Language brain-computer interface decoding method based on long and short range functional connection graph neural network

Abstract

The application discloses a language brain-computer interface decoding method based on a long-short-range functional connection graph neural network, comprising the following steps: obtaining original neural signals, and obtaining a neural signal matrix after preprocessing; constructing a long-range functional connection graph based on the neural signal matrix, and obtaining a long-range pooling graph through a channel pooling mechanism; constructing a short-range functional connection graph based on the neural signal matrix and performing average time pooling; performing graph representation learning on the long-range pooling graph and the pooled short-range functional connection graph; splicing and normalizing long-range graph representation, short-range graph representation and representation extracted from the original neural signals to obtain fused neural representation, which is then input into a downstream decoder to output at least one of a movement behavior category, a speech intention category, a written character category or a cognitive state category. The application can simultaneously capture global stable connections and local transient dynamics, reduce channel redundancy, and improve the accuracy, robustness and interpretability of neural decoding.

Description

Technical Field

[0001] This invention relates to the field of brain-computer interface neural signal decoding, and in particular to a language brain-computer interface decoding method based on long-short-range functional connectivity graph neural networks. Background Technology

[0002] Brain-computer interface (BCI) systems directly read motor, speech, or cognitive intentions from brain neural activity and convert them into control commands or communication content for external devices, providing an important technological pathway for patients with limb paralysis, speech disorders, and other conditions to regain motor control and natural communication abilities. In invasive BCIs, microelectrode arrays can record spike firing or firing rate sequences of multiple neural channels. These high-dimensional neural signals simultaneously contain spatial distribution, temporal variations, and inter-channel synergistic relationships.

[0003] Existing neural decoding methods typically treat neural activity at each moment as a sequence of ordinary vectors and use support vector machines, Kalman filters, recurrent neural networks, temporal convolutional networks, or Transformer networks to learn the mapping relationship between neural activity and behavioral variables. For example, Chinese patent document CN115617180A discloses an invasive brain-computer interface method for decoding fine hand movements. The core model uses a recurrent neural network to decode continuous hand movement regression / dexterity movements. These methods can describe the temporal dynamics of neural signals to some extent, but they often do not explicitly establish the functional connections between different electrode channels and different neuron groups.

[0004] In recent years, graph neural networks have been used to model spatial or functional relationships between neural channels. The basic idea is to treat channels or neurons as graph nodes, and correlations, co-activations, or other statistical connections between channels as graph edges, extracting structured representations of group neural activity through message passing. For example, Chinese patent document CN114611556A discloses a method for classifying EEG motor imagery tasks, using Pearson correlation to obtain a graph structure matrix, which is then input into a graph neural network for motor imagery classification. However, existing graph structure decoding methods often construct single-scale graphs: one type of method focuses on stable relationships across all trials, easily ignoring instantaneous changes in local time segments that are discriminative of behavioral categories; another type of method focuses on local co-activations within short time windows, easily losing the global context that persists throughout the entire trial.

[0005] Neural activity related to language tasks inherently possesses multi-scale functional connectivity. Taking writing and speech tasks as examples, long-range stable connections may correspond to overall intentions such as writing characters, words, or task states, while short-range, rapidly changing connections may correspond to strokes, phonemes, articulation actions, or local kinematic details. If the decoder only utilizes one scale, it is prone to information loss, thereby reducing the decoding accuracy and noise robustness of complex behaviors.

[0006] Furthermore, invasive electrode recordings commonly suffer from issues such as channel redundancy, local noise, neural activity drift, and temporary channel inactivation. If all channels are directly used as equivalent inputs for decoding, redundant or low-quality channels may interfere with model learning, making it difficult for the model to obtain stable and interpretable neural representations. Therefore, it is necessary to design a language brain-computer interface decoding method that simultaneously considers long-range stable connectivity and short-range dynamic connectivity, and is capable of adaptively aggregating functionally similar channels. Summary of the Invention

[0007] This invention provides a language brain-computer interface decoding method based on long- and short-range functional connectivity graph neural networks. It addresses the problem that existing neural decoding algorithms cannot simultaneously capture global stable connectivity and local transient dynamics. In each trial, a long-range functional connectivity graph and a short-range coactivation graph are constructed, and multi-scale structured neural representations are extracted through graph sampling and aggregation networks, thereby improving the decoding performance and robustness of tasks such as speech, writing, and even movement in brain-computer interfaces.

