Autism identification method based on electroencephalogram cross-band coupling

Abstract

The invention discloses an autism identification method based on electroencephalogram cross-band coupling, and relates to the technical field of neuroscience and medical science, and the method comprises the steps: obtaining original electroencephalogram data of an autism patient and a normal person, carrying out the preprocessing, setting labels for the autism patient and the normal person, and carrying out the identification of the autism patient and the normal person; constructing a training sample set for the processed electroencephalogram data and labels; calculating cross-band coupling characteristics of the preprocessed electroencephalogram data; constructing an autism recognition model based on electroencephalogram cross-band coupling, and constructing a classifier by using the cross-band coupling features and labels of the training set and using a deep learning method; and putting the electroencephalogram data features with unknown tags into the classifier to obtain tags corresponding to the electroencephalogram data features, thereby completing the process of identifying autism patients. The cross-band coupling feature of the electroencephalogram reflects the balance state of nerve excitation suppression in the brain, and the autism classification and recognition model constructed based on the balance state has a better classification effect.

Description

Technical Field

[0001] The present invention relates to the technical field of neuroscience and medicine, and in particular to a method for constructing an autism EEG recognition model based on cross-band coupling parameters. Background Art

[0002] Autism spectrum disorder is a complex neurodevelopmental disorder. Studies in recent years have found that excitation-inhibition imbalance in the central nervous system may be one of its important pathogenesis. Under normal circumstances, the brain's excitatory transmitter system (represented by glutamate) and inhibitory transmitter system (represented by γ-aminobutyric acid) work together to maintain the balance of neural circuits. However, in patients with autism, this balance is broken, causing the excitation level to exceed the inhibition level or vice versa, which in turn affects the activity of the neural circuit and may cause core symptoms of autism such as social disorders, restricted interests and repetitive stereotyped behaviors. Excitation-inhibition imbalance, as a neural characteristic of autistic patients, provides a theoretical basis for distinguishing normal people from autistic patients.

[0003] Cross-frequency coupling (CFC) in EEG is an important concept in neuroscience research. It reveals the dynamic interaction between neural oscillations in different frequency bands. This cross-band coupling phenomenon is closely related to the balance of neural excitation and inhibition. Specifically, the balance between neural excitation and inhibition can affect the generation and transmission of neural oscillations, and thus affect the pattern and intensity of cross-band coupling. At the same time, changes in cross-band coupling may also reflect the disorder of neural excitation-inhibition balance. Therefore, the cross-band coupling characteristics of EEG reflect the balance of neural excitation and inhibition in the brain. The autism classification and recognition model constructed based on this has better classification effect, and is expected to provide a new perspective for understanding the pathogenesis of autism and promoting AI-assisted clinical intelligent diagnosis. Summary of the Invention

[0004] Purpose of the invention: In response to the problems existing in the prior art, the purpose of the present invention is to provide an autism identification method based on cross-band EEG coupling, which uses the cross-band coupling characteristics of EEG to reflect the balanced state of neural excitation and inhibition inside the brain. The autism classification and identification model constructed based on this has better classification effect, providing a new perspective for understanding the pathogenesis of autism and promoting AI-assisted clinical intelligent diagnosis.

[0005] In order to solve the above technical problems, the present invention adopts the following technical solutions:

[0006] A method for identifying autism based on EEG cross-band coupling, comprising the following steps:

[0007] S1: Obtain the resting-state raw EEG data of autistic patients and normal people, and mark the data of autistic patients as 1 and the data of normal people as 0;

[0008] S2: Preprocess the original EEG data and use the preprocessed EEG data and labels to construct a training

[0009] training sample set;

[0010] S3: Calculate the cross-band coupling characteristics of the preprocessed EEG data;

[0011] S4: Construct an autism recognition model based on EEG cross-band coupling, use the cross-band coupling features and labels of the training set to build a classifier using machine learning or deep learning methods; put the EEG data features with unknown labels into the classifier, obtain the labels corresponding to the EEG data features, and complete the identification process of autistic patients.

[0012] Furthermore, the preprocessing of EEG data in step S2 specifically includes:

[0013] S2.1: Downsample, filter, notch process, and remove artifacts such as electrooculogram (EOG)

[0014] S2.2: Detect bad sectors on data and interpolate bad sectors

[0015] S2.3: Split data into segments according to a certain length

[0016] Furthermore, the specific steps of calculating the characteristics of cross-band coupling in step S3 are as follows:

[0017] S3.1.1: The time domain signal is represented as X = (x1, x2, ..., x N ), the time evolution of the activity power at the signal X frequency υ is called Y υ To represent the variation of power over time, we estimate the discrete Fourier transform of successive segments extracted from X using a sliding window of M samples:

[0018]

[0019] in is the active power at frequency υ and sampling point n, Represents the Hanning taper. The length of M is five periods relative to the frequency υ, that is, M = 5·Fs / υ;

[0020] S3.1.2: Next, X and Y υ is divided into a set of S consecutive segments with l overlapping time points; X s represents the time series of segment s; Y υ,srepresents the power time evolution of the segment s, that is, the power change sequence of the signal segment in the time dimension; we use the standard fast Fourier transform to calculate their spectrum:

[0021]

[0022] where n FFT =256 represents the frequency resolution of the fast Fourier transform; the center frequency of the FFT transform

[0023] S3.1.3: Next we calculate signal X and signal Y υ Complex coherence between the power envelopes of:

[0024]

[0025] The frequency is “*” indicates complex conjugate; C(υ,f) indicates signal X and signal Y υ The complex coherence between the power envelopes of at frequency point f.

