A method and system for evaluating muscle stiffness based on recurrent graph and width learning

Abstract

This invention discloses a method and system for assessing muscle rigidity based on recursive graphs and width learning, belonging to the field of machine learning technology, to solve the problem of inaccurate classification of input muscle tone signals by existing models. The key technical points of this invention include: collecting time-series muscle tone signals from different individuals and preprocessing the signals; constructing a recursive graph using the preprocessed signals; extracting quantized features based on the recursive graph corresponding to the time-series muscle tone signals from different individuals; inputting the quantized features into a width-learning-based classification model for training to obtain a trained model; extracting the quantized features corresponding to the time-series muscle tone signal to be tested and inputting them into the trained model to obtain a muscle rigidity prediction result. This invention achieves time-series visualization while acquiring muscle tone features with significant discriminative power, meeting the requirements for accuracy and efficiency in muscle rigidity assessment.

Description

Technical Field

[0001] This invention relates to the field of machine learning technology, specifically to a method and system for assessing myotonia based on recursive graphs and width learning. Background Technology

[0002] Myotonia is characterized by involuntary, sustained muscle contractions due to increased muscle tone during movement. It is commonly seen in conditions such as myotonic dystrophy, congenital myotonia, and Parkinson's syndrome. Clinically, myotonia is often diagnosed using methods such as muscle tapping, electromyography (EMG), and the BFMDRS scale. However, objectively assessing and quantifying different degrees of myotonia and spasticity, as well as changes in muscle tone in different parts of the body and muscle groups, is difficult. Even subtle changes in muscle tone are hard to evaluate. In clinical practice, the degree of myotonia in patients relies primarily on the physician's clinical experience for judgment; this subjective assessment is inherently uncertain.

[0003] Muscle tone signals acquired using wearable devices are typically presented as time series. Based on the idea of ​​recursive graphs, the recursive phenomena of dynamic systems can be visualized using two-dimensional images, allowing for the processing of non-stationary time series signals. Recursion is a fundamental property of dynamic systems, used to characterize their behavior in phase space. It indicates that under certain conditions, the system's state is recoverable; after a sufficiently long time, a certain state of the system will return to a state near its initial state. Recursive methods can be used to analyze the periodicity, chaos, and non-stationarity of time series, revealing their internal structure and providing prior knowledge regarding similarity, information content, and predictability.

[0004] Traditional neural networks often suffer from complex network structures, large parameters, and high time and computational resource consumption when performing classification. Width-learning systems, on the other hand, consist only of feature mapping layers and augmentation layers, offering strong model scalability and eliminating the need for gradient descent to update weights, thus simplifying computation. Furthermore, width-learning models can improve classification performance by laterally expanding the number of augmentation and feature nodes in the network. Summary of the Invention

[0005] To address this issue, this invention proposes a method and system for assessing muscle rigidity based on recursive graphs and width learning, in order to solve the problem of inaccurate classification of input muscle tone signals by existing models.

[0006] According to one aspect of the present invention, a method for assessing myotonia based on recursive graphs and width learning is provided, the method comprising the following steps:

[0007] Step 1: Collect muscle tone time-series signals from different individuals and preprocess the muscle tone time-series signals; wherein, the muscle tone time-series signals from different individuals include muscle tone time-series signals corresponding to patients with mild myotonia, patients with severe myotonia, and normal individuals without myotonia symptoms;

[0008] Step 2: Construct a recursive graph using the preprocessed muscle tone time series signal;

[0009] Step 3: Extract the recursive graph-based quantitative features corresponding to the time-series muscle tone signals of different individuals;

[0010] Step 4: Input the quantized features into a width-based learning classification model for training to obtain a trained classification model;

[0011] Step 5: Extract the recursive graph-based quantitative features corresponding to the time series signal of muscle tension to be tested, and input them into the trained classification model to obtain the muscle rigidity prediction results.

[0012] Furthermore, the preprocessing described in step one includes removing outliers from the signal.

[0013] Further, the specific steps of step two include: reconstructing the phase space of the muscle tension time series signal according to the principle of time delay reconstruction; calculating the distance between any two points in the reconstructed phase space, wherein the distance corresponds to a recursive value; and drawing a two-dimensional recursive graph using multiple recursive values.

