An electroencephalogram signal dynamic partitioning method based on phase-locked value time evolution

Abstract

This invention relates to the field of EEG signal processing and pattern recognition technology, and particularly to a dynamic partitioning method for EEG signals based on the temporal evolution of phase-locked values. The method includes: acquiring a feature matrix of EEG phase-locked values ​​under multiple time windows; calculating a temporal merging score based on the feature relationships between brain regions under each time window, incorporating temporal functional connectivity similarity, evolutionary consistency, spatial proximity and drift penalty, and contribution consistency; determining whether to merge brain regions to update the partitioning based on an adaptive dynamic threshold; and performing cross-time window verification on the updated partitions, reverting to the pre-merging state if the classification loss of a certain time window exceeds the threshold. This invention achieves adaptive optimization of brain region partitioning through a time-dimensional driven partitioning and reversion mechanism, improving the accuracy and robustness of EEG emotion recognition.

Description

Technical Field

[0001] This invention relates to the field of electroencephalogram (EEG) signal processing and pattern recognition technology, and in particular to a dynamic partitioning method for EEG signals based on the temporal evolution of phase-locked values. Background Technology

[0002] Electroencephalograms (EEGs) have become important tools in emotion recognition, cognitive state assessment, and brain-computer interface research due to their high temporal resolution and non-invasiveness. In brain functional network analysis, phase-locked value (PLV) is widely used to quantify the phase synchronization between different EEG electrode channels to reflect the strength of functional connectivity between brain regions.

[0003] Existing PLV-based EEG emotion recognition methods typically employ a fixed anatomical partitioning strategy. This involves dividing the electrode array into several fixed brain regions, such as the frontal lobe, temporal lobe, parietal lobe, occipital lobe, and central region, based on brain anatomy, and then performing feature extraction and classification modeling within or between these regions. However, this fixed partitioning strategy has the following significant drawbacks: First, it ignores the differences in functional networks among individuals. The brain functional network organization patterns vary significantly among different individuals, and fixed anatomical divisions cannot adapt to the functional topology of individuals, leading to suboptimal feature extraction.

[0004] Second, it neglects the time-varying characteristics of brain region functional connectivity. In dynamic emotion tasks such as microexpression induction, the functional coupling and separation between brain regions evolve rapidly over time. Most existing methods are based on the average PLV of the entire data set or static partitioning within a single time window, failing to capture the dynamic reorganization process of brain functional networks over time.

[0005] Third, the partitioning decision lacks temporal consistency verification. Traditional methods rely solely on the spatial similarity at the current moment when deciding whether to merge two brain regions, without assessing whether their evolutionary trajectories on continuous time axes are synchronized, leading to unreasonable and drastic abrupt changes in partitioning results within adjacent time windows.

[0006] Fourth, there is a lack of post-validation and fault tolerance mechanisms for the partitioning effect. Existing methods use partitions directly once they are determined. Even if the merging operation within a certain time window impairs the subsequent classification performance, there is no way to backtrack and correct it, resulting in the accumulation and amplification of errors during the feature extraction and classification stages.

[0007] To address the aforementioned issues, researchers have proposed emotion classification methods based on dynamic graph neural networks or dynamic brain functional networks, and patents have also disclosed personalized EEG emotion recognition schemes based on dynamic brain region segmentation. However, these existing schemes still mainly rely on static algorithms such as graph clustering or community detection to partition brain regions at a single point in time, failing to systematically solve the dynamic adaptability problem of brain region segmentation from the perspective of temporal evolution, and also lacking post-validation and backtracking correction mechanisms for partitioning decisions.

[0008] To address this, this invention proposes a dynamic brain region partitioning method based on the temporal evolution of phase-locked values. By constructing a merging and discrimination criterion that integrates multi-dimensional temporal information, introducing an adaptive dynamic threshold, and employing a pioneering verification backoff mechanism, the method achieves adaptive optimization of the brain region partitioning scheme in the temporal dimension. This method can be applied to fields such as EEG emotion recognition induced by micro-expressions, brain-computer interfaces, and neurofeedback systems. Summary of the Invention

[0009] This invention discloses a dynamic partitioning method for electroencephalogram (EEG) signals based on the temporal evolution of phase-locked values, comprising: Obtain the feature matrix of EEG phase-locked values ​​under multiple consecutive time windows; Based on the characteristic relationship between any two current brain regions under each time window, calculate the temporal merging score, which reflects whether the two brain regions are suitable for merging in the time dimension; The temporal merging score is judged based on the adaptively adjusted dynamic threshold to determine whether the two brain regions should be merged, thereby updating the brain region division scheme. The updated brain region segmentation scheme is validated across time windows. If the classification performance loss of any time window exceeds a preset threshold, the time window is reverted to the brain region segmentation state before merging.