[0026] S3.1.4: Phase slope index estimates the slope of the phase difference as a function of frequency within a specified frequency band, where the sign of the slope indicates the direction of the interaction; next calculate the phase amplitude coupling (PAC), where signal X and signal Y υ At the frequency (υ,f j ) is the phase slope index of the power envelope at:

[0027]

[0028] where f j represents the low frequency of signal X, Δf represents the frequency resolution, and Im represents the imaginary part. We use β = 2 Hz to represent the bandwidth for calculating the phase slope;

[0029] S3.1.5: Finally, calculate the low-frequency band B p Phase (θ or α) and high frequency band B a The cross-band coupling value PAC between the amplitudes of (β or γ), that is, all pairwise combinations (f k ,f l ) in the maximum phase slope index:

[0030]

[0031] Using the above phase amplitude coupling (PAC) calculation method, the cross-band coupling between the low-frequency and high-frequency components of the EEG signal collected in a single channel and the cross-band coupling of signals between different channels are calculated. The calculated cross-frequency domain coupling characteristics of the scalp channel are used as EEG features for classification.

[0032] Furthermore, the step S2 further includes:

[0033] Preprocessing of raw EEG data also involves tracing the source of the raw data, obtaining the time-domain EEG signals in the source space, identifying the specific brain areas corresponding to scalp physiological activity through EEG signal source localization technology, and reconstructing the time series of the source signals at the cortical level. This is achieved using the BrainStorm toolkit based on the MATLAB platform and the standardized low-resolution electromagnetic tomography method (sLORETA). Specifically, the mapping relationship between the scalp electrode space and the source space at a certain sampling point can be expressed by Maxwell's equations as follows:

[0034] Φ=MX+ε

[0035] in and They represent the scalp potential vectors recorded from C electrodes (channels) and the source signal vectors composed of V brain regions of interest (ROIs); is the lead field matrix, represents the noise vector. Next, we solve the following inverse problem:

[0036]

[0037] Under the conditions of given M, Φ and regularization parameter α≥0, the explicit solution of the above optimization problem is:

[0038]

[0039] The superscript '+' indicates the Moore-Penrose pseudoinverse. is the center matrix. Normalization is performed to obtain the sLORETA estimation value of each ROI. In this study, we divided the cortex into multiple source space ROIs based on the brain area division map, and extracted the time series of all ROIs by repeating the above process to finally form a reconstructed source signal matrix. where N R represents the number of ROIs, and N represents the number of sampling points.

[0040] The present invention can locate the cortical signal source more accurately after tracing the source. In EEG research, the original sensor-level data is only the potential on the scalp, and tracing the source can map these signals to specific ROIs (Regions of Interest) in the cerebral cortex, thereby analyzing the interaction between brain regions. Calculating cross-frequency coupling after tracing the source can reduce the volume conduction effect and improve the accuracy of cross-frequency coupling analysis between ROIs.

[0041] Furthermore, the specific steps for calculating the characteristics of cross-band coupling described in step S3 are as follows:

[0042] S3.1.1: The time domain signal is represented as X = (x1, x2, ..., x N ), the time evolution of the activity power at the signal X frequency υ is called Y υ To represent the variation of power over time, we estimate the discrete Fourier transform of successive segments extracted from X using a sliding window of M samples:

[0043]

[0044] in is the active power at frequency υ and sampling point n, Represents the Hanning taper. The length of M is five periods relative to the frequency υ, that is, M = 5·Fs / υ;

[0045] S3.1.2: Next, X and Y υ is divided into a set of S consecutive segments with l overlapping time points; X s represents the time series of segment s; Y υ,s represents the power time evolution of the segment s, that is, the power change sequence of the signal segment in the time dimension; we use the standard fast Fourier transform to calculate their spectrum:

[0046]

[0047] where n FFT =256 represents the frequency resolution of the fast Fourier transform;

[0048] S3.1.3: Next we calculate signal X and signal Y υ Complex coherence between the power envelopes of:

[0049]

[0050] The frequency is “*” indicates complex conjugation;

[0051] S3.1.4: Phase slope index estimates the slope of the phase difference as a function of frequency within a specified frequency band, where the sign of the slope indicates the direction of the interaction; next calculate the phase amplitude coupling (PAC), where signal X and signal Y υ At the frequency (υ,f j ) is the phase slope index of the power envelope at:

[0052]

[0053] where f j represents the low frequency of signal X, Δf represents the frequency resolution, and Im represents the imaginary part. We use β = 2 Hz to represent the bandwidth for calculating the phase slope;

[0054] S3.1.5: Finally, calculate the low-frequency band B p Phase (θ or α) and high frequency band B a The cross-band coupling value PAC between the amplitudes of (β or γ), that is, all pairwise combinations (f k ,f l ) in the maximum phase slope index:

[0055]

[0056] Using the above phase amplitude coupling (PAC) calculation method, the cross-band coupling between the low-frequency and high-frequency components of the EEG signal collected in a single channel and the cross-band coupling of signals between different channels are calculated. The calculated cross-frequency domain coupling characteristics of the scalp channel are used as EEG features for classification.

