Autism rTMS treatment effect prediction system based on time variability network

Abstract

The invention discloses an autism rTMS treatment effect prediction system based on a time variability network, and belongs to the field of electroencephalogram data processing. According to the method, the long-term rTMS intervention curative effect of the autism patient is predicted by adopting the time variability network of the resting-state electroencephalogram. Compared with a traditional clinical evaluation method, the electroencephalogram can more objectively and directly record the change of the brain nerve activity, so that the mining of a robust predictive factor is realized. Due to the application of the resting-state electroencephalogram, the difficulty of data acquisition can be simplified, and behavior evaluation errors caused by subjective factors and the like are avoided; the time variability network analysis based on the resting-state electroencephalogram can reveal the complex dynamic characteristics of the brain network loop in time and space, and further successfully analyzes how the brain is recombined in different pathological states. In conclusion, the method breaks away from the framework of traditional clinical evaluation, predicts the long-term rTMS intervention curative effect of autism by applying objective electroencephalogram network characteristics, and has important significance for clinically formulating a dynamically adjustable autism personalized treatment strategy.

Description

Technical Field

[0001] The invention belongs to the field of biomedical engineering, and particularly relates to an autism rTMS treatment efficacy prediction system based on a time variability network. Background Art

[0002] Autism spectrum disorder (ASD) is a neurodevelopmental disorder characterized by persistent social interaction and interaction deficits, restricted interests and activities, and repetitive stereotyped behavior patterns. As a safe and effective noninvasive neural intervention and regulation technology, repetitive transcranial magnetic stimulation (rTMS) can increase the excitability of underactive cortical areas and related network loops, and has potential application prospects in the intervention and treatment of ASD. However, the intervention effect of rTMS regulation is very different between individuals, and current technology cannot determine the response after long-term intervention before the initial treatment of patients, resulting in unnecessary delays. Therefore, identifying robust predictors is of great significance for assisting clinical prediction of the long-term therapeutic efficacy of rTMS for ASD. EEG can measure the synchronous sum of postsynaptic potentials of pyramidal cells in real time through electrodes placed on the surface of the scalp, which has the advantages of non-invasiveness, high temporal resolution, portability and low cost. Compared with the time delay of traditional behavioral assessment, EEG can directly, objectively and in real time record and evaluate the effect of rTMS on brain neural activity; among them, EEG-based time variability network analysis can reveal the complex dynamic characteristics of brain network loops in time and space, which is crucial for understanding how the brain is reorganized under different pathological conditions. This paper proposes a prediction system for the efficacy of rTMS treatment of autism based on time variability network. The system mainly establishes a multivariate linear regression prediction model based on objective biomarkers derived from the EEG time variability network of ASD patients before treatment to achieve the evaluation of its long-term intervention efficacy. In this system, the dynamic fluctuation degree of brain time-varying functional network connections is first calculated by fuzzy entropy method to reveal the rhythmic fluctuation pattern of network loops over time. Based on this, the potential relationship between the fluctuation characteristics of the time variability network and the ASD clinical scale is mined to determine robust biomarkers, and a multivariate linear regression model is further established to predict the long-term efficacy of rTMS intervention. This system breaks away from the limitation of strong subjectivity of traditional clinical evaluation and uses objective EEG network characteristics to predict the efficacy of rTMS intervention, which is of great significance for the clinical formulation of dynamically adjustable personalized treatment strategies for autism. Summary of the invention

[0003] In order to address the problems of large variability in the therapeutic effect of rTMS intervention on autism patients and strong subjectivity in clinical evaluation, the present invention proposes an autism rTMS treatment efficacy prediction system based on a time-variable network.

[0004] The present invention proposes a rTMS treatment efficacy prediction system for autism based on a time-variable network, the system comprising: a preprocessing module, a frequency division module, a time-varying connection matrix calculation module, a time-variable network construction module, a network attribute calculation module, and a prediction module;

[0005] The method of the preprocessing module is:

[0006] The collected head surface EEG signals were preprocessed as follows: reference electrode standardization was used for re-reference, independent component analysis was used to remove artifact components, and bandpass filtering was used to filter out signal noise components;

[0007] The method of frequency division module is:

[0008] Perform frequency division processing on the output signal of the preprocessing module:

[0009] According to its frequency distribution, it is divided into delta: 1-4HZ, theta: 4-8HZ, alpha: 8-13HZ and beta: 13-30HZ bands;