[0010] Furthermore, the temporal merging score is obtained by weighted fusion of the following four dimensions: temporal functional connectivity similarity, temporal evolution consistency, spatial proximity and drift penalty, and temporal contribution consistency.

[0011] Furthermore, the temporal functional connectivity similarity is the weighted average of the functional connectivity matrices between two brain regions under each time window on the time axis, wherein the fusion weight adaptively increases with the order of the time windows.

[0012] Furthermore, the temporal evolution consistency is the correlation coefficient of the regional feature sequences of two brain regions on the time axis, which is used to measure whether their dynamic change trajectories are synchronized.

[0013] Furthermore, the temporal contribution consistency is calculated based on the variance of the contribution of the two brain regions to the preset classification model under each time window; the smaller the variance of the contribution, the higher the consistency.

[0014] Furthermore, the dynamic threshold is adaptively adjusted based on the ratio of the current time window number to the total number of time windows and the current training iteration round, and the adjustment range increases as the ratio and the iteration round increase.

[0015] Furthermore, the preset threshold for the classification performance loss is 1.5%; the rollback operation is executed independently in each time window, and the maximum number of global rollback time windows is 2.

[0016] Furthermore, it also includes placing the updating of the brain region segmentation scheme, the extraction of regional features, and the classification in an end-to-end joint training framework, and introducing a partition stability regularization term in the total loss function to constrain the smooth change of the brain region segmentation scheme between adjacent iteration rounds.

[0017] Furthermore, the EEG phase-locked value feature matrix is ​​used to segment and extract the original EEG signal based on a sliding time window. The number of time windows is dynamically set according to the signal length of the micro-expression induction period, and the initial brain region division of the first time window is determined based on the anatomical location or the clustering result of the phase-locked value of the first time window.

[0018] Due to the adoption of the above technical solutions, the present invention has the following beneficial effects: 1. This invention is the first to base brain region partitioning decisions on the dimension of temporal evolution. Unlike existing static clustering or single-time-window dynamic partitioning methods, this invention integrates four temporal dimension indicators: temporal functional connectivity similarity, temporal evolution consistency, spatial proximity and drift penalty, and temporal contribution consistency. It comprehensively evaluates whether brain regions should be merged from four orthogonal dimensions: functional connectivity pattern, evolutionary trajectory synchronization, spatial smoothness, and functional stability, filling the technological gap of "brain region partitioning based on temporal evolution".

[0019] 2. Introducing a dynamic adaptive threshold mechanism. This invention abandons the approach of fixed thresholds or simple graph density thresholds, and adopts a dynamic threshold that adaptively decreases with the progress of the time window and the number of training iterations. This makes the partitioning strategy more conservative in the early stage of training and more flexible in the later stage, better adapting to the progressive characteristics of the model learning process.

[0020] 3. Pioneering "Verification and Local Rollback" Fault Tolerance Mechanism. After executing the merge decision, this invention reverse-verifies the classification performance of the partitioning scheme in each time window, and independently performs rollback operations for time windows where performance loss exceeds a threshold. This mechanism makes partition merging "reversible," effectively avoiding the accumulation and amplification of errors in subsequent stages and significantly improving the robustness of the partitioning scheme.

[0021] 4. End-to-end joint optimization framework. This invention places partitioning update, feature extraction, and classification in the same framework for joint training, and introduces a partitioning stability regularization term, so that the partitioning scheme is optimized collaboratively with the feature extraction network and the classification network, rather than making partitioning decisions in isolation, thereby obtaining globally optimal recognition performance.

[0022] Other advantages, objectives, and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination, or may be learned from practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description

[0023] The accompanying drawings of this invention are described below.

[0024] Figure 1 This is a flowchart of Example 1.

[0025] Figure 2 This is a flowchart of Example 2.

[0026] Figure 3 This is a flowchart of Example 4. Detailed Implementation

[0027] The present invention will be further described below with reference to the accompanying drawings and embodiments. Example

[0028] like Figure 1 As shown, the present invention provides a dynamic partitioning method for EEG signals based on the temporal evolution of phase-locked values, comprising the following steps: Step S101, Data Acquisition and Preprocessing. Multichannel EEG signals are acquired from the subjects during the micro-expression evoked task. The raw EEG signals undergo preprocessing operations such as denoising, filtering, and oculomotor removal. The preprocessed EEG signals are then divided into preset sliding time windows. The length of each time window can be set according to task requirements (e.g., window length of 2 seconds, step size of 0.5 seconds), resulting in a total of N_t time windows.

