Frequency band-dividing electroencephalogram micro-state analysis method

Abstract

The invention discloses a sub-band electroencephalogram micro-state analysis method, and belongs to the field of neuroscience. According to the method, micro-state analysis is carried out after sub-band filtering, so that the problem that spatial differences of brain activities of different rhythm information of electroencephalogram signals are ignored in a traditional electroencephalogram micro-state analysis method is solved; by comparing micro-state transition probability matrixes, average durations and occurrence frequencies of FOG and non-FOG states under five frequency bands of Delta (1-4Hz), Theta (4-8Hz), Alpha (8-14Hz), Beta (12-31Hz) and Gamma (31-51Hz), a more scientific basis is provided for further revealing different pathogenesis of the FOG, and a new difference degree analysis function is further provided for related index analysis, so that the method has the advantages of being simple in structure, convenient to operate and high in practicability. And the analysis result is more accurate.

Description

Technical Field

[0001] The invention relates to a frequency-band electroencephalogram microstate analysis method and belongs to the field of neuroscience. Background Art

[0002] Parkinson's disease (PD) is a common neurodegenerative disease that affects more than 10 million people worldwide. One of the most distressing symptoms is freezing of gait (FOG), which is characterized by the cessation of movement when patients start or walk. About 50% of PD patients and 80% of advanced PD patients have FOG, which seriously affects their quality of life. Due to the complex pathogenesis of freezing of gait, distinguishing different causes of freezing is of great significance for further effective intervention measures.

[0003] At present, the different causes of freezing of gait in Parkinson's disease can be attributed to the following two ways. One is to objectively monitor the patient's gait data, such as gait rhythm, swing time variation coefficient and swing angle range, through technical means such as wearable devices or sensors, and further determine the possible cause of FOG by analyzing these data; the other is to monitor brain electrical activity through electroencephalogram (EEG) or other electrophysiological techniques to identify changes in neural signals related to FOG, that is, to study the neural electrical activity of the brain in a resting state through the EEG microstate analysis method to explore the functional state and spatiotemporal state of the brain. This method can simultaneously extract spatial patterns and temporal features in EEG activities, providing a scientific basis for determining the cause of freezing of gait.

[0004] Traditional EEG microstate analysis methods usually use the full-band clustering method, that is, all frequency bands are mixed together for analysis. After the full-band signal is mixed with the amplitude and phase information of each frequency band to become a mixed signal, the amplitude and phase information of each frequency band under the sub-band analysis is lost, and further the microstate information of each frequency band signal is lost. This information is important information reflecting the specific brain activity in the frequency band, and may also be important information to reveal the difference in the statistical characteristics of EEG signals in the onset and non-onset states of FOG. Therefore, mixing all signals for full-band microstate analysis has a certain one-sidedness. Summary of the invention

[0005] In order to solve the problems existing in the prior art, the present invention proposes an innovative EEG microstate analysis method, which extracts the spatial and temporal information of EEG activities of different rhythms by performing frequency band filtering on EEG data and then performing microstate analysis. The specific method is as follows:

[0006] Step 1: Data segmentation and sampling;

[0007] The database used in this study is multimodal data collected from 9 FOG patients. The multimodal data set includes electroencephalogram (EEG), electromyography (EMG), acceleration (ACC) and skin charge (SC) data, which are collected using a 32-channel wireless MOVE system with a sampling rate of 1000Hz. FOG episodes are marked by professional doctors. In addition, the data is divided into 70% training set and 30% test set.

[0008] Step 2: Split-band filtering

[0009] Filtering: The EEG data was bandpass filtered into five frequency bands: delta (1-4Hz), theta (4-8Hz), alpha (8-14Hz), beta (12-31Hz) and gamma (31-51Hz). A total of 10 datasets were obtained: data_delta_train, data_delta_test, data_theta_train, data_theta_test, data_alpha_traint, data_alpha_test, data_beta_train, data_beta_test, data_gamma_train and data_gamma_test.

