during the relatively stationary phase in contact with the ground, the spatiotemporal parameter drift is corrected.

[0123] The position of each step is extracted based on gait phase division, where the X-axis coordinate of the horizontal coordinate system at the k-th step is p. X (k), the Y-axis coordinate value is p Y (k), the Z-axis coordinate value is p Z If (k) is a given number, then the formula for calculating the step size is as follows:

[0124]

[0125] Clearance refers to the highest point the foot leaves the ground during the swing phase, specifically the height the foot is off the ground at mid-swing, i.e., the maximum stride height. Clearance reflects a person's lower limb muscle strength, knee and hip joint coordination, and flexibility.

[0126] Height = max(p Z (k))

[0127] The process is traversed based on each pose information, and joint vectors are defined. Joint angle information is calculated using joint vectors, and the calculation formula is as follows:

[0128]

[0129] in and These are the adjacent joint vectors.

[0130] Step 303: Use TCN to extract features of inertial signals, realize the perception of gait time series information through TCN, and use dilated convolution to expand the receptive field of signals in Parkinson's gait analysis; according to the multi-layer residual connection structure, realize the stepwise calculation of inertial signals from local to global, and obtain gait parameters that can characterize Parkinson's disease and other movement disorders and stroke.

[0131] The above calculations yield gait parameters including gait time, space, and joint angles.

[0132] Step 400: Using the tremor features, first bradykinesia features, second bradykinesia features obtained in step 200 and the gait parameters obtained in step 300, obtain a movement disorder feature set, and use the medical scale scores of neurology experts as labels; and randomly divide the movement disorder feature set into a movement disorder feature training set and a movement disorder feature test set according to the proportion, and randomly select a part of the movement disorder feature training set as the movement disorder feature validation set.

[0133] Step 500: Based on a fully connected neural network, establish a quantitative assessment network for movement disorder symptoms; train the quantitative assessment network for movement disorder symptoms using a movement disorder feature training set, and test the network using a movement disorder feature test set to finally obtain a quantitative assessment model for movement disorder symptoms.

[0134] In this step, the movement disorder symptom quantification assessment network is a classifier algorithm based on a fully connected neural network.

[0135] Method of using the quantitative assessment model for movement disorder symptoms obtained in this invention:

[0136] The system collects patient motion data and videos, and uses a motion recognition model to classify the patient's motion data and output the motion category of the motion segment. It also uses a motion disorder symptom quantification assessment model to quantify the patient's motion disorder symptoms.

[0137] The invention will be further explained and illustrated below with examples:

[0138] For example, human hand movement data was collected using two wearable sensor nodes. Each sensor node included a three-axis accelerometer, a three-axis magnetometer, a three-axis gyroscope, and a dual-channel surface electromyography (EMG) sensor, with a sampling frequency of 1024 Hz. Human gait data was collected using seven wearable sensor nodes, worn on the waist, left and right thighs, left and right calves, and left and right ankles. Subjects were required to remain still for one minute at the start of the experiment for sensor initialization. The designed movement was a 10-meter gait test, where participants walked 10 meters at their comfortable and maximum speed without assistance. A total of 40 subjects, aged 41-95, participated. The two sensor nodes were placed on the right and left wrists of the subjects, respectively. The designed movements included: (a) placing both hands on chair armrests; (b) raising both hands horizontally; (c) forearm rotation; (d) finger tapping; (e) palm movement; (f) finger-to-nose; and (g) the 10-meter walk test. Video motion data was collected using a camera during the movement data acquisition process. In addition, the collected raw data will be processed in Python and combined with the written recognition algorithm to obtain the final recognition result. This example collected a total of 560 motion sequences (40 people × 2 sensors × 7 actions) and 40 video frequency bands.

[0139] Then, the collected action sequences are segmented into windows. For example, for the i-th action sequence, the size of the segmentation window is 3072, that is, every 3072 sampling points constitute a data window.

[0140] If the length of the action sequence matrix is ​​L A Then the motion data matrix can be divided into There are 3072×11-dimensional segmented data matrices within each data window. Each feature vector is considered a data sample, and all samples are compiled into a dataset for identification and classification.

[0141] This example primarily utilizes simple cross-validation to evaluate the recognition algorithm. First, data from 30 randomly selected subjects are used as training data for model training. During training, the model randomly selects 6 subjects' data (excluding training data) as the validation set for model validation. The remaining 10 subjects' data serve as test data. Model parameter settings: variables are used in parallel as input, with each time series input length being 3072 frames. The filter sizes of the convolutional layers in the model are set to 128, 128, 256, and 256. The LSTM time step size is set to 32. An attention mechanism is implemented before the LSTM, followed by a dropout layer with a decrease rate of 0.5 to prevent overfitting. The cost function δ and λ are set to 0.01 and 0.2, respectively.