[0029] Step S102, PLV Feature Matrix Construction. For the multi-channel EEG signal within each time window, calculate the phase lock value (PLV) between any two electrode channels. PLV calculation can employ methods such as Hilbert transform, Fourier transform, or wavelet transform. Taking Hilbert transform as an example, perform a Hilbert transform on the signal of each channel to obtain the analytic signal, extract the instantaneous phase, and then calculate the average magnitude of the instantaneous phase difference between the two channels, which is the PLV value. Organize the calculation results into a PLV feature matrix of dimension E×E×B, where E is the number of electrodes and B is the number of frequency bands (such as sub-bands like theta wave, alpha wave, beta wave, and gamma wave). Each time window corresponds to one PLV feature matrix, thus obtaining a sequence of N_t PLV feature matrices.

[0030] Step S103, Initial Brain Region Division. Determine the initial brain region division scheme for the first time window. The initial division can be based on anatomical location, dividing the electrodes into five brain regions: frontal lobe, temporal lobe, parietal lobe, occipital lobe, and central region. As an alternative, clustering (such as K-means clustering or spectral clustering) can be performed based on the PLV feature matrix of the first time window, automatically clustering the electrodes into several functional groups as initial partitions.

[0031] Step S104, Partition Iterative Optimization. For the current brain region partitioning scheme, traverse all brain region pairs and perform the following operations for each pair: extract the PLV feature submatrices of the two brain regions under each time window, calculate the temporal merging score; determine whether to merge based on the current dynamic threshold; if merging, update the partitioning scheme; after merging, verify and back off all time windows. The above process is executed once every T iterations in the end-to-end training framework (T is a preset hyperparameter, e.g., T=5), until the model converges or reaches the preset maximum number of iterations. Example

[0032] like Figure 2 As shown, the temporal merging score is the core criterion of this invention. The calculation process for the temporal merging score of any two brain regions i and j is as follows: S201. Calculate temporal functional connectivity similarity. For each time window τ (τ=1,2,...,N_t), extract the functional connectivity submatrix FCS^(τ)_ij between brain region i and brain region j (with dimensions n_i×n_j×B, where n_i and n_j are the number of electrodes in the two brain regions, and B is the number of frequency bands). The formula for calculating TFCS is: TFCS(i,j)=Σ_{τ=1}^{N_t}α_τ×mean(FCS^(τ)_ij) S202, Computing the consistency of temporal evolution For each brain region r, its average eigenvectors under each time window are arranged in chronological order to form a temporal feature sequence V_r=[v_r(1),v_r(2),...,v_r(N_t)], where v_r(τ) is the average eigenvector of brain region r under time window τ (obtained by averaging the PLV values ​​of all electrodes and all frequency bands of that brain region). The formula for calculating TEC is: TEC(i,j)=corr(V_i,V_j) Where corr(·,·) is the Pearson correlation coefficient. The higher the TEC value, the more synchronized the dynamic trajectories of the two brain regions are, and the more likely they are to merge.

[0033] S203, Computational Spatial Proximity and Drift Penalty The TSP consists of two parts: spatial proximity score and drift penalty term.

[0034] Spatial proximity score is calculated based on the physical spatial location of electrodes contained in two brain regions. Let the spatial coordinates of the electrodes in brain region i be set P_i, and the spatial coordinates of the electrodes in brain region j be set P_j. Then, the spatial proximity score is defined as the negative exponent of the average distance between the sets of electrode coordinates of the two brain regions: SpatialProx(i,j)=exp(-dist_mean(P_i,P_j) / σ) Where dist_mean(·,·) is the average Euclidean distance between the two sets, and σ is the spatial scale parameter.

[0035] The drift penalty term is used to constrain the magnitude of change in the partitioning scheme between adjacent time windows. Let the merging state of brain regions i and j in the previous iteration be M_prev(i,j) (1 indicates merged, 0 indicates not merged), and the merging decision in the current round be M_curr(i,j). Then the drift penalty is: DriftPenalty(i,j)=γ×|M_curr(i,j)-M_prev(i,j)| Where γ is the drift penalty coefficient. The comprehensive calculation formula for TSP is: TSP(i,j)=SpatialProx(i,j)-DriftPenalty(i,j) S204, Calculate the consistency of time-series contributions For each brain region r, calculate its contribution c_r(τ) to the preset classification model at each time window τ. The contribution can be quantified in various ways, such as calculating the gradient norm of the classification model's loss function with respect to the features of brain region r, or utilizing the attention weights corresponding to brain region r in the attention mechanism. A preferred implementation is as follows: c_r(τ)=||∂L / ∂f_r(τ)|| Where L is the classification loss, f_r(τ) is the feature representation of brain region r under time window τ, and ||·||2 is the L2 norm.