[0010] Step 3: Spatial cluster analysis

[0011] The k-means clustering algorithm is used to divide the 5 training data sets into 4 microstates S0, S1, S2, and S3. A total of 5 models are obtained.

[0012] Step 4: Microstate Allocation

[0013] Use the predict function of pycrostates to assign microstates to EEG signals, that is, use these five models to predict the corresponding training sets and data sets to obtain time series.

[0014] The specific calculation method is: according to the 4 (or other number) microstate templates that have been obtained so far, the function calculates the similarity between the input data and each microstate template at each time point. This is usually achieved by calculating the Pearson correlation coefficient, Euclidean distance or cosine similarity between the data and the microstate template. The higher the similarity, the closer the EEG pattern at that time point is to a specific microstate. Finally, ModKmeans adopts the winner-takes-all method, that is, each moment is assigned and only one microstate with the highest similarity to it. Finally, we get time_series_delta_train, time_series_delta_test, time_series_theta_train, time_series_theta_test, time_series_alpha_train, time_series_alpha_test, time_series_beta_train, time_series_beta_test, time_series_gamma_train and

[0015] time_series_gamma_test has a total of 10 time series.

[0016] Step 5: Statistical analysis

[0017] Use three indicators: transition probability, duration, and frequency of occurrence

[0018] (1) Transition probability matrix

[0019] Assuming that the current moment is in microstate Sa, the next moment may be S0, S1, S2, S3. Count all changes of Sa, and the final transition probability of Sa-Sb is:

[0020]

[0021] Calculate all transition probabilities in turn and draw the transition probability matrix.

[0022] (2) Average duration

[0023] We count the occurrence of the same microstate Sn for several consecutive moments as an occurrence period of Sn, and count the number of all occurrence periods of Sn. Then we can calculate the average duration of each microstate as:

[0024]

[0025] Where Sn is the nth identified microstate.

[0026] (3) Frequency of occurrence

[0027] Calculate and record the frequency of occurrence of each microstate:

[0028]

[0029] Then, a designed algorithm (similar to calculating the difference in proportions, such as a 25% difference between 5 and 4) is used to analyze the difference between the diseased state and the non-disease state (significance evaluation), and to analyze the difference between the results of the training set and the test set (consistency evaluation). Ultimately, the higher the significance and the higher the consistency, the better the effect.

[0030] Step 6: Significance and consistency assessment

[0031] The significance and consistency of EEG activities in different states are to compare the differences between two sets of data. The difference is that the significance directly takes the difference, while the consistency takes the similarity (1-difference). The general difference evaluation can be measured by indicators such as Euclidean distance, Pearson correlation coefficient, and cosine similarity. Here we also use the Pearson correlation coefficient to measure: get two sets of results A and B, convert them into one-dimensional vectors, and calculate the Pearson correlation coefficient of the two vectors; use the Pearson correlation coefficient to measure the linear correlation of the two sets of results. The Pearson correlation coefficient of data with linear correlation is 1. We use result A as the horizontal coordinate and result B as the vertical coordinate to draw a plane graph.

[0032] like Figure 3 As shown in the figure, the Pearson correlation coefficients of the three groups of data are all 1, but only the blue group of data is truly consistent. Therefore, the Pearson coefficient does not necessarily measure similarity and can only be used as a reference.

[0033] Here, the present application has designed a more appropriate algorithm to measure similarity, which is required to meet the following properties:

[0034] (1) The basic idea is to use a ratio to measure the difference, such as 7 is +40% different from 5. The result is a percentage, not a distance like the Euclidean distance, which cannot directly tell whether the difference is high or low.

[0035] (2) The actual degree of difference increases monotonically with the difference index value, and the data index values ​​of different degrees of difference cannot be the same.

[0036] (3) When measuring the difference between two sets of vectors, the influence of a single component on the overall difference is limited. Because of the influence of uncontrollable factors such as noise, which may cause a large difference in a single component, it cannot be concluded that the two vectors are very different. For example, A = (0, 1, 2, 3, 4, 5), B = (10, 1, 2, 3, 4, 5), it should be said that the difference between A and B is not large.