[0008] A language brain-computer interface decoding method based on long-short-range functional connectivity graph neural networks includes the following steps: (1) Obtain the raw neural signals of the brain-computer interface task, and obtain the neural signal matrix after preprocessing the raw neural signals. , where C represents the number of recording channels or neurons, and T represents the time step; (2) Based on the neural signal matrix X, a long-range functional connectivity graph is constructed, the stable correlation between channels is calculated within the entire trial range, and a long-range pooling graph is obtained through the channel pooling mechanism. This long-range pooling graph is constrained by structural reconstruction loss and link-assisted loss. (3) Divide the neural signal matrix X into multiple short time windows with overlapping parts, construct a short-range functional connectivity graph sequence based on the co-activation relationship of neurons in each window, and perform average time pooling on adjacent short-range functional connectivity graphs to eliminate redundant information. (4) Graph representation learning is performed on the long-range pooled graph and the pooled short-range functional connection graph sequence using graph sampling and aggregation networks respectively, to obtain long-range graph representation and short-range graph representation; (5) The long-range graph representation, the short-range graph representation, and the representation extracted from the original neural signal are spliced ​​and normalized to obtain the fused neural representation, which is then input into the downstream decoder and outputs at least one of the following categories: motor behavior category, speech intention category, written character category, or cognitive state category.

[0009] In step (2), a long-range functional connectivity graph is constructed based on the neural signal matrix X, specifically including: For the neural signal matrix X, take the first... Full-time sequence of each channel and the Full-time sequence of each channel Calculate the Pearson correlation coefficient between the two. Based on relevant thresholds Or the top-K related neighbors of each node determine the graph topology, when Greater than the threshold or channel It belongs to the channel When collecting Top-K related neighbors, in the channel With channel Establish edges between them.

[0010] In step (2), the stable correlation between channels is calculated across the entire trial range, specifically including: While maintaining the topology of the long-range functional connectivity graph, the time series data of two connected channels are concatenated and input into a shared multilayer perceptron to obtain learnable edge weights. and will and Together as a long-range weighted adjacency matrix The weights of the corresponding edges are calculated using the following formula: ; ; in, This represents a shared multilayer perceptron. That is, the channel and channels Stable long-range functional connections between them.

[0011] In step (2), the channel pooling mechanism includes: Generating soft-assignment matrices using graph neural networks Where K is the preset upper limit of the number of functional clusters; according to Obtain pooled node features ,in The node feature matrix of the long-range functional connectivity graph; based on Obtain the adjacency matrix of the long-range pooling graph ,in For stable long-range functional connections between two channels.

[0012] In step (2), the structural reconstruction loss The link-assisted loss is used to constrain long-range pooling graphs to maintain the original long-range graph topology. The loss function used to constrain the predictive ability of edge existence for node embedding pairs is as follows: ; ; in, regarded as a node With nodes The probability of an edge between two points is log-odd, which is determined by... Calculated; Represents the standard binary cross-entropy with logarithmic odds; For nodes , Long-range functional connections between them; This represents the Frobenius norm structural loss.

[0013] In step (3), constructing the short-range functional connectivity graph specifically includes: The neural signal matrix X is divided into multiple short-time sub-matrices by sliding partitioning with window length w and step size s; for any short-time sub-matrix m, the channels are calculated. and channels The frequency of simultaneous activation within the corresponding window is used as the co-activation score. And determine the edges of the short-range functional connectivity graph based on the co-activation threshold τ; Node features are extracted from each short-time submatrix m using average firing rate, total pulse count, and energy statistics. These features are then input into a shared multilayer perceptron encoder to generate node features. ,in, For encoder dimensions.

[0014] In step (3), average time pooling is performed on the adjacent short-range function connectivity graph, specifically including: Time pooling connects two adjacent short-range functions in the graph. The node features are averaged, and the adjacency matrix is ​​maximized element by element, so that the pooled short-range functional connection graph inherits significant short-range connections in adjacent time windows.