[0057] Preferably, the characteristic step of calculating the cross-band coupling in step S3 may also be:

[0058] S3.2.1: The time domain signal is represented as X = (x1, x2, ..., x N ), the time evolution of the activity power at the signal X frequency υ is called Y υ To represent the variation of power over time, we estimate the discrete Fourier transform of successive segments extracted from X using a sliding window of M samples:

[0059]

[0060] in is the active power at frequency υ and sampling point n, The length of M is five periods of the frequency υ, that is, M = 5·Fs / υ.

[0061] S3.2.2: Next, X and Y υ is divided into a set of S consecutive segments with l overlapping time points. s represents the time series of segment s; Y υ,s represents the power time evolution of the segment s, that is, the power change sequence of the signal segment in the time dimension. Our standard fast Fourier transform calculates their spectrum:

[0062]

[0063] where n FFT =256 represents the frequency resolution of the fast Fourier transform.

[0064] S3.2.3: Next we calculate signal X and signal Y υ Complex coherence between the power envelopes of:

[0065]

[0066] The frequency is “*” indicates complex conjugation.

[0067] S3.2.4: Phase Slope Index estimates the slope of the phase difference as a function of frequency within a specified frequency band, where the sign of the slope indicates the direction of the interaction. Quantification of Phase Amplitude Coupling (PAC) Using Jiang's method, we calculate Signal X and Signal Y υ At the frequency (υ,f j ) is the phase slope index of the power envelope at:

[0068]

[0069] where f j represents the low frequency of signal X, Δf represents the frequency resolution, and Im represents the imaginary part. We use β = 2 Hz to represent the bandwidth for calculating the phase slope.

[0070] S3.2.5: Finally, we calculate the cross-band coupling of ROI signals with different source spaces, and use the calculated cross-band coupling features of ROI signals with different source spaces as EEG features for classification. We calculate the low-frequency band B p Phase (θ or α) and high frequency band B a The cross-frequency coupling value PAC between the amplitudes of (β or γ), that is, all pairwise combinations (f k ,f l ) in the maximum phase slope index:

[0071]

[0072] In this invention, we use the signals corresponding to the traced ROI to calculate the cross-frequency domain coupling within the network and between networks to help extract the dynamic changes of different functional networks themselves and the changes in the coupling relationship between each other, which reflects the driving relationship and excitation-inhibition balance between different functional networks. The autism classification and recognition model constructed based on this has better classification effect.

[0073] Furthermore, in step S3.2.6, the cross-frequency coupling characteristics within and between source space networks can be calculated and used as features for classification. Based on the distribution of ROIs, the source signals within each ROI are averaged, and then mapped to the seven functional network systems proposed by Yeo based on the spatial distribution characteristics of multiple ROIs in the source space. The intra-network phase amplitude coupling (PAC) is defined as the mean PAC value of all ROIs within a specific functional network, and the calculation formula is as follows:

[0074]

[0075] The calculation of PAC (Phase-Amplitude Coupling) refers to the low-frequency band B p The phase and high frequency band B a The coupling between the amplitudes, Indicates a specific network net i The time course signal of the hth ROI (region of interest), H represents the total number of ROIs contained in the network;

[0076] At the same time, the cross-frequency coupling characteristics between source space networks are calculated. First, the connection strength is quantified by calculating the β-θ frequency band PAC: the phase of the low-frequency component (β band) of a certain ROI in one network is coupled with the amplitude of the high-frequency component (θ band) of a certain ROI in another network. Second, the average of all point-to-point connections between two networks with the same phase-amplitude coupling direction is taken as the inter-network connection strength based on network characteristics. The calculation formula is as follows:

[0077]

[0078] Where H1 and H2 represent the network i with net j The total number of ROIs included.

[0079] The cross-band coupling characteristics of time domain signals of different ROIs in the source space, and the cross-band coupling characteristics within and between networks can be used as features for classification.

[0080] Furthermore, step S4 includes:

[0081] Use a two-layer transformer network for modeling and classification. Use sliced data of 64×1000, where 64 is the number of channels and 1000 is the number of features. Put the data of the training set into the two-layer transformer model for learning. The steps are as follows:

[0082] S4.1 Position encoding: We use sine and cosine functions to encode position information.

[0083]

[0084] set up Represents the position encoding matrix, where pos is the position index, 0≤pos≤N c , 0≤pos≤N c ,2i or 2i+1 is the dimension index, 0≤2i≤2i+1≤N T As shown in the above formula, the position code is directly calculated from the data segment x;

[0085] To introduce temporal information, the position code is superimposed on the original input data O;

[0086]

[0087] Embed position encoded data It is then fed into a multi-head self-attention layer

[0088] S4.2 Multi-head self-attention layer; In the Transformer encoder, the multi-head attention module captures the dependencies between vectors at different positions; In the VAE encoder module, the input vector It is replicated into three copies, which are called query, key, and value, and are recorded as Q, K, and V respectively; here Q, K, and V are actually the same Subsequently, Q, K, and V are divided into H segments, and subsequent calculations are performed within each sub-segment. This module is called H heads.