[0014] Furthermore, the quantification features mentioned in step three include recursion rate, diagonal length entropy, and capture time; wherein, recursion rate is a measure of the density of recursive points in the recursion graph; diagonal length entropy is the Shannon entropy of the probability distribution of the diagonal line of length l in the recursion graph; and capture time is the weighted average of the lengths of the vertical line segments in the recursion graph.

[0015] Furthermore, the formula for calculating the recursion rate is as follows:

[0016]

[0017] In the formula, N represents the trajectory length in phase space; R i,j The recursive value is a square matrix consisting of 0s and 1s.

[0018] Furthermore, the formula for calculating the diagonal length entropy is as follows:

[0019]

[0020] In the formula, l represents the length of the diagonal; P(l) represents the probability distribution of the diagonal of length l.

[0021] Furthermore, the formula for calculating the capture time is as follows:

[0022]

[0023] In the formula, v represents a line segment perpendicular to the main diagonal, and P(v) represents the probability distribution of a perpendicular line segment of length v. min This represents the minimum length of a line segment.

[0024] Furthermore, the loss function of the width-based classification model described in step four during training includes the square of the difference between the predicted value and the true value, plus a regularization term based on ridge regression theory. Therefore, the loss function is expressed as:

[0025]

[0026] In the formula, Z n H represents the sequence of mapping features obtained through nonlinear mapping; m W represents the sequence of mapped feature layers; m This represents the output weight matrix; λ represents the regularization coefficient. Represents the actual value;

[0027] The output weight matrix corresponding to the minimum value of the loss function is:

[0028]

[0029] In the formula, I represents the identity matrix.

[0030] Furthermore, the width-based classification model described in step four improves model performance during training by adding additional augmentation nodes. These augmentation nodes are calculated based on the feature layer, and the new augmentation layer is represented as follows:

[0031]

[0032] in, It is the weight matrix and bias matrix that generate new enhanced nodes from the mapped feature nodes; ξ represents a nonlinear function.

[0033] According to another aspect of the present invention, a myotonia assessment system based on recursive graphs and width learning is provided, the system comprising:

[0034] The data acquisition module is configured to acquire muscle tone time-series signals from different individuals and preprocess the muscle tone time-series signals; wherein, the muscle tone time-series signals from different individuals include muscle tone time-series signals corresponding to patients with mild myotonia, patients with severe myotonia, and normal individuals without myotonia symptoms.

[0035] A recursive graph construction module, configured to construct a recursive graph using preprocessed muscle tone time-series signals;

[0036] The feature extraction module is configured to extract recursive graph-based quantitative features corresponding to the time-series muscle tension signals of different individuals.

[0037] The model training module is configured to input the quantized features into a width-based learning classification model for training, and obtain a trained classification model.

[0038] The prediction module is configured to extract the recursive graph-based quantitative features corresponding to the time series signal of muscle tension to be tested, and input them into the trained classification model to obtain the muscle rigidity prediction result.

[0039] The beneficial technical effects of this invention are:

[0040] Clinically, the degree of muscle rigidity is often determined by doctors based on clinical experience. Slight changes in muscle tone are easily overlooked, and the magnitude of muscle tone is difficult to quantify. This invention designs a muscle tone detection device that collects muscle tone signals during specific movements of a body part driven by external force, enabling objective analysis of muscle tone. Time-series signals have large data volumes and high dimensionality, and direct analysis places high demands on the performance of algorithms and hardware. This invention uses a recursive graph to quantify and analyze muscle tone signals, achieving visualization of time series data while obtaining muscle tone feature parameters with significant discriminative power. Width-learning neural networks, which do not depend on depth, have a simple and flexible structure, fast computation speed, and strong nonlinearity, meeting the requirements for accuracy and efficiency in the assessment of muscle rigidity. Attached Figure Description

[0041] The present invention can be better understood by referring to the description given below in conjunction with the accompanying drawings, which together with the following detailed description are included in and form part of this specification, and are used to further illustrate preferred embodiments of the invention and explain the principles and advantages of the invention.

[0042] Figure 1 This is a flowchart of a muscle rigidity assessment method based on recursive graphs and width learning, according to an embodiment of the present invention.