[0015] In step (5), the downstream decoder is a multilayer perceptron that includes nonlinear activation ReLU and dropout regularization, and the training objective is a weighted sum of cross-entropy loss, structural reconstruction loss and link-assisted loss.

[0016] Compared with the prior art, the present invention has the following beneficial effects: First, long-range functional connectivity maps and short-range coactivation maps complement each other, enabling the simultaneous utilization of global context and local dynamic details in neural activity. Second, learnable channel pooling reduces channel redundancy and forms interpretable functional clusters. Third, structural reconstruction loss and link-assisted loss stabilize long-range graph representation learning. Fourth, the lightweight splicing and normalization fusion method is less prone to overfitting on limited-scale neural decoding data. Fifth, this method is applicable to various brain-computer interface tasks such as writing, speech, and center-outward movement, exhibiting good generalization ability and noise resistance. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a schematic diagram of the overall process of the language brain-computer interface decoding method based on long-short-range functional connectivity graph neural network of the present invention.

[0019] Figure 2 This is a schematic diagram illustrating the process of constructing a long-range functional connection graph and a learnable channel pooling mechanism according to the present invention.

[0020] Figure 3 This invention provides a schematic diagram of the short-range functional connection graph and the time pooling process.

[0021] Figure 4 The bar chart shows the decoding performance comparison between the present invention and baseline methods on writing, speech, and center-out motion datasets, where an asterisk (*) indicates that the present invention's method has a significant improvement over other baseline methods.

[0022] Figure 5 This is a schematic diagram illustrating that the fusion representation learned by this invention has better class separability compared to the original representation. Detailed Implementation

[0023] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] It should be noted that, unless otherwise specified, the features in the following embodiments and implementation methods can be combined with each other.

[0025] The datasets used in this example include the handwriting neural decoding dataset from the 2021 Nature paper "High-performance brain-to-text communication via handwriting," the speech neural decoding dataset from the 2023 Nature paper "A high-performance speech neuroprosthesis," and the Center-Out dataset for the center-to-out motion decoding task. The neural activity in the handwriting dataset was collected by two sets of microelectrode arrays implanted in the motor cortex, containing a total of 192 recording channels. During the experiment, participants imagined their non-paralyzed hand writing corresponding characters based on prompts displayed on a computer screen. The task categories included 26 English letters and several punctuation marks or symbols, totaling 31 categories. This example uses the Handwriting I subset dataset, which contains 837 trial samples. The neural activity of the speech dataset was also collected by two sets of microelectrode arrays implanted in the motor cortex, each array containing 64 channels, for a total of 128 recording channels. During the experiment, participants attempted to speak corresponding words or perform a "silent" condition based on screen prompts. The task categories included 39 word categories and 1 silence category, for a total of 40 categories. The speech subset used in this example contains 800 trial samples. The center-out dataset was used to verify the applicability of this method in the motion direction decoding task. Its neural activity was collected by a 96-channel Utah array implanted in the primary motor cortex of macaques. During the experiment, the macaques needed to control a joystick to move a red cursor on the display screen from an initial central position to one of eight specified directional targets on the periphery. Therefore, this dataset contains 8 categories of center-out motion direction tasks. This example uses the Center-Out (M) II subset dataset, which contains 243 trial samples.

[0026] like Figure 1 As shown, a language brain-computer interface decoding method based on long-range and short-range functional connectivity graph neural networks mainly includes five steps: preprocessing, long-range graph construction, short-range graph construction, graph representation learning, and fusion feature decoding. The input can be a spike sequence recorded by an invasive microelectrode array, or a firing rate sequence or other neural activity features calculated from the spike sequence. The output can be a discrete behavior category, or a speech or writing intention category.

[0027] (1) Neural signal preprocessing.

[0028] Suppose the original neural signal contains C recording channels. After aligning the timestamps according to behavioral cues, start of movement, pronunciation cues, start of writing, or other events, the effective time segment of each trial is extracted. Trials that are too short are padded with zeros, and trials that are too long are truncated or segmented using a sliding window to obtain a neural signal matrix of uniform length T. .