[0010] The method of the time-varying connection matrix calculation module is:

[0011] The signal of each frequency band output by the frequency division module is processed in segments. For the signal of each frequency band, a sliding window with a window length of q and an overlap of D is used to segment the signal. For each data segment, the phase locking value method is used to evaluate the synchronization strength of the EEG signals u(t) and v(t) of electrodes at two different positions. Specifically, for the i-th segment data, the phase locking value w between the EEG signals of all electrode pairs is calculated. PLV , forming a time-varying connection matrix w(i);

[0012] The specific method of the time-variable network construction module is:

[0013] Calculate the fuzzy entropy of each connection edge sequence of the time-varying connection matrix w(i) of each frequency band to measure the time complexity of the EEG functional network, so as to construct a time-variable network;

[0014] The specific method of the network attribute calculation module is:

[0015] Based on the time-variable network, the following network properties were calculated using the Brain Functional Network Toolbox:

[0016] Characteristic path length

[0017] Local efficiency

[0018] Global efficiency

[0019] Clustering coefficient Among them, d ij is the length of the weighted shortest path between nodes i and j, θ represents the set of all network nodes, N represents the number of nodes, and w ij Represents the weighted value of the connection strength between nodes i and j;

[0020] The specific method of the prediction module is:

[0021] Based on the network attribute data of the four frequency bands, a stepwise multivariate linear regression model was established and combined with leave-one-out cross-validation to predict the clinical status of autistic patients after rTMS intervention in the subsequent time period.

[0022] Furthermore, the method for calculating the phase locking value between paired electrode signals in the time-varying connection matrix calculation module is as follows:

[0023] Step S31: Perform Hilbert transform on the signals u(t) and v(t) to obtain the corresponding analytical signal H u (t) = u(t) + iHT u (t) and H v (t) = v (t) + iHT v (t), where and Where PV is the Cauchy principal value, t' represents the derivative of t;

[0024] Step S32: Calculate the analytical signal H respectively u (t) and H v The phase of (t) and

[0025] Step S33: Calculate the phase lock value Where N is the number of sampling points, j is the jth sampling point of the signal, t and Δt are the time point and sampling period respectively, and w PLV is the network connectivity weight.

[0026] Furthermore, the method for calculating the fuzzy entropy in the time variability network building module is as follows:

[0027] Step S41: After the sliding window segmentation, M functional networks are obtained, and a window with a length of q is set to obtain a connection matrix sequence in represents the average value of the connection matrix sequence;

[0028] Step S42: Calculate two adjacent connection matrix sequences W i q and W l q Similarity index in W i q and W l q The maximum absolute difference of the scalar, r is the similarity tolerance;

[0029] Step S43: For each vector W i q ,calculate Expected neighboring vector W l q The average similarity

[0030] Step S44: For the time series W i , 1≤i≤M, the fuzzy entropy is calculated as:

[0031]

[0032] Further statistical estimation yields

[0033] Furthermore, the construction method of the multivariate linear regression prediction model in the prediction module is as follows:

[0034] Step S61: using the leave-one-out cross-validation method, the data of P-1 patients are selected as the training set, a stepwise multivariate linear regression prediction model is established, and the data of the remaining patient is used to verify the prediction result;

[0035] Step S62: In each cycle of the model, set the dependent variable Y and the network attributes of the four frequency bands, a total of 16 independent variables x 1 , x 2 , …, x 16 , and randomly introduce any initial independent variables in the initial stage;

[0036] Step S63: For the variable x already in the equation ik , calculate its partial regression sum of squares Where Q(...) represents the residual sum of squares of the regression model of the variables in brackets, and r represents the number of variables already in the equation;

[0037] Step S64: Calculate all Minimum value of Get the independent variable x that has the smallest impact on the dependent variable i0 ;

[0038] Step S65: Use the univariate analysis of F statistics to test x i0 Whether the influence on the dependent variable is significant, the insignificant variables are eliminated according to the set threshold p; i0 The relevant F statistic is expressed as Further Used to determine whether it is significant. The F statistic follows the distribution F~F(1,nr-2). If Then it means x i0 The effect on the dependent variable is not significant, so it is eliminated;

[0039] Step S66: For the variable x that does not exist in the equation jk , calculate its partial regression sum of squares Where Q(...) represents the residual sum of squares of the regression model of the variables in brackets, and r represents the number of variables already in the equation;

[0040] Step S67: Calculate all The maximum value of Get the independent variable x that has the greatest impact on the dependent variable j0 ;

[0041] Step S68: Use the univariate analysis of F statistics to test x j0 Whether the influence on the dependent variable is significant, a significant variable is introduced according to the set threshold p; j0 The relevant F statistic is expressed as Further Used to determine whether it is significant. The F statistic follows the distribution F~F(1,nr-2). If Then it means x j0 The effect on the dependent variable is significant, so it is added to the model;

[0042] Step S69: Repeat step S63 until no new variables are eliminated or introduced;

[0043] Step S610: Repeat steps S61-S69 until all subjects are treated as a test set and obtain predicted values.