[0043] Figure 2 This is a hardware connection diagram of the muscle tension detection device used in the embodiments of the present invention.

[0044] Figure 3 This is a schematic diagram of the method for selecting phase space reconstruction parameters when constructing a recursive graph in an embodiment of the present invention; wherein, Figure (a) corresponds to the method for determining the delay time; and Figure (b) corresponds to the method for determining the embedding dimension.

[0045] Figure 4This is an example diagram of the recursive graph constructed from the collected muscle tone signals of three types of subjects in an embodiment of the present invention; wherein, Figure (a) corresponds to healthy people; Figure (b) corresponds to mild patients; and Figure (c) corresponds to severe patients.

[0046] Figure 5 The following are example diagrams of the feature results extracted when performing quantitative analysis on the recursive graph in an embodiment of the present invention; wherein, Figure (a) corresponds to the recursion rate; Figure (b) corresponds to the diagonal length entropy; and Figure (c) corresponds to the capture time.

[0047] Figure 6 This is a diagram of the width learning model structure with added additional enhancement nodes used in this embodiment of the invention.

[0048] Figure 7 This is an example diagram of the confusion matrix of the classification results obtained in the embodiments of the present invention. Detailed Implementation

[0049] To enable those skilled in the art to better understand the present invention, exemplary embodiments or examples of the present invention will be described below in conjunction with the accompanying drawings. Obviously, the described embodiments or examples are merely some, not all, of the embodiments or examples of the present invention. All other embodiments or examples obtained by those skilled in the art based on the embodiments or examples of the present invention without inventive effort should fall within the scope of protection of the present invention.

[0050] This invention proposes a method and system for assessing muscle rigidity based on recursive graphs and width learning. By quantitatively analyzing muscle tone signals, it solves the problem of the difficulty in objectively assessing changes in muscle tone.

[0051] This invention proposes a method for assessing muscle rigidity based on recursive graphs and width learning, such as... Figure 1 As shown, the method includes the following steps:

[0052] Step 1: Collect muscle tone time-series signals from different individuals and preprocess the muscle tone time-series signals; wherein, the muscle tone time-series signals from different individuals include muscle tone time-series signals corresponding to patients with mild myotonia, patients with severe myotonia, and normal individuals without myotonia symptoms;

[0053] Step 2: Construct a recursive graph using the preprocessed muscle tone time series signal;

[0054] Step 3: Extract the recursive graph-based quantitative features corresponding to the time-series muscle tone signals of different individuals;

[0055] Step 4: Input the quantized features into a width-based learning classification model for training to obtain a trained classification model;

[0056] Step 5: Extract the recursive graph-based quantitative features corresponding to the time series signal of muscle tension to be tested, and input them into the trained classification model to obtain the muscle rigidity prediction results.

[0057] In step one, muscle tone signals are acquired, and outliers in the data are preprocessed.

[0058] According to an embodiment of the present invention, for data acquisition: muscle tone data of specific body parts during movement of the test subject are collected through clinical trials. During specific reciprocating movements of the body parts driven by external force, the resistance and position information experienced by the device are recorded. Starting with the arm extended, a motor drives the upper arm to perform specific movements such as uniform low-speed movement and uniform high-speed movement. The angle of arm rotation is collected by an encoder. The magnitude of the force output by the motor when driving the arm movement is equal to the resistance of the arm. The hardware connection of the detection device is as follows... Figure 2 As shown. Muscle tone data of the right arm were collected from healthy individuals, patients with mild myotonia, and patients with severe myotonia.

[0059] For data preprocessing: Due to unavoidable noise, equipment malfunctions, and human error during data acquisition, the acquired data often contains various problems that affect the data analysis results. Therefore, data preprocessing is performed first to improve data quality. An outlier detection method based on statistical characteristics is used to check the motor output torque data for each test subject during the test. The average absolute value of the torque under each motion command is calculated. If the average value of a certain cycle is greater than that of other cycles, the motion data for that cycle is very likely affected by the force actively applied by the test subject, and such data is discarded.