[0029] In one specific embodiment, the peak data can be binarized or the distribution rate can be statistically calculated over a fixed time width, and then smoothed using a Gaussian kernel or moving average. Each channel can be standardized using the z-score method, so that the mean of each channel is 0 and the variance is 1. After preprocessing, the trials are divided into training set, validation set and test set in an 8:1:1 ratio.

[0030] (2) Construction of long-range graph.

[0031] Long-range plots aim to characterize the overall stable channel relationships over a single trial. For a neural signal matrix X, take the [missing information - likely a specific plot or plot]... Full-time sequence of each channel and the Full-time sequence of each channel Calculate the Pearson correlation coefficient between the two. The correlation matrix R is obtained by calculating the correlation between all channels pairwise.

[0032] The initial adjacency relationships in a long-range graph can be determined either by a fixed relevance threshold θ or by the Top-K relevance neighbors of each node. When using the Top-K strategy, the K neighbors with the highest relevance coefficients for each node are retained. This avoids the complete isolation of some nodes when the overall relevance level is low, making it more suitable for migration between different numbers of channels and different task data.

[0033] ; ,like or ;otherwise ; in, The initial adjacency weights for the long-range graph. For channel With channel The correlation coefficient between them. For the retained edges, this invention further learns nonlinear edge weights. Specifically, the time series of the two channels are concatenated and input into a shared multilayer perceptron. Obtain the edge weights between 0 and 1. and with Together they form a long-range weighted adjacency matrix: ; ; in, This represents a stable long-range functional connection between two channels.

[0034] The aforementioned long-range weighted adjacency matrix not only preserves linear correlations but also supplements nonlinear channel interaction information through learnable edge weights. For complex tasks such as writing and speech, long-range graphs can encode task intent, global semantics, or stable motion planning information that persists throughout the entire trial.

[0035] In long-range graph construction, learnable channel pooling can be further implemented. For example... Figure 2 As shown, due to the overlapping responses of adjacent channels, low-quality channels, and functional redundancy in invasive electrode arrays, this invention incorporates a learnable channel pooling module on the long-range graph. Specifically, a graph neural network is used to generate a soft-assignment matrix. S in the first The line represents the first The probability distribution of each neural channel belonging to a functional cluster, where K is the upper limit of the number of functional clusters. Matrix multiplication yields the pooled node features and adjacency matrix. This pooling method allows the model to automatically discover functionally similar neural channel groups during end-to-end training. ; ; in, This represents the graph neural network used to generate the soft assignment matrix, where H is the node feature matrix of the long-range functional connectivity graph. For pooling node features, That is, the adjacency matrix of the long-range pooling graph.

[0036] To prevent the pooling process from disrupting the original long-range connectivity structure, this invention introduces a structure reconstruction loss. This loss constrains the pooling graph to maintain the main topological structure of the original graph by backprojecting the pooled graph back into the original channel space and comparing it with the original long-range adjacency matrix. ; in, This represents the Frobenius norm structural loss.

[0037] Furthermore, this invention also introduces link-assisted loss. Specifically, a link score matrix is ​​constructed using long-range graph node embedding H. Using the initial adjacency relationship as a supervision signal, a binary cross-entropy loss with logical values ​​is applied to constrain connected nodes to have higher similarity and non-connected nodes to have lower similarity. ; in, regarded as a node With nodes The probability of an edge connecting them is log odd. Represents the standard binary cross-entropy with logarithmic odds.

[0038] The two regularization terms mentioned above can improve the structural stability and separability of long-range branch representations.

[0039] (3) Construction of short-range graphs.

[0040] Long-range plots are insufficient for depicting rapidly changing local collaborative activities; therefore, this invention further establishes a sequence of short-range plots. For example... Figure 3 As shown, firstly, based on the window length and step length By sliding partitioning X, we obtain M short-time submatrices. Each short-time submatrix covers the same set of channels, but only contains activities within a local time range.

[0041] Calculate the channel in the m-th window. With channel The proportion of simultaneous activation is used as the co-activation score. .when When the value is greater than the preset threshold τ, connect channel i and channel j in the m-th short plot: ; in, For indicator functions. If ,but Otherwise, it is 0.

[0042] Node features may include the average firing rate within the window, total firing count, energy, variance, or features encoded by a shared multilayer perceptron.