[0089] In each head, the query, key, and value sub-segments are fed into three separate linear layers; the outputs of the linear layers are processed by an attention mechanism.

[0090]

[0091] Then, the attention calculation results from H heads are concatenated and mapped from high dimensions to the same dimension as the input through another fully connected layer. The same dimension is then processed by the dropout layer; the process can be expressed as:

[0092] MHA(Q,K,V)=[head0;…;head H-1 ]

[0093] head i =Attention(Q i ,K i ,V i )

[0094] where represents multi-head attention, and represents the query, key, and value obtained by linear transformation in the i-th head respectively;

[0095] After passing through two layers of transformers in S4.3, the data is fed into a linear layer to obtain the classification label.

[0096] The loss function of the network described in S4.4 is the cross-entropy loss function, which calculates the cross-entropy loss between the probability distribution of the network output and the target probability distribution; the optimizer used is the Adam optimizer; the evaluation metric is accuracy, that is, the ratio of correctly predicted samples to the total number of samples;

[0097] S4.5 Further, the network is trained. The specific process is as follows: the input data passes through each layer of the neural network in sequence, that is, after forward propagation, back propagation is performed. The back propagation process includes calculating the loss, that is, using the cross entropy loss function to calculate the cross entropy loss between the probability distribution of the network output and the target probability distribution, calculating the gradient, that is, calculating the gradient of the network weight matrix and weight coefficient according to the cross entropy loss function, and updating the parameters, that is, using the Adam optimizer to update the network weight matrix and weight coefficient according to the calculated gradient; training is completed after 200 rounds;

[0098] S4.6 For the data in the test set, put the data into the classification network model to obtain the label corresponding to the data to achieve the purpose of autism identification.

[0099] Compared with the prior art, the present invention has the following beneficial effects:

[0100] The present invention reflects the balance state of neural excitation and inhibition in the brain through the cross-band coupling characteristics of EEG. The autism classification and recognition model constructed based on this has better classification effect.

[0101] On the one hand, tracing the source can accurately locate the source of cortical signals. In EEG research, the original sensor-level data is only the potential on the scalp, and tracing the source can map these signals to specific ROIs (Regions of Interest) in the cerebral cortex, thereby analyzing the interaction between brain regions. Calculating cross-frequency coupling after tracing the source can reduce the volume conduction effect and improve the accuracy of cross-frequency coupling analysis between ROIs.

[0102] On the other hand, different networks are closely related to different functions, such as the visual network, sensorimotor network, dorsal attention network, ventral attention network, limbic network, frontoparietal control network, and default mode network. We use the signals corresponding to the traced ROI to calculate the cross-frequency domain coupling within and between networks to help extract the dynamic changes of different functional networks themselves and the changes in the coupling relationship between each other. This reflects the driving relationship and excitation-inhibition balance between different functional networks. The autism classification and recognition model constructed based on this has better classification effect.

[0103] In this study, the cross-frequency coupling method we used employed a phase slope index in the final calculation step. The phase slope index estimates the interaction between signals, and the ± of the slope reflects the direction of the interaction, making our cross-frequency coupling feature directional. The cross-frequency coupling features we calculated for inter-channel, intra-channel, inter-network, and intra-network networks not only include information on coupling strength but also on coupling direction, more accurately reflecting the excitation-inhibition balance within the brain. Therefore, autism classification and recognition models based on our cross-frequency coupling features are more effective. BRIEF DESCRIPTION OF THE DRAWINGS

[0104] Figure 1 This is an overall flow chart of an autism identification method based on EEG cross-band coupling according to the present invention;

[0105] Figure 2 Flowchart of cross-band coupling feature extraction in an embodiment of the present invention. DETAILED DESCRIPTION

[0106] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments.

[0107] Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the invention as claimed, but rather merely represents selected embodiments of the present invention. All other embodiments derived by persons of ordinary skill in the art based on the embodiments of the present invention without creative effort shall fall within the scope of protection of the present invention.

[0108] The present invention will be described in detail below with reference to the accompanying drawings:

[0109] Reference Figure 1 The present invention discloses a method for identifying autism based on EEG cross-band coupling, comprising the following steps:

[0110] S1: Obtain the resting-state raw EEG data of autistic patients and normal people, and mark the data of autistic patients as 1 and the data of normal people as 0;

[0111] S2: Preprocess the original EEG data and use the preprocessed data and labels to construct a training sample set;

[0112] In this embodiment, the preprocessing of data includes:

[0113] S2.1 downsamples, filters, notches, and removes artifacts such as electrooculogram (EOG)

[0114] Then perform bad track detection on the data and interpolate the bad tracks

[0115] S2.2: Split data into segments according to a certain length

[0116] S2.3: Trace the original data to obtain the time domain EEG signal of the source space

[0117] EEG signal source localization technology can be used to identify specific brain areas corresponding to scalp physiological activity and reconstruct the time series of source signals at the cortical level. This process is based on the BrainStorm toolkit on the MATLAB platform and is implemented using the standardized low-resolution electromagnetic tomography method (sLORETA). Specifically, the mapping relationship between the scalp electrode space and the source space at a certain sampling point can be expressed by Maxwell's equations as follows:

[0118] Φ=MX+ε

[0119] in and They represent the scalp potential vectors recorded from C electrodes (channels) and the source signal vectors composed of V brain regions of interest (ROIs); is the lead field matrix, represents the noise vector. Next, we solve the following inverse problem:

[0120]

[0121] Under the conditions of given M, Φ and regularization parameter α≥0, the explicit solution of the above optimization problem is:

[0122]

[0123] The superscript '+' indicates the Moore-Penrose pseudoinverse. is the center matrix. Normalization is performed to obtain the sLORETA estimation value of each ROI. In this study, we divided the cortex into multiple source space ROIs based on the brain area division map, and extracted the time series of all ROIs by repeating the above process to finally form a reconstructed source signal matrix. where N R represents the number of ROIs, and N represents the number of sampling points.