[0142] Table 1 shows the classification accuracy obtained under different recognition methods. This example presents the recognition results obtained when using different variables to identify test samples; additionally, it also presents the recognition results obtained by other similar methods, including Nearest Neighbor Classifier (KNN) and Support Vector Machine (SVM). The results show that the method proposed in this invention achieves the highest recognition accuracy.

[0143] Table 1. Action classification accuracy obtained under different recognition methods

[0144] method SVM KNN MSCNN Tapnet MLSTMFCN This invention Recognition rate 60.16％ 62.50％ 55.20％ 78.13％ 91.40％ 92.20％

[0145] For feature extraction of identified data segments, taking the right-hand dataset as an example, as mentioned earlier, the 25 features extracted for tremor include: y-axis acceleration, y-axis angular velocity, and the root mean square, sample entropy, mean, variance, covariance, skewness, and kurtosis of the difference between the two-channel sEMG. The correlation coefficient and coherence coefficient of the two-channel sEMG are also included. The CEEMDAN algorithm is used to decompose the y-axis angular velocity to obtain the amplitude and frequency of the tremor component. For slow-motion extraction, the 74 features include: peak-to-peak value, rectified mean, variance, standard deviation, skewness, kurtosis, root mean square, peak factor, margin factor, and fuzzy entropy of the three-axis acceleration and angular velocity; mutual information entropy, cross-correlation coefficient, KL divergence of the two-channel sEMG signals, and the fuzzy entropy and permutation entropy of the two-channel sEMG and their difference; the roll and pitch angles of the wrist; the resultant acceleration and the arithmetic square root of the sum of squares of the three-axis angular velocities; and the right elbow joint angle based on human joints.

[0146] Pearson correlation analysis was performed on the extracted tremor features and the MDS-UPDRS tremor score. The results showed that the amplitude and frequency of the tremor components, the sample entropy of the difference between the two-channel sEMG, 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 were significantly correlated with the tremor score (p<0.01). Pearson correlation analysis was also performed on the extracted bradykinesia features and the MDS-UPDRS bradykinesia score. The results showed that 68 of the extracted bradykinesia features were highly significantly correlated with the MDS-UPDRS bradykinesia score (p<0.001). Among these, the root mean square of the y-axis and z-axis angular velocities, the peak factor of the x-axis acceleration, the fuzzy entropy of the x-axis and y-axis acceleration, and the fuzzy entropy of the sEMG signal from channel 1 did not show statistically significant differences with the MDS-UPDRS bradykinesia score (p>0.05).

[0147] This paper processes raw gait data acquired by inertial sensors in Matlab, extracting gait temporal, spatial, and joint parameter features based on the quaternion-based extended Kalman filter method. First, inertial data is fused based on acceleration and angular velocity information, and sensor attitude is calculated through information complementarity. The intensity of the inertial signal at each moment is calculated, and the inertial signal worn on the foot is used to determine whether the user is in a stationary phase. Each gait cycle is divided based on phase information. The sensor's acceleration signal is converted to the navigation coordinate system based on attitude information to calculate the user's foot position in space. Double integral is used to calculate the displacement of acceleration in the navigation coordinate system, and zero-velocity update is used to correct position drift. Using a human skeletal model and forward kinematics methods, the attitude of each joint is calculated by ergonomic method, reconstructing human movement and calculating joint angle changes. Phase information is used to segment spatiotemporal and joint parameter information to obtain discrete gait parameter features. TCN is used to extract inertial signal features, enabling perception of gait time-series information. Dilated convolution is used to expand the receptive field of signals in Parkinson's gait analysis. The inertial signal is calculated stepwise from local to global based on the multi-layer residual connection structure, and gait parameters that can characterize Parkinson's disease are obtained. The extracted gait features can effectively characterize the gait characteristics of Parkinson's patients.

[0148] This example primarily utilizes simple cross-validation to evaluate the assessment algorithm. A movement disorder feature set is established based on gait, tremor, and bradykinesia characteristics. Patient disease levels are categorized into three levels: primary, intermediate, and advanced. The dataset is divided into training and test sets in a 3:1 ratio. The proposed classification model is used for quantitative evaluation, and multiple trials on the test set yielded a recognition rate exceeding 90%.

[0149] This example provides a method for recognizing clinical movement disorder test movements in Parkinson's disease and stroke patients based on inertial and electromyographic information. Addressing the complexity of data from Parkinson's disease and stroke patients, which differs significantly from normal individuals, this method proposes a bimodal data fusion-based approach for recognizing clinical test movements in these patients. A recognition model is built based on FCN and LSTM, with the addition of an attention mechanism and an improved cost function, thereby enhancing the model's ability to capture important features and resolve class imbalance problems, thus improving algorithm accuracy. This invention also provides a method for extracting movement disorder features from Parkinson's disease and stroke patients. This method can quantitatively extract features such as gait parameters, tremor, and bradykinesia, and the effectiveness of the extracted features is verified using Pearson correlation analysis. Furthermore, the various movement disorder features extracted by the proposed method show significant correlations with neurological expert scale scores. In addition, the quantitative grading assessment method for movement disorder symptoms in Parkinson's disease and stroke patients provided by this invention can objectively quantify and grade patient movement disorder symptoms, effectively overcoming the drawbacks of manual assessment.