[0043] Short-range maps reflect rapidly occurring neural coordination activities within a local window. For example, in writing tasks, short-range maps can capture local dynamics at the beginning, turning, and ending stages of a stroke; in speech tasks, short-range maps can capture instantaneous changes in phonemes and articulation movements.

[0044] Furthermore, short-range graph construction also includes short-range graph temporal pooling. To reduce the length of the short-range graph sequence, improve efficiency, and preserve significant transient connections, this invention merges two adjacent short-range graphs into a single temporal pooling graph. During merging, node features are averaged, and the adjacency matrix is ​​maximized element-wise. This preserves significant co-activation connections that have appeared in any window within a slightly wider temporal context, while reducing the computational cost of subsequent graph neural networks. The specific formula is as follows: ; ; ; in, It is a simplified representation of the short-term dynamic characteristics of neurons, which complements the long-range graph by encoding rapidly changing local structures.

[0045] (4) Graph representation learning.

[0046] The long-range and short-range branches employ a graph sampling and aggregation network (GraphSAGE), respectively. GraphSAGE, by sampling and aggregating neighboring nodes, can stably extract graph representations from neural data with a large number of channels and graph structure varying with trials. The long-range branch takes a pooled long-range graph as input and outputs long-range graph node representations; the short-range branch takes a sequence of time-pooled short-range graphs as input, extracts node representations for each short-range graph, and then averages them over time. ; ; Then, mean pooling is performed at each node of each branch to obtain a vector of fixed dimensions: ; ; in, and This is used to characterize the final long-range and short-range plots.

[0047] (5) Fusion feature decoding.

[0048] The long-range and short-range plots have been obtained through the above steps. and In this step, the original signal branch is obtained through a multilayer perceptron. Then, the three are spliced ​​together and subjected to layer normalization to obtain the fused neural representation. : ; in, Representative layer normalization steps.

[0049] A lightweight fusion approach using splicing and layer normalization can calibrate three types of representation scales without significantly increasing the number of parameters, thus reducing the risk of overfitting on small-scale neural decoding data.

[0050] In the fusion feature decoding, the fusion neural representations are... Inputting the data into a multilayer perceptron decoder yields the prediction result. : ; For classification tasks, the decoder outputs the probability distributions for each category, and cross-entropy loss is used for supervised training. ; in, This represents the cross-entropy loss operator.

[0051] The overall loss during the training phase is obtained by weighted summation of the downstream task loss, the structural reconstruction loss, and the link-assisted loss: ; Among them, weight and The appropriate optimization method can be selected based on the performance on the validation set. Training can employ the Adam optimizer, cosine annealing learning rate scheduling, dropout regularization, and early stopping strategies to improve model stability.

[0052] (6) Model training and testing.

[0053] In a preferred embodiment, the number of Top-K neighbors K in the long-range graph is 2 to 4; the number of long-range pooling clusters can be set to 64, 128, or 256 depending on the number of channels; and the short-range graph co-activation threshold... The value is set to 0.8 to 0.85; the short-range window length w is set to 2 to 3 time units, and the step size s is set to ⌊w / 2⌋. The above parameters are preferred settings and do not constitute a limitation on the scope of protection of this invention. Those skilled in the art can adjust them according to the sampling rate, trial length, and task type.

[0054] The model training and testing process includes the following steps. First, all trials in each dataset are divided into training, validation, and test sets according to trial level, in a ratio of 8:1:1. Each trial belongs to only one set to avoid overlap between training and test samples. Second, the neural signals of each trial in the training set are preprocessed to obtain a neural signal matrix X, and then labeled according to the corresponding task. In the handwriting task, the labels are 31 character categories; in the speech task, the labels are 40 speech or silent intention categories; and in the center-to-outward motion task, the labels are 8 motion directions.