[0037] Finally, the consistency evaluation algorithm is summarized as follows:

[0038] S1: First, the input data A and B are converted into one-dimensional vectors, where A is used as the standard vector, and then the difference between B and A is calculated.

[0039] S2: Take the absolute value of each component of A to get A'

[0040] S3: Calculate the mean value a of A'

[0041] S4: Calculate the difference vector C of the two input vectors and take the absolute value of each component of C.

[0042] S5: Each component of C is divided by a to get C', which is the ratio of the difference of the algebraic value of each component to the mean of the standard vector A'. For example, A = (0, 1, 2, 3, 4), the mean a = 2. The second component of B is 2, |2-1| / 2 = 50%, which means that the difference of this component is 50%. The contribution of the difference of this component to the final difference can be calculated later.

[0043] S6: compress the difference. This step is to achieve the above-mentioned goal (3) and suppress the excessive difference of a single component. The following function is used for the difference of each component:

[0044]

[0045] like Figure 4 As shown, (0,1) is approximately y = x, (1,∞) is approximately y = 1, and C' is processed by this function to get C"

[0046] S7: C" is divided by a to obtain C"', and each component of C"' is the contribution of the difference of the component at that location to the final difference.

[0047] S8: Sum the components of vector C'' to obtain the final difference.

[0048] This algorithm is used to calculate the difference, and then (1-difference) is used to get the similarity. A similarity of 0-30% means there is similarity, 30%-70% means good similarity, and 70%-100% means very high similarity.

[0049] The beneficial effects of the present invention are:

[0050] Through the frequency-band microstate analysis method, the problem of ignoring the diversity of EEG signal spectrum in traditional microstate analysis methods is solved, and the full utilization of the entire frequency band is achieved. By comparing the microstate transition probability matrix, average duration and occurrence frequency of FOG and non-FOG states in five frequency bands: Delta (1-4Hz), Theta (4-8Hz), Alpha (8-14Hz), Beta (12-31Hz) and Gamma (31-51Hz), a deeper understanding of the neural mechanism of FOG is provided, which further reveals the different pathogenesis of FOG and provides a basis for precise intervention. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0052] Figure 1 This is a typical EEG power spectrum density map calculated using subject 008’s data;

[0053] Figure 2 This is the EEG frequency band clustering result diagram;

[0054] Figure 3 It is a data graph with three groups of Pearson correlation coefficients of 1;

[0055] Figure 4 is the component difference map;

[0056] Figure 5 It is the transition probability train-test comparison chart;

[0057] Figure 6 It is the average duration train-test comparison chart;

[0058] Figure 7 It is a train-test comparison chart of average occurrence frequency. DETAILED DESCRIPTION

[0059] In order to make the objectives, technical solutions and advantages of the present invention more clear, the embodiments of the present invention will be further described in detail below with reference to the accompanying drawings.

[0060] First, the basic theory of the present invention is introduced as follows:

[0061] 1. EEG signal rhythm information;

[0062] EEG signal rhythm information refers to electrical signals with specific frequency and amplitude generated by the synchronous activities of brain neuron groups, which can reflect the electrical activity patterns of the brain in different states.

[0063] EEG rhythm information is usually divided into different frequency bands, including delta waves (1-4Hz), theta waves (4-8Hz), alpha waves (8-14Hz), beta waves (14-31Hz) and gamma waves (31-51Hz); different waves have different characteristics and represent different meanings:

[0064] ①Delta wave (1-4Hz);

[0065] Characteristics: Delta waves are the slowest brain waves and are usually associated with deep sleep and slow-wave sleep. In the waking state, high activity of delta waves usually indicates abnormalities in the brain, such as brain damage or certain pathological conditions.

[0066] Significance: Analysis of delta waves can provide information about deep neural activity, and changes in delta waves may reflect the brain's deep regulatory mechanisms for motor control and postural stability.