[0060] In step two, a recursive graph is constructed to analyze the muscle tone time series signal. Specific steps include:

[0061] (1) Phase space reconstruction: The time series is reconstructed in phase space using time delay reconstruction technology;

[0062] (2) Distance calculation: Calculate the distance between any two points in the reconstructed phase space;

[0063] (3) Recursive value calculation: The distance between any two points in phase space corresponds to a recursive value;

[0064] (4) Drawing the recursion graph: Draw the recursive values ​​as a two-dimensional recursion graph.

[0065] According to an embodiment of the present invention, let the muscle tone time series signal be u i For each i = 1, 2, ..., n, a recursive graph is constructed to analyze the time series signal of muscle tone. The specific steps include:

[0066] (1) Phase space reconstruction: Let X = {X1, X2, X3, ..., X} N}, X i =[u i ,u i+τ ,u i+2τ ,...,u i+(m-1)τ ], where i = 1, 2, ..., N, N = n - (m - 1)τ, and N is the trajectory length of X in the reconstructed phase space. m is the embedding dimension, determined using the spurious nearest neighbor method, such as Figure 3 As shown in (a), m = 3. τ is the delay time, the value corresponding to the first minimum value of the mutual information function in the average mutual information method, as shown in... Figure 3 As shown in (b), τ = 5.

[0067] (2) Distance calculation: Calculate the distance r between any two points in the reconstructed phase space. ij =||X i -X j ||, where i,j=1,2,...,N, and ||·|| denotes the norm.

[0068] (3) Recursive value calculation: The distance between any two points in phase space corresponds to a recursive value, R. i,j =Θ(ε-||X) i -X j ||)=Θ(ε-r ij ), where i,j=1,2,...,N, ε is the preset threshold distance, and Θ(x) is the Heaviside function.

[0069] (4) Recursion graph drawing: R i,j Let R be an N×N square matrix consisting of 0s and 1s, with i as the x-coordinate and j as the y-coordinate on a two-dimensional coordinate axis. i,j When R is 1, the pixel at coordinate (i,j) is 0. i,j When the value is 0, the pixel at (i,j) is 255. The recursive graph showing the conversion of muscle tone signals among the three types of subjects is as follows: Figure 4 As shown.

[0070] Through the Figure 4 Analysis reveals that all three types of recursion graphs exhibit a clear diagonal line parallel to the main diagonal, corresponding to the fixed instruction cycle in the instruction settings. The recursion graphs of healthy individuals and mildly ill patients show some vertical and horizontal lines, indicating that the system's phase space trajectory remains approximately constant during a certain process, with minimal change in torque values, corresponding to the pause time after each instruction execution. The recursion graph of severely ill patients shows numerous arcs, indicating that two closely spaced trajectories in phase space changed at a certain time.

[0071] In step three, quantization features based on recursive graphs are extracted from the time-series muscle tone signals of different individuals, and the graphical features of the recursive graphs are analyzed using recursive quantization. Based on the density of recursive points and structural features such as diagonals and vertical lines in the recursive graph, certain properties of the recursive graph are quantized. Recursive features with high discriminative power are retained.

[0072] According to an embodiment of the present invention, recursive quantization analysis is performed on certain features of the recursive graph, and the results are as follows: Figure 5 As shown, the specific description of the features is as follows:

[0073] (1) The recursion rate is a simple measure of the density of recursive points in a recursive graph, expressed as: N is the trajectory length in phase space.

[0074] (2) The diagonal length entropy is the Shannon entropy of the probability distribution of the diagonal of length l in the recursive graph, expressed as: P(l) represents the probability distribution along the diagonal of length l. It reflects the complexity of the recursion graph relative to the diagonal. For uncorrelated time series, such as white noise, the entropy value is very low, indicating that the signal has low complexity.

[0075] (3) The capture time refers to the weighted average of the lengths of the vertical line segments in the recursive graph, expressed as... Let P(v) represent the line segment perpendicular to the main diagonal, where v represents the line segment perpendicular to the main diagonal, and P(v) is the probability distribution of the perpendicular line segment of length v. min It is the minimum line segment length, typically taken as 2. This value reflects the duration for which the system remains in a specific state and can be used to evaluate the system's stability.

[0076] Depend on Figure 5 It can be seen that the quantitative features of muscle tone signals based on recursion graphs have good regional differentiation for the three types of subjects.

[0077] In step four, the quantized features are input into the width-based learning classification model for training to obtain the trained classification model.