[0036] For brain region r, its contribution sequence C_r = [c_r(1), c_r(2), ..., c_r(N_t)]. The formula for calculating TCC is: TCC(i,j)=1-(var(C_i)+var(C_j)) / 2 Where var(·) represents the variance. A higher TCC value indicates that the functional contributions of the two brain regions are more stable and consistent over time.

[0037] S205, Weighted Fusion The weighted sum of the indicators from the four dimensions above yields the final time-series merged score: MergeScore_time(i,j)=w1×TFCS(i,j)+w2×TEC(i,j)+w3×TSP(i,j)+w4×TCC(i,j) Where w1, w2, w3, and w4 are preset weight coefficients, satisfying w1 + w2 + w3 + w4 = 1. Each weight can be optimized through cross-validation according to the specific task. In a preferred implementation, the weights can be set to w1 = 0.30, w2 = 0.25, w3 = 0.20, and w4 = 0.25. Example

[0038] The dynamic threshold in this invention adaptively adjusts as the training process progresses. The formula for calculating the dynamic threshold is: θ=θ_base-λ×(mean_tau / N_t)-η×(N_t / N_max) The meanings and preferred value ranges of each parameter are as follows: θ_base: Baseline threshold, ranging from 0.5 to 0.9, preferably 0.75. The baseline threshold reflects the conservatism of the algorithm in the initial stage of the merging operation; mean_tau: The average index of the current time window being considered. In the early stages of training, mean_tau is smaller, and the threshold is higher; as training progresses, mean_tau increases, and the threshold gradually decreases. N_t: Total number of time windows; N_max: Maximum number of training iterations, set according to the size of the dataset and the complexity of the model; λ: Time decay coefficient, with a value ranging from 0.1 to 0.5, preferably 0.3, used to control the rate of decrease of the threshold as the time window progresses; η: Iterative decay coefficient, with a value ranging from 0.05 to 0.2, preferably 0.1, used to control the rate at which the threshold decreases as the training rounds increase.

[0039] The physical significance of this dynamic threshold mechanism is that in the early stages of training, the model's understanding of the data distribution is not yet sufficient, so a higher threshold should be used to conservatively merge partitions and avoid incorrect decisions. As training progresses, the model gradually learns the distribution characteristics of the data, and at this point, lowering the threshold allows for more flexible region recombination to capture a more refined functional network structure. Example

[0040] like Figure 3 As shown, the verification and partial rollback mechanism of the present invention includes the following sub-steps: S401. Post-merging performance evaluation. After completing one round of brain region merging decisions, the updated partitioning scheme R_new is obtained. R_new is applied to all N_t time windows, and region features are extracted for each time window τ. These features are then input into a pre-defined classification model, and the classification performance Metric_new(τ) (e.g., classification accuracy) is calculated.

[0041] S402. Performance Loss Calculation. Obtain the classification performance Metric_old(τ) for each time window under the partition scheme R_old before merging. Calculate the performance loss for each time window: Loss(τ)=Metric_old(τ)-Metric_new(τ) S403. Backtracking Decision. For each time window τ, determine whether Loss(τ) is greater than a preset threshold δ (preferably δ=1.5%). If it is greater, it is determined that the merging operation of that time window has impaired the classification performance, triggering a backtrack.

[0042] S404. Independent rollback execution. For each time window that triggers rollback, its partition is restored to its state R_old before merging. The rollback decisions for each time window are independent of each other, that is, rolling back time window A does not affect the partition state of time window B.

[0043] S405. Rollback Quantity Constraint. To prevent excessive rollbacks from causing over-fragmentation of the partitioning scheme, a global rollback quantity limit K_max is set (preferably K_max=2). If the number of time windows that trigger rollbacks exceeds K_max, rollbacks are only performed on the K_max time windows with the greatest performance loss, while the remaining time windows maintain the merged partitioning state.