[0067] ②Theta wave (4-8Hz);

[0068] Characteristics: Theta waves are often associated with light sleep, meditation, memory formation, and information processing. During wakefulness, high activity in theta waves may be associated with cognitive tasks and emotional processing.

[0069] Significance: Changes in theta waves can reveal changes in cognitive load and emotional state. Analysis of theta waves may help understand how cognitive and emotional factors affect the occurrence of gait freezing.

[0070] ③alpha wave (8-14Hz);

[0071] Characteristics: Alpha waves occur primarily in the occipital lobe when resting with eyes closed, and are associated with states of relaxation and alertness. A decrease in alpha waves usually means increased brain activity, such as during task execution or under stress.

[0072] Significance: Alpha wave activity can reflect the brain's alertness and relaxation states. Analysis of alpha waves helps to understand the changes in patients' attention and relaxation states before and after FOG occurs.

[0073] ④ beta wave (14-31Hz);

[0074] Characteristics: Beta waves are associated with wakefulness, alertness, concentration, and active thinking. High frequencies of beta waves may be associated with anxiety, stress, and overexcitement.

[0075] Significance: Beta wave activity can reflect the brain's alertness and motor readiness. Its analysis can provide information about the patient's tension and motor readiness before and after FOG occurs.

[0076] ⑤gamma wave (31-51Hz);

[0077] Characteristics: Gamma waves are associated with higher cognitive functions such as perception, consciousness, memory, and information processing. Their activity may reflect the synchronization process between different brain regions.

[0078] Significance: Gamma wave activity can reveal the brain's advanced functions in complex tasks and cognitive processes. It may help understand changes in the brain during motor execution and cognitive integration.

[0079] 2. EEG microstate template;

[0080] Microstate templates (also called microstate maps or microstate patterns) refer to a series of short-lived, stable scalp potential field topologies extracted from electroencephalogram (EEG) data by cluster analysis in the resting state of the brain. These templates represent specific brain networks activated in a specific time period and are divided into microstates S0, S1, S2, and S3, where:

[0081] Microstate S0 is often associated with language / speech processing. It is identified in the alpha band of the EEG and is associated with activity in certain specific areas of the brain, such as the parietal and temporal lobes.

[0082] Microstate S1 is associated with visual processing. It shows higher occurrence and duration during visual tasks and exhibits specific activity patterns in the visual network.

[0083] Microstate S2 is related to subjective interoceptive-autonomic processing. It involves brain areas related to emotions and the autonomic nervous system, such as the limbic system.

[0084] Microstate S3 is associated with attention redirection. It shows higher activity during tasks that require attention shifting and is associated with the frontoparietal network of the brain.

[0085] The present invention uses a frequency-band clustering analysis method to analyze the connection status of different brain rhythms, which provides a possibility for further revealing different working states of the brain and thus providing a more detailed explanation of the different mechanisms of FOG. The specific introduction is as follows:

[0086] Embodiment 1:

[0087] This embodiment provides a frequency-divided band EEG microstate analysis method, the method comprising:

[0088] Step 1: Data segmentation and collection

[0089] The database used in the present invention is the multimodal data collected from 9 FOG patients. The multimodal data set includes electroencephalogram (EEG), which is collected using a 32-channel wireless MOVE system with a sampling rate of 1000Hz. FOG attacks are marked by professional doctors. This data set is subsequently used for model training and verification. The data in the data set is divided into 70% training set and 30% test set.

[0090] Step 2: Split-band filtering

[0091] Filtering: The EEG data in the dataset is divided into five frequency bands through bandpass filtering: delta (1-4Hz), theta (4-8Hz), alpha (8-14Hz), beta (12-31Hz) and gamma (31-51Hz). A total of 10 datasets are obtained: data_delta_train, data_delta_test; data_theta_train, data_theta_test; data_alpha_traint, data_alpha_test; data_beta_train, data_beta_test; data_gamma_train, data_gamma_test.