[0078] According to an embodiment of the present invention, a width learning model is first constructed. Input data X is processed through a nonlinear mapping to obtain a feature layer Zn, which in turn generates an enhancement layer Hm. The enhancement layer and the feature layer are connected in parallel and processed through an output weight matrix Wm to obtain the predicted value output by the width learning model, which is the assessment result of myotonia. The network structure is as follows: Figure 6 As shown.

[0079] (1) Let the input data be X, and then use a nonlinear mapping... The mapping features are obtained, where It is a non-linear activation function, let and The input weight matrix and bias vector are randomly generated. To make the generated mapping features more concise, a sparse autoencoder concept is introduced to optimize the input weight matrix. Let Z n =[Z1,Z2,...,Z n This allows random features to be generated into a more sparse and compact feature layer.

[0080] (2) Mapping feature layer through Generate enhanced nodes, where and It is randomly generated, a nonlinear function ξ(x) = 1 / (1+e -x Let the mapping layer be H. m =[H1,H2,...,H m ].

[0081] (3) The enhancement layer and the mapping layer are connected in parallel and passed through the output weight matrix W. m Get the output, use To represent the true value, the predicted value Y = [Z] of the width learning network. n |H m W m .

[0082] Then, the dataset is divided, and a width learning model is trained. During training, the loss function is defined as the square of the difference between the predicted value and the true value. Weight matrix W m The solution is obtained using the least squares method. To avoid [Z n |H m When the matrix is ​​not a full column rank matrix, it can cause a large error. Based on ridge regression theory, a regularization term is added to the loss function, and the loss function becomes: λ represents the regularization coefficient; the output weight matrix corresponding to the minimum value of the loss function is... I represents the identity matrix.

[0083] During training, if the performance of width learning does not meet expectations, model performance can be improved by adding additional augmentation nodes. These new nodes are calculated based on the feature layers. The representation of the new augmentation layer is as follows: in It is the weight matrix and bias matrix that generate new enhanced nodes from the mapped feature nodes.

[0084] The quantized feature set of the muscle tone signal recurrence map was divided into training and test sets using a five-fold cross-validation method. The constructed width learning algorithm was trained using the training set. When the network performance reached the expected standard, the number of nodes and the relevant values ​​of the output weight matrix were saved. The performance of the method was validated using data from the test set. The evaluation results for each test set were statistically analyzed, showing a precision of 89.06%, an accuracy of 83.57%, and a recall of 82.26%. A confusion matrix was plotted, as shown below. Figure 7 As shown.

[0085] Another embodiment of the present invention proposes a myotonia assessment system based on recursive graphs and width learning, the system comprising:

[0086] The data acquisition module is configured to acquire muscle tone time-series signals from different individuals and preprocess the muscle tone time-series signals; wherein, the muscle tone time-series signals from different individuals include muscle tone time-series signals corresponding to patients with mild myotonia, patients with severe myotonia, and normal individuals without myotonia symptoms.

[0087] A recursive graph construction module, configured to construct a recursive graph using preprocessed muscle tone time-series signals;

[0088] The feature extraction module is configured to extract recursive graph-based quantitative features corresponding to the time-series muscle tension signals of different individuals.

[0089] The model training module is configured to input the quantized features into a width-based learning classification model for training, and obtain a trained classification model.

[0090] The prediction module is configured to extract the recursive graph-based quantitative features corresponding to the time series signal of muscle tension to be tested, and input them into the trained classification model to obtain the muscle rigidity prediction result.

[0091] The functionality of the myotonia assessment system based on recursive graphs and width learning in this embodiment of the invention can be described by the aforementioned myotonia assessment method based on recursive graphs and width learning. Therefore, for the parts not detailed in the system embodiment, please refer to the above method embodiment, and they will not be repeated here.

[0092] Although the invention has been described with respect to a limited number of embodiments, those skilled in the art will understand from the foregoing description that other embodiments are conceivable within the scope of the invention described herein. The disclosure of the invention is illustrative and not restrictive, and the scope of the invention is defined by the appended claims.