[0044] Furthermore, in the time-varying connection matrix calculation module, a sliding window with a window length q of 20 and an overlap degree D of 0.98 is set to segment the signal.

[0045] Furthermore, in step S42 of the method for calculating the fuzzy entropy, r is set to 0.2 multiplied by the standard deviation of the time series.

[0046] Furthermore, in the method for constructing the multivariate linear regression prediction model, the threshold value p=0.05; if p≥0.05 indicates that it is not significant, then x is eliminated.i0 If p < 0.05, it indicates significant, then x j0 variable.

[0047] The present invention uses the time variability network of resting-state EEG to predict the efficacy of long-term rTMS intervention in patients with autism. Compared with traditional clinical evaluation methods, EEG can more objectively and directly record the changes in brain neural activity, thereby realizing the mining of robust predictive factors. In addition, the application of resting-state EEG can simplify the difficulty of data collection and avoid behavioral evaluation errors caused by subjective factors, etc.; and the time variability network analysis based on resting-state EEG can reveal the complex dynamic characteristics of brain network loops in time and space, and further successfully analyze how the brain is reorganized under different pathological conditions. In summary, the present invention breaks away from the framework of traditional clinical evaluation and uses objective EEG network characteristics to predict the efficacy of long-term rTMS intervention for autism, which is of great significance for the clinical formulation of dynamically adjustable personalized treatment strategies for autism. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 This is a schematic diagram of the use of an autism rTMS treatment efficacy prediction system based on a time-variable network proposed by the present invention.

[0049] Figure 2 A schematic diagram of the steps for establishing a system for predicting the efficacy of rTMS treatment of autism based on a time-variable network provided by the present invention.

[0050] Figure 3 A schematic diagram of the specific steps for establishing the stepwise multiple linear regression model provided by the present invention.

[0051] Figure 4 This is the prediction effect of the autism rTMS treatment efficacy prediction system built by the present invention. DETAILED DESCRIPTION

[0052] The implementation of the present invention is further described below in conjunction with the accompanying drawings.

[0053] In the first aspect, according to an embodiment of the present invention, a method for predicting the efficacy of rTMS treatment for autism based on a time-variable network is proposed. Figure 1 , including the following steps:

[0054] Step S1: Preprocess the EEG signals collected by the head meter in the following steps:

[0055] The reference electrode standardization technique was used to re-reference the EEG signal, independent component analysis was used to remove artifact components, and band-pass filtering was used to filter out signal noise components;

[0056] Step S2: Divide the preprocessed EEG signals into delta (1-4 Hz), theta (4-8 Hz), alpha (8-13 Hz) and beta (13-30 Hz) frequency bands according to their frequency distribution;

[0057] Step S3: For each frequency band of the EEG signal, use a sliding window with a window length of 20s and an overlap of 98% to segment the signal. For each data segment, the phase locking value method is used to evaluate the synchronization strength of the EEG signals u(t) and v(t) of two electrodes at different locations. Specifically, for the i-th segment data, calculate the phase locking value w between the EEG signals of all electrode pairs PLV , forming a W i Connection Matrix;

[0058] The steps to calculate the phase lock value between paired electrode signals are as follows:

[0059] Step S31: Perform Hilbert transform on the signals u(t) and v(t) to obtain the corresponding analytical signal H u (t) = u(t) + iHT u (t) and H v (t) = v (t) + iHT v (t), where and Where PV is the Cauchy principal value, t' represents the derivative of t;

[0060] Step S32: Calculate the analytical signal H respectively u (t) and H v The phase of (t) and

[0061] Step S33: Calculate the phase lock value Where N is the number of sampling points, j is the jth sampling point of the signal, t and Δt are the time point and sampling period respectively, and w PLV is the network connectivity weight.