[0124] Reference Figure 2 , S3: Calculate the cross-band coupling characteristics of the preprocessed EEG data. The steps are as follows:

[0125] S3.1.1: The time domain signal is represented as X = (x1, x2, ..., x N ), the time evolution of the activity power at the signal X frequency υ is called Y υ To represent the variation of power over time, we estimate the discrete Fourier transform of successive segments extracted from X using a sliding window of M samples:

[0126]

[0127] in is the active power at frequency υ and sampling point n, Represents the Hanning taper. The length of M is five periods relative to the frequency υ, that is, M = 5·Fs / υ;

[0128] S3.1.2: Next, X and Y υ is divided into a set of S consecutive segments with l overlapping time points; X s represents the time series of segment s; Y υ,s represents the power time evolution of the segment s, that is, the power change sequence of the signal segment in the time dimension; we use the standard fast Fourier transform to calculate their spectrum:

[0129]

[0130] where n FFT =256 represents the frequency resolution of the fast Fourier transform; the center frequency of the FFT transform

[0131] S3.1.3: Next we calculate signal X and signal Y υ Complex coherence between the power envelopes of:

[0132]

[0133] The frequency is “*” indicates complex conjugate; C(υ,f) indicates signal X and signal Y υ The complex coherence between the power envelopes of at frequency point f.

[0134] S3.1.4: Phase slope index estimates the slope of the phase difference as a function of frequency within a specified frequency band, where the sign of the slope indicates the direction of the interaction; next calculate the phase amplitude coupling (PAC), where signal X and signal Y υ At the frequency (υ,f j ) is the phase slope index of the power envelope at:

[0135]

[0136] where f j represents the low frequency of signal X, Δf represents the frequency resolution, and Im represents the imaginary part. We use β = 2 Hz to represent the bandwidth for calculating the phase slope;

[0137] S3.1.5: Finally, calculate the low-frequency band B p Phase (θ or α) and high frequency band B a The cross-band coupling value PAC between the amplitudes of (β or γ), that is, all pairwise combinations (f k ,f l ) in the maximum phase slope index:

[0138]

[0139] Using the above phase amplitude coupling (PAC) calculation method, the cross-band coupling between the low-frequency and high-frequency components of the EEG signal collected in a single channel and the cross-band coupling of signals between different channels are calculated. The calculated cross-frequency domain coupling characteristics of the scalp channel are used as EEG features for classification.

[0140] Reference Figure 2 In this embodiment, the specific steps of calculating the cross-band coupling characteristics in step S3 can also be performed by tracing the original data.

[0141] S3.2.1: The time domain signal is represented as X = (x1, x2, ..., x N ), the time evolution of the activity power at the signal X frequency υ is called Y υ To represent the variation of power over time, we estimate the discrete Fourier transform of successive segments extracted from X using a sliding window of M samples:

[0142]

[0143] in is the active power at frequency υ and sampling point n, The length of M is five periods of the frequency υ, that is, M = 5·Fs / υ.

[0144] S3.2.2: Next, X and Y υ is divided into a set of S consecutive segments with l overlapping time points. s represents the time series of segment s; Y υ,s represents the power time evolution of the segment s, that is, the power change sequence of the signal segment in the time dimension. Our standard fast Fourier transform calculates their spectrum:

[0145]

[0146] where n FFT =256 represents the frequency resolution of the fast Fourier transform, the center frequency of the FFT transform

[0147] S3.2.3: Next we calculate signal X and signal Y υ Complex coherence between the power envelopes of:

[0148]

[0149] The frequency is “*” indicates complex conjugate, C(υ,f) indicates signal X and signal Y υ The complex coherence between the power envelopes of at frequency point f.

[0150] S3.2.4: Phase Slope Index estimates the slope of the phase difference as a function of frequency within a specified frequency band, where the sign of the slope indicates the direction of the interaction. Quantification of Phase Amplitude Coupling (PAC) Using Jiang's method, we calculate Signal X and Signal Y υ At the frequency (υ,f j ) is the phase slope index of the power envelope at:

[0151]

[0152] where f j represents the low frequency of signal X, Δf represents the frequency resolution, and Im represents the imaginary part. We use β = 2 Hz to represent the bandwidth for calculating the phase slope.