Claims

1. A method for assessing muscle rigidity based on recursive graphs and width learning, characterized in that, Includes the following steps: Step 1: Collect muscle tone time-series signals from different individuals and preprocess the muscle tone time-series signals; wherein, the muscle tone time-series signals from different individuals include muscle tone time-series signals corresponding to patients with mild myotonia, patients with severe myotonia, and normal individuals without myotonia symptoms; Step 2: Construct a recursive graph using the preprocessed muscle tone time series signal, including: reconstructing the phase space of the muscle tone time series signal according to the time delay reconstruction principle; calculating the distance between any two points in the reconstructed phase space, where each distance corresponds to a recursive value; and drawing a two-dimensional recursive graph using multiple recursive values. Step 3: Extract the recursive graph-based quantitative features corresponding to the time-series muscle tone signals of different individuals; Step 4: Input the quantized features into a width-based learning classification model for training to obtain a trained classification model; the loss function of the width-based learning classification model during training includes the square of the difference between the predicted value and the true value, plus a regularization term based on ridge regression theory, then the loss function is expressed as: In the formula, Z n H represents the sequence of mapping features obtained through nonlinear mapping; m W represents the sequence of mapped feature layers; m This represents the output weight matrix; λ represents the regularization coefficient. Represents the actual value; The output weight matrix corresponding to the minimum value of the loss function is: In the formula, I represents the identity matrix; Step 5: Extract the recursive graph-based quantitative features corresponding to the time series signal of muscle tension to be tested, and input them into the trained classification model to obtain the muscle rigidity prediction results.

2. The method for assessing muscle rigidity based on recursive graphs and width learning according to claim 1, characterized in that, The preprocessing described in step one includes removing outliers from the signal.

3. The method for assessing muscle rigidity based on recursive graphs and width learning according to claim 2, characterized in that, The quantification features described in step three include recursion rate, diagonal length entropy, and capture time; where recursion rate is a measure of the density of recursive points in the recursion graph; diagonal length entropy is the Shannon entropy of the probability distribution of the diagonal line of length l in the recursion graph; and capture time is the weighted average of the lengths of the vertical line segments in the recursion graph.

4. The method for assessing muscle rigidity based on recursive graphs and width learning according to claim 3, characterized in that, The formula for calculating the recursion rate is as follows: In the formula, N represents the trajectory length in phase space; R i,j The recursive value is a square matrix consisting of 0s and 1s.

5. The method for assessing muscle rigidity based on recursive graphs and width learning according to claim 3, characterized in that, The formula for calculating the diagonal length entropy is as follows: In the formula, l represents the length of the diagonal; P(l) represents the probability distribution of the diagonal of length l.

6. The method for assessing muscle rigidity based on recursive graphs and width learning according to claim 3, characterized in that, The formula for calculating the capture time is as follows: In the formula, v represents a line segment perpendicular to the main diagonal, and P(v) represents the probability distribution of a perpendicular line segment of length v. min This represents the minimum length of a line segment.

7. The method for assessing muscle rigidity based on recursive graphs and width learning according to claim 1, characterized in that, The width-based classification model described in step four improves its performance during training by adding additional augmentation nodes. These augmentation nodes are calculated based on the feature layers, and the new augmentation layer is represented as follows: in, It is the weight matrix and bias matrix that generate new enhanced nodes from the mapped feature nodes; ξ represents a nonlinear function.

8. A muscle rigidity assessment system based on recursive graphs and width learning, characterized in that, The system is used to implement the muscle rigidity assessment method based on recursive graphs and width learning as described in any one of claims 1-7; the system comprises: The data acquisition module is configured to acquire muscle tone time-series signals from different individuals and preprocess the muscle tone time-series signals; wherein, the muscle tone time-series signals from different individuals include muscle tone time-series signals corresponding to patients with mild myotonia, patients with severe myotonia, and normal individuals without myotonia symptoms. A recursive graph construction module, configured to construct a recursive graph using preprocessed muscle tone time-series signals; The feature extraction module is configured to extract recursive graph-based quantitative features corresponding to the time-series muscle tension signals of different individuals. The model training module is configured to input the quantized features into a width-based learning classification model for training, and obtain a trained classification model. The prediction module is configured to extract the recursive graph-based quantitative features corresponding to the time series signal of muscle tension to be tested, and input them into the trained classification model to obtain the muscle rigidity prediction result.

Images