[0092] Step 3: Spatial cluster analysis

[0093] The k-means clustering algorithm is used to divide the five training data sets into four microstates S0, S1, S2, and S3, and five models for delta, theta, alpha, beta, and gamma are obtained.

[0094] Step 4: Microstate Allocation

[0095] Use the predict function of pycrostates to assign microstates to the EEG signals, that is, use the five models obtained in step 3 to predict the corresponding training set and test set to obtain the time series.

[0096] The specific calculation method is: based on the 4 (or other number) microstate templates that have been obtained so far, the function calculates the similarity between the input data and each microstate template at each time point.

[0097] Here, similarity is usually calculated by calculating the Pearson correlation coefficient, Euclidean distance or cosine similarity between the data and the microstate template. The higher the similarity, the closer the EEG pattern at that time point is to a specific microstate. Finally, ModKmeans adopts the winner-takes-all method, that is, each moment is assigned and only one microstate with the highest similarity to it is assigned. Finally, 10 time series are obtained: time_series_delta_train, time_series_delta_test; time_series_theta_train, time_series_theta_test; time_series_alpha_train, time_series_alpha_test; time_series_beta_train, time_series_beta_test; time_series_gamma_train and time_series_gamma_test.

[0098] Step 5: Statistical analysis

[0099] Use three indicators: transition probability, duration, and frequency of occurrence

[0100] (1) Transition probability matrix

[0101] Assuming that the current moment is in microstate Sa, the next moment may be S0, S1, S2, S3. Count all changes of Sa, and the final transition probability of Sa-Sb is:

[0102]

[0103] Calculate all transition probabilities in turn and draw the transition probability matrix;.

[0104] (2) Average duration

[0105] We count the occurrence of the same microstate Sn for several consecutive moments as an occurrence period of Sn, and count the number of all occurrence periods of Sn. Then we can calculate the average duration of each microstate as:

[0106]

[0107] Where Sn is the nth identified microstate.

[0108] (3) Frequency of occurrence

[0109] Calculate and record the occurrence frequency of each microstate.

[0110]

[0111] Then, a designed algorithm (similar to calculating the difference in proportions, such as a 25% difference between 5 and 4) is used to analyze the difference between the diseased state and the non-disease state (significance evaluation), and to analyze the difference between the results of the training set and the test set (consistency evaluation). Ultimately, the higher the significance and the higher the consistency, the better the effect.

[0112] Step 6: Significance Assessment and Consistency Assessment

[0113] Calculate the significance of each indicator of each microstate, where each indicator is the transition probability, duration, and occurrence frequency calculated in step 5 above;

[0114] S1: First, the input data A and B are converted into one-dimensional vectors, where A is used as the standard vector, and then the difference between B and A is calculated;

[0115] S2: Take the absolute value of each component of A to obtain A';

[0116] S3: Calculate the mean a of A';

[0117] S4: Calculate the difference vector C of the two input vectors, and take the absolute value of each component of C;

[0118] S5: Each component of C is divided by a to obtain C', which is the ratio of the difference in the algebraic value of each component compared to the mean of the standard vector A'.

[0119] For example, A=(0, 1, 2, 3, 4), mean a=2. The second component of B is 2, |2-1| / 2=50%, which means that the difference of this component is 50%. The contribution of the difference of this component to the final difference can be calculated later;

[0120] S6: compress the difference. This step is to achieve the above-mentioned goal (3) and suppress the excessive difference of a single component. The following function is used for the difference of each component:

[0121]

[0122] like Figure 4 As shown, (0,1) is approximately y = x, (1,∞) is approximately y = 1, and C' is processed by this function to obtain C".

[0123] S7: C" is divided by a to obtain C"', and each component of C"' is the contribution of the difference of the component at that location to the final difference.

[0124] S8: Sum the components of vector C'' to obtain the final difference.