[0044] The aforementioned verification and rollback mechanism makes the partition merging operation of this invention "reversible," effectively avoiding the cumulative amplification of errors caused by an improper merging decision in subsequent processing, and significantly improving the robustness of the partitioning scheme and the final classification performance. Example

[0045] This invention places the updating of brain region segmentation schemes, region feature extraction, and classification within an end-to-end joint training framework. Specifically, the loss function of the entire training framework consists of three parts: 1. Classification loss L_cls: The cross-entropy loss function is used to measure the difference between the mood category predicted by the model and the true label.

[0046] 2. Partition stability regularization term L_stab: Used to constrain the smooth change of brain region partitioning schemes between adjacent iterations, avoiding drastic oscillations in partitions during training. The formula for calculating L_stab is: L_stab=β×Σ_{i <j}|M_t(i,j)-M_{t-1}(i,j)| Where M_t(i,j) is the merged state of brain region i and brain region j in the t-th iteration, and β is the regularization coefficient (preferably β=0.01).

[0047] 3. Total loss function: L_total = L_cls + L_stab Through the aforementioned end-to-end joint training, the parameters of the partitioning scheme, feature extraction network, and classification network are updated collaboratively within the same optimization framework. This ensures that the partitioning decision always evolves in the direction of improving the final recognition accuracy, rather than partitioning in isolation, thereby achieving globally optimal recognition performance.

[0048] This invention can be applied to EEG emotion recognition tasks induced by micro-expressions. In a specific test case, the method of this invention was used to process an EEG dataset. Experimental results showed that the method of this invention achieved an emotion classification accuracy of over 92%, which is a significant improvement compared to traditional fixed partitioning methods and existing dynamic partitioning methods. Ablation experiments further verified the effectiveness of each core module: adding temporal merging scores improved performance by about 2-3 percentage points; introducing dynamic thresholds further improved performance by about 1 percentage point; the addition of verification and backoff mechanisms further improved performance by about 0.5-1 percentage points, and effectively suppressed temporal jitter in the partitioning results.

[0049] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A dynamic partitioning method for electroencephalogram (EEG) signals based on the temporal evolution of phase-locked values, characterized in that, include: Obtain the feature matrix of EEG phase-locked values ​​under multiple consecutive time windows; Based on the characteristic relationship between any two current brain regions under each time window, calculate the temporal merging score, which reflects whether the two brain regions are suitable for merging in the time dimension; The temporal merging score is judged based on the adaptively adjusted dynamic threshold to determine whether the two brain regions should be merged, thereby updating the brain region division scheme. The updated brain region segmentation scheme is validated across time windows. If the classification performance loss of any time window exceeds a preset threshold, the time window is reverted to the brain region segmentation state before merging.

2. The method according to claim 1, characterized in that, The temporal merging score is obtained by weighted fusion of the following four dimensions: temporal functional connectivity similarity, temporal evolution consistency, spatial proximity and drift penalty, and temporal contribution consistency.

3. The method according to claim 2, characterized in that, The temporal functional connectivity similarity is the weighted average of the functional connectivity matrices between two brain regions in each time window on the time axis, wherein the fusion weights adaptively increase with the order of the time windows.

4. The method according to claim 2, characterized in that, The temporal evolution consistency is the correlation coefficient of the regional feature sequences of two brain regions on the time axis, which is used to measure whether their dynamic change trajectories are synchronized.

5. The method according to claim 2, characterized in that, The temporal contribution consistency is calculated based on the variance of the contribution of the two brain regions to the preset classification model under each time window. The smaller the contribution variance, the higher the consistency.

6. The method according to claim 1, characterized in that, The dynamic threshold is adaptively adjusted based on the ratio of the current time window number to the total number of time windows and the current training iteration round. The adjustment range increases as the ratio and the iteration round increase.

7. The method according to claim 1, characterized in that, The preset threshold for the classification performance loss is 1.5%; the rollback operation is executed independently in each time window, and the maximum number of global rollback time windows is 2.

8. The method according to claim 1, characterized in that, It also includes placing the updating of the brain region partitioning scheme, the extraction of regional features and classification in an end-to-end joint training framework, and introducing a partition stability regularization term in the total loss function to constrain the smooth change of the brain region partitioning scheme between adjacent iteration rounds.

9. The method according to claim 1, characterized in that, The EEG phase-locked value feature matrix is ​​used to segment and extract the original EEG signal based on a sliding time window. The number of time windows is dynamically set according to the signal length during the micro-expression induction period, and the initial brain region division of the first time window is determined based on the anatomical location or the clustering result of the phase-locked value of the first time window.

Images