[0055] During the training phase, for each trial in a mini-batch, the model dynamically constructs long-range and short-range functional connectivity graphs based on the current trial. The long-range graph is obtained from channel correlations across all trials and then processed by learnable channel pooling to form a long-range pooled graph. The short-range graph is obtained from channel co-activation relationships within a sliding time window and significant transient connections are preserved through temporal pooling of adjacent windows. Subsequently, the long-range and short-range graphs are input into the GraphSAGE network to obtain long-range and short-range graph representations, respectively; the original neural signal is then processed by a multilayer perceptron to obtain the original signal representation. The three representations are concatenated and normalized by layers to obtain a fused neural representation, which is finally input into the downstream multilayer perceptron decoder to output the class prediction result.

[0056] The training objective is composed of downstream classification cross-entropy loss, structural reconstruction loss, and link-assisted loss: ; in, Used to constrain the class prediction results to be consistent with the true labels. Used to constrain long-range pooling graphs to maintain the main topology of the original long-range graph. This is used to enhance the ability of node embeddings to express channel connectivity relationships. The Adam optimizer can be used during training, with an initial learning rate set to... The batch size was set to 128, the maximum number of training epochs was set to 300, and cosine annealing learning rate scheduling, Dropout regularization, and early stopping strategy on the validation set were combined. The validation set was used to select model hyperparameters and the optimal number of training epochs, while the test set was only used for final performance evaluation and did not participate in model training or parameter selection.

[0057] To verify the effectiveness of the method of this invention, it was compared with methods such as SVM, GRU, Topology, ModernTCN, AMAG, and iTransformer under the same dataset and training and testing settings. The results are as follows. Figure 4 As shown, the method achieved an accuracy of 96.1% ± 0.42% on the Handwriting I dataset, higher than the best baseline SVM's 92.8% ± 0.71%, based on the mean accuracy ± standard deviation of five randomized experiments. On the Speech dataset, the method achieved an accuracy of 75.6% ± 0.84%, higher than the best baseline AMG's 63.3% ± 0.91%. On the Center-Out (M) II dataset, the method achieved an accuracy of 89.2% ± 0.63%, higher than the best baseline ModernTCN's 79.4% ± 0.70%. These results demonstrate that the present invention, by jointly modeling long-range stable functional connections and short-range dynamic co-activation, can achieve higher classification accuracy in writing, speech, and motion decoding tasks.

[0058] Furthermore, this invention also verifies the effectiveness of the fused representations. Taking the handwriting task as an example, the classification accuracy on the Handwriting I dataset is 92.2% ± 0.92% when using only the original signal representation; after fusing the original signal representation, long-range graph representation, and short-range graph representation, the accuracy improves to 93.3% ± 1.30%. On the more challenging Handwriting II dataset, the accuracy using only the original signal representation is 67.6% ± 2.44%, while the accuracy of the fused representation improves to 83.1% ± 2.67%. These results indicate that the long-range graph branch and the short-range graph branch can provide complementary information to the original neural signal, thereby improving the class discrimination ability of the neural representation.

[0059] like Figure 5 As shown, the fused representation learned by this invention has clearer category boundaries and a more compact intra-class distribution compared to the original representation. For letters with similar writing trajectory shapes, such as g and s, h and n and r, x and y, inter-class confusion is more likely to occur in the original representation; however, after fusing long-range and short-range graph information, the separation degree of different character categories in the low-dimensional representation space is improved, and off-diagonal errors in the confusion matrix are reduced. This phenomenon indicates that this invention can not only improve the accuracy of neural decoding, but also form more discriminative and interpretable neural representations.

[0060] The embodiments described above provide a detailed explanation of the technical solutions and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the present invention. Any modifications, additions, and equivalent substitutions made within the scope of the principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A language brain-computer interface decoding method based on long-short-range functional connectivity graph neural networks, characterized in that, Includes the following steps: (1) Obtain the raw neural signals of the brain-computer interface task, and obtain the neural signal matrix after preprocessing the raw neural signals. , where C represents the number of recording channels or neurons, and T represents the time step; (2) Based on the neural signal matrix X, a long-range functional connectivity graph is constructed, the stable correlation between channels is calculated within the entire trial range, and a long-range pooling graph is obtained through the channel pooling mechanism. This long-range pooling graph is constrained by structural reconstruction loss and link-assisted loss. (3) Divide the neural signal matrix X into multiple short time windows with overlapping parts, construct a short-range functional connectivity graph sequence based on the co-activation relationship of neurons in each window, and perform average time pooling on adjacent short-range functional connectivity graphs. (4) Graph representation learning is performed on the long-range pooled graph and the pooled short-range functional connection graph sequence using graph sampling and aggregation networks respectively, to obtain long-range graph representation and short-range graph representation; (5) The long-range graph representation, the short-range graph representation, and the representation extracted from the original neural signal are spliced ​​and normalized to obtain the fused neural representation, which is then input into the downstream decoder and outputs at least one of the following categories: motor behavior category, speech intention category, written character category, or cognitive state category.