[0150] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications to the technical solutions described in the foregoing embodiments, or equivalent substitutions for some or all of the technical features, do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A modeling method for a quantitative assessment model of movement disorder symptoms, characterized in that, Includes the following processes: Step 100: Collect human motion data and establish a human motion recognition model; Step 100 includes the following steps 101-103: Step 101: Collect human motion data using N inertial sensor nodes fixed to different parts of the human body; collect human motion video using video acquisition equipment; Step 102: Using the sliding window segmentation technique, the human motion data collected by each wearable sensor node is segmented into windows to obtain multiple motion data segments for each wearable sensor node. Step 103: Using multiple motion data segments from each inertial sensor node, form a motion dataset. Randomly divide the motion dataset into a motion data training set and a motion data test set according to the proportion. Randomly select a portion of the motion data training set as the motion data validation set. Step 104: Establish an action recognition network; the action recognition network includes: an FCN module and an LSTM module; the number of input variables is v; variables are used as input in parallel, and the input length of each time series is v. l frame; The FCN module includes K convolutional layers; The time step of the LSTM module is set to m Set up an attention mechanism before the LSTM module and a dropout layer after the LSTM module; The attention mechanism in the LSTM module is shown in the following equation, which focuses the attention of the LSTM model on variables in different dimensions: ； in, for The sum of weights, each weight Defined as , For Hadamard product, It is a fully connected layer; Cost functions include L 2 The loss function consists of two parts: regularization and class-imbalanced loss function. ； Among them, the category is i hour, =1, otherwise =0; k It is the number of categories. p It is the output of the model; Calculate according to the following formula: ； Among them, it means j Number of samples in the class δ For hyperparameters; Step 105: Train the action recognition network using the action data training set to obtain an action recognition model; validate the action recognition model using the action data validation set; test the action recognition model using the action recognition model test set to obtain the best-performing human action recognition model; wherein, the action categories recognized by the human action recognition model include: tremor, bradykinesia, or gait. Step 200: Extract the patient's tremor and bradykinesia characteristics and perform Pearson correlation analysis; Step 300: Obtain the patient's gait parameters; Step 300 includes the following steps 301-303: Step 301: Calculate the gait characteristics required for gait analysis using acceleration and angular velocity data; Step 302: Transform the acceleration in the sensor coordinate system to the navigation coordinate system and perform double integration to obtain the spatial position in the navigation coordinate system; and correct for the spatiotemporal parameter drift. Step 303: Use TCN to extract features of inertial signals, realize the perception of gait time series information through TCN, and use dilated convolution to expand the receptive field of signals in Parkinson's gait analysis; according to the multi-layer residual connection structure, realize the stepwise calculation of inertial signals from local to global, and obtain gait parameters that can characterize Parkinson's movement disorder and stroke. Step 400: Using the tremor features, first bradykinesia features, second bradykinesia features obtained in Step 200 and the gait parameters obtained in Step 300, obtain a movement disorder feature set, and use the medical scale scores of neurology experts as labels; and randomly divide the movement disorder feature set into a movement disorder feature training set and a movement disorder feature test set according to the proportion, and randomly select a part of the movement disorder feature training set as the movement disorder feature validation set. Step 500: Based on a fully connected neural network, establish a quantitative assessment network for movement disorder symptoms; train the quantitative assessment network for movement disorder symptoms using a movement disorder feature training set, and test the network using a movement disorder feature test set to finally obtain a quantitative assessment model for movement disorder symptoms.

2. The modeling method for the quantitative assessment model of movement disorder symptoms according to claim 1, characterized in that, Step 200 includes the following steps 201-204: Step 201: Extract the tremor features from the tremor action segments; The tremor characteristics include: root mean square of the difference between acceleration, angular velocity and surface electromyography signal, sample entropy, mean, variance, covariance, skewness and kurtosis; Step 202: For motion segments with slowed motion, extract the first motion slowness feature; and extract the second motion slowness feature for the corresponding human motion video segment; The first motion sluggishness features include: peak-to-peak value, rectified average value, variance, standard deviation, skewness, kurtosis, root mean square, peak factor, margin factor, and fuzzy entropy of acceleration and angular velocity; mutual information entropy, cross-correlation coefficient, KL divergence of sEMG signals, and fuzzy entropy and permutation entropy of sEMG; roll and pitch angles of the wrist; resultant acceleration and the arithmetic square root of the sum of squares of the three-axis angular velocities; The second characteristic of bradykinesia includes: information on human joint points and joint angles; Step 203: Perform Pearson correlation analysis on the tremor features obtained in Step 201, the first bradykinesia feature and the second bradykinesia feature obtained in Step 202, and the medical scale scores assessed by neurology experts to obtain the Pearson correlation coefficient, and determine the validity of the proposed features based on the Pearson correlation coefficient.

Images