[0125] For the transition probability matrix of each microstate, A and B are four-component vectors composed of the transition probabilities of a certain microstate to each microstate in the non-disease state and the diseased state, respectively, and B is the corresponding vector corresponding to the diseased state; for example, when calculating the significance of the transition probability matrix indicator of a certain microstate (such as S0), A is a four-component vector composed of the relevant parts of the transition probability matrix S0 in the non-disease state (P(S0→S0), P(S0→S1), P(S0→S2), P(S0→S3)), and B is the corresponding vector of the diseased state of S0.

[0126] For the average duration of each microstate, A and B are the average duration of a microstate in the non-pathological state and the pathological state, respectively (1-component vector);

[0127] For the occurrence frequency of each microstate, A and B are the occurrence frequencies of a microstate in the non-pathological state and the pathological state, respectively (1-component vector);

[0128] In the process of analyzing the correctness of the method, the consistency of all indicators of each microstate in two repeated experiments was also calculated. This step is not required in the normal process. Taking S0 as an example, A is the S0-related part of the transition probability difference matrix between the diseased state and the non-disease state in the first experiment, that is: (P(S0→S0), P(S0→S1), P(S0→S2), P(S0→S3))(disease)-(P(S0→S0), P(S0→S1), P(S0→S2), P(S0→S3))(non-disease), and B is the corresponding vector of the second experiment.

[0129] The data obtained by the above method are shown in the following table:

[0130] Table 1: Average discrimination between disease data and non-disease data based on an indicator of a microstate in a certain EEG frequency band of a subject

[0131]

[0132] Table 2 Average discrimination between disease data and non-disease data based on microstate transition probability matrix for a certain EEG frequency band of a subject

[0133]

[0134] Table 3 Average discrimination between disease data and non-disease data based on the average duration of microstates in a certain EEG frequency band of a subject

[0135]

[0136]

[0137] Table 4 Average discrimination between disease data and non-disease data based on the frequency of microstates in a certain EEG frequency band of a subject

[0138]

[0139] According to Table 1, each data in Table 1 is the average change amplitude of three indicators (transition probability, average duration, average frequency of occurrence). We can measure the change amplitude of EEG signals according to the change amplitude of the indicators. It can be seen that EEG signals in different bands have different change amplitudes in the diseased state and the non-disease state. Among them, delta and gamma have the largest change amplitudes, which are significantly higher than the change amplitudes of the other three frequency bands.

[0140] According to Table 2, the microstate transition probability matrix indicator generally has a low degree of differentiation between disease data and non-disease data. If you want to distinguish between disease data and non-disease data, the transition probability matrix is ​​not a good choice. We can even preliminarily conclude that the transition probability between a person's EEG microstates may be stable for a considerable period of time and will not be affected by the onset of FOG.

[0141] According to Table 3, the average duration of microstates also generally has a low degree of differentiation between onset data and non-onset data. However, the delta frequency band indicator can barely distinguish between onset data and non-onset data.

[0142] According to Table 4, the average frequency of occurrence of microstates has a high degree of differentiation between disease data and non-disease data. Among them, the average frequency of occurrence index of delta and gamma frequency bands has the largest change range.

[0143] In summary, the analysis of delta and gamma frequency band indicators can play a certain auxiliary role in determining the occurrence of FOG. The average duration of the delta band microstate, the average frequency of occurrence of the delta band, and the average frequency of occurrence of the gamma band have a high degree of differentiation between the diseased state and the non-disease state data.

[0144] Embodiment 2:

[0145] In order to verify the reproducibility of the method proposed in this invention, the same method was used to conduct two experiments on 008 at different times, one as a training set and one as a test set. The index values ​​of the training set are plotted on the horizontal axis and the index values ​​of the test set are plotted on the vertical axis. Figure 5-7 As shown, the Pearson correlation coefficient is calculated; it can be seen from the figure that the person correlation coefficients are all above 0.75, which is very high.