[0153] S3.2.5: Finally, we calculate the cross-band coupling of ROI signals with different source spaces, and use the calculated cross-band coupling features of ROI signals with different source spaces as EEG features for classification. We calculate the low-frequency band B p Phase (θ or α) and high frequency band B a The cross-frequency coupling value PAC between the amplitudes of (β or γ), that is, all pairwise combinations (f k ,f l ) in the maximum phase slope index:

[0154]

[0155] Step S3.2.6: Calculate the cross-frequency coupling characteristics within and between source-space networks. First, calculate the network distribution based on the ROIs, average the source signals within each ROI, and then map them to the functional network system proposed by Yeo based on the spatial distribution characteristics of multiple ROIs in the source space. The functional network system includes but is not limited to the visual network (VN), somatomotor network (SMN), dorsal attention network (DAN), ventral attention network (VAN), limbic network (LN), frontoparietal network (FPN), and default mode network (DMN). The intra-network phase-amplitude coupling (PAC) is defined as the mean PAC value of all ROIs within a specific functional network, calculated as follows:

[0156]

[0157] The calculation of PAC (Phase-Amplitude Coupling) refers to the low-frequency band B p The phase and high frequency band B a The coupling between the amplitudes, Indicates a specific network net i The time course signal of the hth ROI (region of interest), H represents the total number of ROIs contained in the network;

[0158] Then, the cross-frequency coupling characteristics between source space networks were calculated. First, the connection strength was quantified by calculating the β-θ frequency band PAC: the phase of the low-frequency component (β band) of a certain ROI in one network was coupled with the amplitude of the high-frequency component (θ band) of a certain ROI in another network. Second, the average of all point-to-point connections between two networks with the same phase-amplitude coupling direction was taken as the inter-network connection strength based on network characteristics. The calculation formula is as follows:

[0159]

[0160] Where H1 and H2 represent the network i with net j The total number of ROIs included.

[0161] The calculated cross-band coupling features within and between networks or the cross-band coupling features of ROI time domain signals, and the cross-band coupling features within and between networks can all be used as features for classification.

[0162] S4: Build an autism recognition model based on EEG cross-band coupling, and use the training set data and labels to build a classifier using machine learning or deep learning methods. Here is an example of using a two-layer transformer network for modeling and classification. For example, the slice data is 64 (number of channels)

[0163] ×1000 (number of features), we put the data of the training set into the two-layer transformer model for learning. The specific steps are:

[0164] S4.1 Position encoding: We use sine and cosine functions to encode position information.

[0165]

[0166] set up Represents the position encoding matrix, where pos is the position index, 0≤pos≤N c , 0≤pos≤N c ,2i or 2i+1 is the dimension index, 0≤2i≤2i+1≤N T As shown in the above formula, the position code is directly calculated from the data segment x.

[0167] In order to introduce temporal information, the position code is superimposed on the original input data O.

[0168]

[0169] Embed position encoded data It is then fed into the multi-head self-attention layer.

[0170] self-attention layer)

[0171] S4.2 Multi-head self-attention layer; In the Transformer encoder, the multi-head attention module captures the dependencies between vectors at different positions. In the VAE encoder module, the input vector is replicated into three copies.

[0172] Since these three copies have different meanings in computing, they are called query, key, and value, and are denoted as Q, K, and V. Here, Q, K, and V are actually the same.

[0173] Subsequently, Q, K, and V are divided into H segments, and subsequent calculations are performed in each sub-segment. This module is called H heads.

[0174] In each head, the query, key, and value sub-segments are fed into three separate linear layers. The output of the linear layer is processed by the attention mechanism.

[0175]

[0176] Then, the attention calculation results from H heads are concatenated and mapped from high dimensions to the same dimension as the input through another fully connected layer. The same dimension is then processed by the dropout layer. The process can be expressed as:

[0177] MHA(Q,K,V)=[head0;…;head H-1 ]

[0178] head i =Attention(Q i ,K i ,V i )

[0179] where represents multi-head attention, and represents the query, key, and value obtained by linear transformation in the i-th head.

[0180] S4.3 After going through two layers of transformer, the data is fed into a linear layer to obtain the classification label.

[0181] The loss function of the network described in S4.4 is the cross-entropy loss function, which calculates the cross-entropy loss between the probability distribution of the network output and the target probability distribution; the optimizer used is the Adam optimizer; the evaluation indicator is accuracy, that is, the proportion of correctly predicted samples to the total number of samples.

[0182] S4.5 further trains the network. The specific process is: the input data passes through each layer of the neural network in turn, that is, after forward propagation, back propagation is performed. The back propagation process includes calculating the loss, that is, using the cross entropy loss function to calculate the cross entropy loss between the probability distribution of the network output and the target probability distribution, calculating the gradient, that is, calculating the gradient of the network's weight matrix and weight coefficient according to the cross entropy loss function, updating the parameters, that is, using the Adam optimizer to update the network's weight matrix and weight coefficient according to the calculated gradient; training is completed after 200 rounds.

[0183] S4.6 For the data in the test set, put the data into the classification network model to obtain the label corresponding to the data to achieve the purpose of autism identification.

[0184] The above embodiments are preferred implementation modes of the present invention, but the implementation modes of the present invention are not limited to the above embodiments. Any other changes, modifications, substitutions, combinations, and simplifications that do not deviate from the spirit and principles of the present invention should be considered as equivalent replacement methods and are included in the scope of protection of the present invention.