[0062] Step S4: Calculate the fuzzy entropy of each connection edge of the time-varying connection matrix obtained in step S3 to measure the time complexity of the EEG functional network, thereby obtaining a time-variable network;

[0063] Step S41: Assuming that M functional networks are obtained after the sliding window segmentation in step S4, a window with a length of q is set to obtain a connection matrix sequence in Represents the average value of the connection matrix sequence. A reasonable q can ensure a more robust reconstruction of the dynamic process. Here we set q to 2;

[0064] Step S42: Calculate two adjacent connection matrix sequences W i q and W l q Similarity index in W i q and W l q The maximum absolute difference of the scalar, r is the similarity tolerance. Considering that a smaller r will bring noise, and a larger r will cause information loss, r is set to 0.2 times the standard deviation of the time series;

[0065] Step S43: For each vector W i q (i=1,2,…,M-q+1), calculate Expected neighboring vector W l q The average similarity

[0066] Step S44: For the time series W i (1≤i≤M), the fuzzy entropy is calculated as:

[0067]

[0068] Further statistical estimation yields

[0069] Step S5: Based on the variability network obtained in step S4, the following network properties are calculated using the brain functional network toolbox:

[0070] Characteristic path length

[0071] Local efficiency

[0072] Global efficiency

[0073] Clustering coefficient Among them, d ij is the length of the weighted shortest path between nodes i and j, θ represents the set of all network nodes, N represents the number of nodes, and w ij Represents the weighted value of the connection strength between nodes i and j;

[0074] Step S6: The calculated network attributes of the four frequency bands are input into the constructed multiple linear regression model to predict the score of the patient's clinical condition one month after the end of the rTMS intervention treatment.

[0075] Step S7: Calculate the Pearson correlation coefficient between the patient's actual clinical scale score and the predicted value Root mean square error And the normalized root mean square error Evaluate the prediction effect of the model, where Y represents the predicted value, X represents the true value, and P represents the number of subjects. Represents the mean of the true clinical scores of all patients.

[0076] The present invention was applied to the head surface EEG data of 24 autistic children before treatment to predict the efficacy of the patients after one month of rTMS intervention treatment. The results obtained after Pearson correlation analysis of the predicted value and the true value of the clinical scale are as follows: Figure 4 shown. Figure 4 The results showed that the predicted values ​​and true values ​​of the clinical scales of patients with autism were significantly correlated, indicating that using the network properties of the time-variable network to predict the effect of rTMS treatment can achieve a good effect.

Claims

1. A system for predicting the efficacy of rTMS treatment for autism based on a time-variable network, the system comprising: Preprocessing module, frequency division module, time-varying connection matrix calculation module, time-variable network construction module, network attribute calculation module, prediction module; The method of the preprocessing module is: The collected head surface EEG signals were preprocessed as follows: reference electrode standardization was used for re-reference, independent component analysis was used to remove artifact components, and bandpass filtering was used to filter out signal noise components; The method of frequency division module is: Perform frequency division processing on the output signal of the preprocessing module: According to its frequency distribution, it is divided into delta: 1-4HZ, theta: 4-8HZ, alpha: 8-13HZ and beta: 13-30HZ bands; The method of the time-varying connection matrix calculation module is: The signal of each frequency band output by the frequency division module is processed in segments. For the signal of each frequency band, a sliding window with a window length of q and an overlap of D is used to segment the signal. For each data segment, the phase locking value method is used to evaluate the synchronization strength of the EEG signals u(t) and v(t) of electrodes at two different positions. Specifically, for the i-th segment data, the phase locking value w between the EEG signals of all electrode pairs is calculated. PLV , forming a time-varying connection matrix w(i); The specific method of the time-variable network construction module is: Calculate the fuzzy entropy of each connection edge sequence of the time-varying connection matrix w(i) of each frequency band to measure the time complexity of the EEG functional network, so as to construct a time-variable network; The specific method of the network attribute calculation module is: Based on the time-variable network, the following network properties were calculated using the Brain Functional Network Toolbox: Characteristic path length Local efficiency Global efficiency Clustering coefficient Among them, d ij is the length of the weighted shortest path between nodes i and j, θ represents the set of all network nodes, N represents the number of nodes, and w ij Represents the weighted value of the connection strength between nodes i and j; The specific method of the prediction module is: Based on the network attribute data of the four frequency bands, a stepwise multivariate linear regression model was established and combined with leave-one-out cross-validation to predict the clinical status of autistic patients after rTMS intervention in the subsequent time period.