[0146] In addition, the difference analysis function designed by this application is used to calculate and analyze:

[0147]

[0148] (0-30% consistency, 30%-70% high consistency, 70%-100% very high consistency)

[0149] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A frequency-band EEG microstate analysis method, characterized in that: The method comprises: Step 1: Collect multimodal EEG data of FOG patients, and add labels 0 or 1 to the collected multimodal EEG data to indicate FOG attack or no attack, respectively. Use the labeled data to construct a data set and divide it into training set and test set; Step 2: Perform frequency band filtering on the collected multimodal EEG data to obtain frequency band signals corresponding to delta waves, theta waves, alpha waves, beta waves, and gamma waves; Step 3: Use the spatial cluster analysis method to perform cluster analysis on the frequency band signals corresponding to delta waves, theta waves, alpha waves, beta waves, and gamma waves; Step 4: Match the microstate template according to the cluster analysis results obtained in step 3; Step 5: Based on the matching results of step 4, statistically analyze the transition probability matrix, average duration, and occurrence frequency of each microstate; Step 6: Perform significance assessment and consistency assessment based on the analysis results of step 5.

2. The method according to claim 1, characterized in that The step 2 comprises: The collected multimodal EEG data were divided into five frequency bands through bandpass filtering. The frequency bands corresponding to delta waves, theta waves, alpha waves, beta waves, and gamma waves were 1-4 Hz, 4-8 Hz, 8-14 Hz, 12-31 Hz, and 31-51 Hz, respectively. The training set and test set are divided into training sets and test sets for delta waves, theta waves, alpha waves, beta waves, and gamma waves respectively.

3. The method according to claim 2, characterized in that The step 3 comprises: The k-means clustering algorithm is used to divide the five training set data obtained in step 2 into four microstates S0, S1, S2 and S3, and five k-means clustering models for delta, theta, alpha, beta and gamma are obtained respectively.

4. The method according to claim 3, characterized in that The step 4 comprises: The similarity between the input data in the five training sets and the test set and each microstate template is calculated at each time point to obtain the time series corresponding to the training sets and the test sets predicted by the five models.

5. The method according to claim 4, characterized in that The transition probability matrix P in step 5 Sa→Sb , average duration t Sn_average , frequency of occurrence f Sn The calculation formula is as follows: Among them, Sb, Sb and Sn each represent one of the four microstates S0, S1, S2 and S3.

6. The method according to claim 5, characterized in that The step 6 comprises: S1: First, the input data A and B are converted into one-dimensional vectors, where A is used as the standard vector, and then the difference between B and A is calculated; For the transition probability matrix of each microstate, A and B are vectors with four components composed of the transition probabilities of a microstate to each microstate in the non-disease state and the disease state, respectively, and B is the corresponding vector corresponding to the disease state; For the average duration of each microstate, A and B are the average duration of a microstate in the non-pathological state and the pathological state, respectively; Regarding the frequency of occurrence of each microstate, A and B are the frequency of occurrence of a microstate in the non-pathological state and the pathological state, respectively; S2: Take the absolute value of each component of vector A to obtain A'; S3: Calculate the mean a of A'; S4: Calculate the difference vector C of the two input vectors, and take the absolute value of each component of C; S5: Each component of C is divided by a to obtain C', that is, the ratio of the difference of the algebraic value of each component to the mean of the standard vector A'; so that the contribution of the difference of the component to the final difference can be calculated later; S6: compress the difference; use the following function for each component difference: (0,1) is approximately y = x, (1,∞) is approximately y = 1, and C' is processed by this function to obtain C"; S7: C" divided by a is C"', and each component of C"' is the contribution of the difference of the component at that location to the final difference; S8: sum the components of vector C'' to obtain the final difference; After obtaining the final difference, the similarity is obtained by using 1-difference; among them, 0-30% similarity means there is similarity, 30%-70% means good similarity, and 70%-100% means extremely high similarity.

7. The method according to claim 6, characterized in that The multimodal data set in step 1 includes EEG data, and the ratio of the training set to the test set is 7:

3.

8. The method according to claim 7, characterized in that In step 4, the predict function of pycrostates is used to assign microstates to the EEG signal.

Images