Claims

1. A method for identifying autism based on EEG cross-band coupling, characterized in that: The following steps are involved: S1. Obtain the original EEG data of autistic patients and normal people in the resting state, and label them respectively; S2: Preprocess the raw EEG data and use the preprocessed EEG data and labels to construct a training sample set; S3: Calculate the cross-band coupling characteristics of the preprocessed EEG data; S4: Construct an autism recognition model based on EEG cross-band coupling. Use the cross-band coupling features and labels of the training set to build a classifier through machine learning or deep learning methods. Put the EEG data features with unknown labels into the classifier to obtain the labels corresponding to the EEG data features, thereby completing the identification process of autistic patients.

2. The autism identification method based on EEG cross-band coupling according to claim 1, characterized in that: The preprocessing of the raw EEG data in step S2 specifically includes: S2.

1. Downsampling, filtering, notching, and removing oculoscopic artifacts on EEG data; S2.2, perform bad track detection and interpolation on EEG data; S2.

3. Split the EEG data proportionally.

3. The autism identification method based on EEG cross-band coupling according to claim 2, characterized in that: Step S2 preprocesses the raw EEG data and also includes tracing the source of the raw data, obtaining the time-domain EEG signal in the source space, identifying the specific brain area corresponding to the scalp physiological activity through EEG signal source localization technology, and reconstructing the time series of the source signal at the cortical level. The specific method is: the mapping relationship between the scalp electrode space and the source space at a certain sampling point is expressed by Maxwell's equations as follows: Φ=MX+ε in and They represent the scalp potential vector recorded from C electrode channels and the source signal vector composed of V brain regions of interest; is the lead field matrix, represents the noise vector; Solve the following inverse problem: Under the conditions of given M, Φ and regularization parameter α≥0, the above explicit solution is: Where, the superscript '+' indicates the Moore-Penrose pseudoinverse, is the center matrix, Perform standardization to obtain the sLORETA estimation value of each ROI; divide the cortex into multiple source space ROIs according to the brain area division map, extract the time series of all ROIs by repeating the above process, and finally form a reconstructed source signal matrix where N R represents the number of ROIs, and N represents the number of sampling points.

4. The autism identification method based on EEG cross-band coupling according to claim 1, characterized in that: The specific steps for calculating the characteristics of cross-band coupling described in step S3 are as follows: S3.1.1: The time domain signal is represented as X = (x1, x2, ..., x N ), the time evolution of the activity power at the signal X frequency υ is called Y υ Representing the variation of power over time, is estimated using a sliding window of M samples over the discrete Fourier transform of successive segments extracted from X: in is the active power of frequency υ at sampling point n, Represents the Hanning taper, where the length of M is five periods relative to the frequency υ, i.e., M = 5·Fs / υ; S3.1.2: Next, X and Y υ is divided into a set of S consecutive segments with l overlapping time points; X s represents the time series of segment s; Y υ,s represents the power time evolution of segment s, that is, the power change sequence of the signal segment in the time dimension; the standard fast Fourier transform is used to calculate their spectrum: where n FFT =256 represents the frequency resolution of the fast Fourier transform, the center frequency of the FFT transform S3.1.3: Next, calculate signal X and signal Y υ Complex coherence between the power envelopes of: The frequency is "*" represents complex conjugate, C(υ,f) represents signal X and signal Y υ The complex coherence between the power envelopes of at frequency point f; S3.1.4: Phase Slope Index Estimate the slope of the phase difference as a function of frequency within a specified frequency band, where the sign of the slope indicates the direction of the interaction; Next, calculate the phase-amplitude coupling where signal X and signal Y υ At frequency (υ,f j ) is the phase slope index of the power envelope at: where f j represents the low frequency of signal X, Δf represents the frequency resolution, Im represents the imaginary part, and β = 2 Hz is used to represent the bandwidth for calculating the phase slope; S3.1.5: Finally, calculate the low-frequency band B p Phase (θ or α) and high frequency band B a The cross-band coupling value PAC between the amplitudes of (β or γ), that is, all pairwise combinations (f k ,f l ) in the maximum phase slope index: The phase-amplitude coupling calculation method is used to calculate the cross-band coupling between the low-frequency and high-frequency components of the EEG signal collected in a single channel, and the cross-band coupling of signals between different channels. The calculated cross-frequency domain coupling characteristics of the scalp channel are used as EEG features for classification.

5. The autism identification method based on EEG cross-band coupling according to claim 3, characterized in that: The specific steps for calculating the characteristics of cross-band coupling described in step S3 are as follows: S3.2.1: The time domain signal is represented as X = (x1, x2, ..., x N ), the time evolution of the activity power at the signal X frequency υ is called Y υ To represent the variation of power over time, we estimate the discrete Fourier transform of successive segments extracted from X using a sliding window of M samples: in is the active power at frequency υ and sampling point n, Represents the Hanning taper, where the length of M is five periods relative to the frequency υ, i.e., M = 5·Fs / υ; S3.2.2: Next, X and Y υ is divided into a set of S consecutive segments with l overlapping time points, X s represents the time series of segment s; Y υ,s Represents the power time evolution of segment s, that is, the power change sequence of the signal segment in the time dimension, and we use the standard fast Fourier transform to calculate their spectrum: where n FFT =256 represents the frequency resolution of the fast Fourier transform; S3.2.3: Next, calculate signal X and signal Y υ Complex coherence between the power envelopes of: The frequency is "*" indicates complex conjugation; S3.2.4: Phase Slope Index Estimate the slope of the phase difference as a function of frequency within a specified frequency band, where the sign of the slope indicates the direction of the interaction. Phase-amplitude coupling is quantified using the Jiang method and is calculated for signal X and signal Y. υ At frequency (υ,f j ) is the phase slope index of the power envelope at: where f j represents the low frequency of signal X, Δf represents the frequency resolution, Im represents the imaginary part, and β = 2 Hz is used to represent the bandwidth for calculating the phase slope; S3.2.5: Finally, the cross-band coupling of ROI signals with different source spaces is calculated, and the cross-band coupling features of ROI signals with different source spaces are used as EEG features for classification. The low-frequency band B is calculated. p Phase (θ or α) and high frequency band B a The cross-frequency coupling value PAC between the amplitudes of (β or γ), that is, all pairwise combinations (f k ,f l ) in the maximum phase slope index:

6. The autism identification method based on EEG cross-band coupling according to claim 5, characterized in that: The method also includes calculating the cross-frequency coupling characteristics within and between source space networks, and using the calculated cross-band coupling characteristics within and between networks as features for classification. Based on the distribution of ROIs, the source signals within each ROI are averaged, and then mapped to the seven functional network systems proposed by Yeo based on the spatial distribution characteristics of multiple ROIs in the source space. The definition of intra-network phase amplitude coupling is: the mean PAC value of all ROIs in a specific functional network, calculated as follows: The calculation of PAC phase-amplitude coupling refers to the low-frequency band B p The phase and high frequency band B a The coupling between the amplitudes, Indicates a specific network net i The time course signal of the h-th ROI in , H represents the total number of ROIs contained in the network; At the same time, the cross-frequency coupling characteristics between source space networks are calculated. First, the connection strength is quantified by calculating the PAC in the β-θ frequency band: the phase of the low-frequency component of a certain ROI in one network is coupled with the amplitude of the high-frequency component of a certain ROI in another network. Second, the average of all point-to-point connections between two networks with the same phase-amplitude coupling direction is taken as the inter-network connection strength based on network characteristics. The calculation formula is as follows: Where H1 and H2 represent the network i with net j The total number of ROIs included; The cross-band coupling characteristics of time domain signals of different ROIs in the source space, and the cross-band coupling characteristics within and between networks can be used as features for classification.

7. The autism identification method based on EEG cross-band coupling according to claim 6, characterized in that: The network system includes but is not limited to the visual network, sensorimotor network, dorsal attention network, ventral attention network, limbic network, frontoparietal control network and default mode network.

8. The autism identification method based on EEG cross-band coupling according to any one of claims 1 to 7, characterized in that: Step S4 includes: Use a two-layer transformer network for modeling and classification. Use sliced data of 64×1000, where 64 is the number of channels and 1000 is the number of features. Put the data of the training set into the two-layer transformer model for learning. The steps are as follows: S4.1 Position coding: Sine and cosine functions are used to encode position information. set up Represents the position encoding matrix, where pos is the position index, 0≤pos≤N c , 0≤pos≤N c ,2i or 2i+1 is the dimension index, 0≤2i≤2i+1≤N T As shown in the above formula, the position code is directly calculated from the data segment x, and the timing information is introduced to superimpose the position code on the original input data O; Embed position encoded data It is then fed into a multi-head self-attention layer; S4.2 Multi-head self-attention layer; In the Transformer encoder, the multi-head attention module captures the dependencies between vectors at different positions; In the VAE encoder module, the input vector is replicated into three copies; the three copies are called query, key, and value in the calculation, and are recorded as Q, K, and V, where Q, K, and V are actually the same Subsequently, Q, K, and V are divided into H segments, and subsequent calculations are performed within each sub-segment. This module is called H heads. In each head, the query, key, and value sub-segments are fed into three separate linear layers, and the outputs of the linear layers are processed by the attention mechanism. Then, the attention calculation results from H heads are concatenated and mapped from high-dimensional to the same dimension as the input through another fully connected layer. The same dimension, and then processed by the dropout layer, the process can be expressed as: MHA(Q,K,V)=[head0;…;head H-1 ] head i =Attention(Q i ,K i ,V i ) where represents multi-head attention, and represents the query, key, and value obtained by linear transformation in the i-th head; After passing through two layers of transformers in S4.3, the data is fed into a linear layer to obtain the classification label. The loss function of the network described in S4.4 is the cross-entropy loss function, which calculates the cross-entropy loss between the probability distribution of the network output and the target probability distribution; the optimizer evaluation metric used is accuracy, that is, the ratio of correctly predicted samples to the total number of samples; S4.5 further trains the network. The specific process is as follows: the input data passes through each layer of the neural network in sequence, that is, after forward propagation, back propagation is performed. The back propagation process includes calculating the loss, that is, using the cross entropy loss function to calculate the cross entropy loss between the probability distribution of the network output and the target probability distribution, calculating the gradient, that is, calculating the gradient of the network weight matrix and weight coefficient according to the cross entropy loss function, and updating the parameters, that is, using the Adam optimizer to update the network weight matrix and weight coefficient according to the calculated gradient; The training ends after 200 rounds of training; S4.6 For the data in the test set, put the data into the classification network model to obtain the label corresponding to the data to achieve the purpose of autism identification.

Images