2. The autism rTMS treatment efficacy prediction system based on time-variable network as claimed in claim 1, characterized in that: The method for calculating the phase locking value between paired electrode signals in the time-varying connection matrix calculation module is as follows: Step S31: Perform Hilbert transform on the signals u(t) and v(t) to obtain the corresponding analytical signal H u (t) = u(t) + iHT u (t) and H v (t) = v (t) + iHT v (t), where and Where PV is the Cauchy principal value, t' represents the derivative of t; Step S32: Calculate the analytical signal H respectively u (t) and H v The phase of (t) and Step S33: Calculate the phase lock value Where N is the number of sampling points, j is the jth sampling point of the signal, t and Δt are the time point and sampling period respectively, and w PLV is the network connectivity weight.

3. The autism rTMS treatment efficacy prediction system based on time-variable network as claimed in claim 1, characterized in that: The method for calculating fuzzy entropy in the time-variable network building module is as follows: Step S41: After the sliding window segmentation, M functional networks are obtained, and a window with a length of q is set to obtain a connection matrix sequence in represents the average value of the connectivity matrix sequence; Step S42: Calculate two adjacent connection matrix sequences and Similarity index in W i q and The maximum absolute difference of the scalar, r is the similarity tolerance; Step S43: For each vector calculate Expected neighboring vectors The average similarity Step S44: For the time series W i , 1≤i≤M, and the fuzzy entropy is calculated as Further statistical estimation yields 4. The autism rTMS treatment efficacy prediction system based on time-variable network as claimed in claim 1, characterized in that: The construction method of the multivariate linear regression prediction model in the prediction module is as follows: Step S61: using the leave-one-out cross-validation method, the data of P-1 patients are selected as the training set, a stepwise multivariate linear regression prediction model is established, and the data of the remaining patient is used to verify the prediction result; Step S62: In each cycle of the model, set the dependent variable Y and the network attributes of the four frequency bands, a total of 16 independent variables x1, x2, ..., x 16 , and randomly introduce any initial independent variables in the initial stage; Step S63: For the variable x already in the equation ik , calculate its partial regression sum of squares Where Q(...) represents the residual sum of squares of the regression model of the variables in brackets, and r represents the number of variables already in the equation; Step S64: Calculate all Minimum value of Get the independent variable x that has the smallest impact on the dependent variable i0 ; Step S65: Use the univariate analysis of F statistics to test x i0 Whether the influence on the dependent variable is significant, the insignificant variables are eliminated according to the set threshold p; i0 The relevant F statistic is expressed as Further Used to determine whether it is significant. The F statistic follows the distribution F~F(1,nr-2). If Then it means x i0 The effect on the dependent variable is not significant, so it is eliminated; Step S66: For the variable x that does not exist in the equation jk , calculate its partial regression sum of squares Where Q(...) represents the residual sum of squares of the regression model of the variables in brackets, and r represents the number of variables already in the equation; Step S67: Calculate all The maximum value of Get the independent variable x that has the greatest impact on the dependent variable j0 ; Step S68: Use the univariate analysis of F statistics to test x j0 Whether the influence on the dependent variable is significant, a significant variable is introduced according to the set threshold p; j0 The relevant F statistic is expressed as Further Used to determine whether it is significant. The F statistic follows the distribution F~F(1,nr-2). If Then it means x j0 The effect on the dependent variable is significant, so it is added to the model; Step S69: repeat step S63 until no new variables are eliminated or introduced; Step S610: Repeat steps S61-S69 until all subjects are treated as a test set and obtain predicted values.

5. The autism rTMS treatment efficacy prediction system based on time-variable network as claimed in claim 1, characterized in that: In the time-varying connection matrix calculation module, a sliding window with a window length q of 20 and an overlap degree D of 0.98 is set to segment the signal.

6. The autism rTMS treatment efficacy prediction system based on time-variable network as claimed in claim 1, characterized in that: Steps of the method for calculating fuzzy entropy In step S42, r is set to 0.2 multiplied by the standard deviation of the time series.

7. The autism rTMS treatment efficacy prediction system based on time-variable network as claimed in claim 1, characterized in that: In the method for constructing the multivariate linear regression prediction model, the threshold value p=0.05; if p≥0.05 indicates insignificance, then x is eliminated. i0 variable; If p < 0.05 indicates significant, then x j0 variable.

Images