2. The language brain-computer interface decoding method based on long-short-range functional connectivity graph neural networks according to claim 1, characterized in that, In step (2), a long-range functional connectivity graph is constructed based on the neural signal matrix X, specifically including: For the neural signal matrix X, take the first... Full-time sequence of each channel and the Full-time sequence of each channel Calculate the Pearson correlation coefficient between the two. Based on relevant thresholds Or the top-K related neighbors of each node determine the graph topology, when Greater than the threshold or channel It belongs to the channel When collecting Top-K related neighbors, in the channel With channel Establish edges between them.

3. The language brain-computer interface decoding method based on long-short-range functional connectivity graph neural networks according to claim 2, characterized in that, In step (2), the stable correlation between channels is calculated across the entire trial range, specifically including: While maintaining the topology of the long-range functional connectivity graph, the time series data of two connected channels are concatenated and input into a shared multilayer perceptron to obtain learnable edge weights. and will and Together as a long-range weighted adjacency matrix The weights of the corresponding edges are calculated using the following formula: ； ； in, This represents a shared multilayer perceptron. That is, the channel and channels Stable long-range functional connections between them.

4. The language brain-computer interface decoding method based on long-short-range functional connectivity graph neural networks according to claim 3, characterized in that, In step (2), the channel pooling mechanism includes: Generating soft-assignment matrices using graph neural networks Where K is the preset upper limit of the number of functional clusters; according to Obtain pooled node features ,in The node feature matrix of the long-range functional connectivity graph; based on Obtain the adjacency matrix of the long-range pooling graph ,in For stable long-range functional connections between two channels.

5. The language brain-computer interface decoding method based on long-short-range functional connectivity graph neural networks according to claim 4, characterized in that, In step (2), the structural reconstruction loss The link-assisted loss is used to constrain long-range pooling graphs to maintain the original long-range graph topology. The loss function used to constrain the predictive ability of edge existence for node embedding pairs is as follows: ； ； in, regarded as a node With nodes The probability of an edge between two points is log-odd, which is determined by... Calculated; Represents the standard binary cross-entropy with logarithmic odds; For nodes , Long-range functional connections between them; This represents the Frobenius norm structural loss.

6. The language brain-computer interface decoding method based on long-short-range functional connectivity graph neural networks according to claim 1, characterized in that, In step (3), constructing the short-range functional connectivity graph specifically includes: The neural signal matrix X is divided into multiple short-time sub-matrices by sliding partitioning with window length w and step size s; for any short-time sub-matrix m, the channels are calculated. and channels The frequency of simultaneous activation within the corresponding window is used as the co-activation score. And determine the edges of the short-range functional connectivity graph based on the co-activation threshold τ; Node features are extracted from each short-time submatrix m using average firing rate, total pulse count, and energy statistics. These features are then input into a shared multilayer perceptron encoder to generate node features. ,in, For encoder dimensions.

7. The language brain-computer interface decoding method based on long-short-range functional connectivity graph neural networks according to claim 6, characterized in that, In step (3), average time pooling is performed on the adjacent short-range function connectivity graph, specifically including: Time pooling connects two adjacent short-range functions in the graph. The node features are averaged, and the adjacency matrix is ​​maximized element by element, so that the pooled short-range functional connection graph inherits significant short-range connections in adjacent time windows.

8. The language brain-computer interface decoding method based on long-short-range functional connectivity graph neural networks according to claim 1, characterized in that, In step (5), the downstream decoder is a multilayer perceptron that includes nonlinear activation ReLU and dropout regularization, and the training objective is a weighted sum of cross-entropy loss, structural reconstruction loss and link-assisted loss.

Images