FNIRS brain function state decoding method and system based on graph information bottleneck

Abstract

The invention discloses an FNIRS brain function state decoding method and system based on a graph information bottleneck, and belongs to the technical field of brain-computer interface and neural signal processing, and the method mainly comprises the steps: constructing a brain network graph based on a channel space position of an fNIRS data segment; a data segment is input into a deep learning model which sequentially comprises a graph convolutional network (GCN) module, a graph information bottleneck (GIB) module and a time sequence processing module, a unique cascade processing flow of'spatial modeling-point-by-point compression-time modeling 'is provided, and an information bottleneck mechanism is embedded between spatial feature extraction and time feature extraction. And carrying out instant compression and regularization on the spatial representation of each time point. Through a space-time separation processing flow and embedding a point-by-point information bottleneck, fNIRS space-time feature representation which is most useful and robust for a decoding task can be effectively learned, and the decoding accuracy and the model generalization ability are remarkably improved.

Description

Technical Field

[0001] This invention relates to the fields of brain-computer interface and neural signal processing technology, and in particular to a method and system for decoding FNIRS brain functional states based on graph information bottleneck. Background Technology

[0002] Functional near-infrared spectroscopy (fNIRS) is a non-invasive brain imaging technique that indirectly reflects neural activity by measuring changes in hemoglobin concentration in the cerebral cortex. Due to its advantages such as high safety, portability, and strong resistance to motion interference, fNIRS has broad application prospects in fields such as brain-computer interfaces (BCI), cognitive neuroscience, clinical rehabilitation, and artificial intelligence.

[0003] However, fNIRS signals also suffer from low signal-to-noise ratios and are susceptible to physiological noise (such as heartbeats and respiration) and motion artifacts. These factors severely affect signal quality and the reliability of subsequent analysis. Traditional feature extraction methods often rely on manually designed time-domain, frequency-domain, or time-frequency-domain features, such as mean, variance, and power spectrum. These methods often depend on expert experience and are difficult to fully exploit the complex spatiotemporal patterns contained in fNIRS data.

[0004] In recent years, deep learning technology has demonstrated powerful capabilities in brain signal processing. In particular, Graph Convolutional Networks (GCNs) can effectively model and learn the spatial topological relationships between fNIRS channels, while Recurrent Neural Networks (RNNs) and their variants (such as LSTM and GRU) excel at capturing the temporal dependencies of captured signals. Models combining GCNs and RNNs offer a new approach to processing fNIRS data with spatiotemporal structure. However, learning feature representations from high-dimensional, noisy fNIRS signals that are both effective and have good generalization ability for decoding specific brain states remains a key challenge. Existing methods still have the following shortcomings: 1. Feature redundancy and overfitting: fNIRS data has high dimensionality and a lot of noise, making it easy for the model to learn redundant features that are irrelevant to the task, leading to overfitting and poor generalization ability.

[0005] 2. Insufficient fusion of spatiotemporal features: Most methods either extract spatial features first and then process the time series, or mix spatiotemporal features as input, failing to achieve true spatiotemporal separation and collaborative modeling.

[0006] 3. Lack of interpretability and compression mechanism: The model often lacks interpretability constraints on the feature representation, making it difficult to ensure that the learned features are both concise and discriminative.

[0007] Therefore, there is an urgent need for a brain functional state decoding method that can effectively integrate the spatiotemporal features of fNIRS, has strong generalization ability, and is interpretable. Summary of the Invention

[0008] The purpose of this invention is to overcome the problems existing in the prior art and to provide a method and system for decoding FNIRS brain functional states based on graph information bottlenecks.

[0009] The objective of this invention is achieved through the following technical solution: Firstly, a method for decoding FNIRS brain functional states based on graph information bottlenecks is provided, comprising the following steps: S1. Acquire functional near-infrared spectral data and preprocess the functional near-infrared spectral data to obtain an analysis data segment containing multiple channels and multiple time points; S2. Construct a brain network graph structure based on the spatial location information of the functional near-infrared spectral data channels, and define the connection relationships between nodes in the brain network graph structure; S3. Input the analysis data segment into a deep learning model, the deep learning model including a graph convolutional network module, a graph information bottleneck module and a time series processing module connected in sequence; S4. The node features at each time point in the analysis data segment are processed by the graph convolutional network module, and combined with the brain network graph structure, a probability distribution parameter is output for the instantaneous representation of each node; S5. The graph information bottleneck module samples the instantaneous representation of each node according to the probability distribution parameters through a reparameterization method to obtain the node bottleneck representation; S6. The time series processing module processes the time series composed of the bottleneck representations of all time points to extract the time series dynamic features; S7. Based on the aforementioned temporal dynamic features, the decoded brain functional state is output through a classifier.

[0010] Preferably, the functional near-infrared spectral data includes signals of changes in oxygenated hemoglobin concentration and / or changes in deoxygenated hemoglobin concentration.

[0011] Preferably, the preprocessing steps include bandpass filtering and data segmentation.

[0012] Preferably, the construction of the brain network graph structure includes: Based on the three-dimensional spatial coordinates between functional near-infrared spectral data channels, the K-nearest neighbor (KNN) algorithm is used to construct the coordinates.

[0013] Preferably, the probability distribution parameters include the mean and logarithmic standard deviation of a Gaussian distribution.

[0014] Preferably, the loss function of the graph information bottleneck module includes: The task loss term measures the difference between the classifier output and the true state. The regularization term is used to minimize the mutual information between the node bottleneck representation and the input data, while maximizing the mutual information between it and the brain functional state label.

[0015] Preferably, the regularization term is implemented by introducing a KL divergence term into the loss function. The KL divergence term calculates the difference between the posterior distribution of the node bottleneck representation and a preset prior distribution. The KL divergence loss term encourages the model to compress irrelevant information while retaining task-related information.

[0016] Preferably, the time series processing module is a long short-term memory network or a gated recurrent unit.

[0017] Preferably, before inputting the time series of the node bottleneck representation into the time series processing module, a spatial aggregation operation (such as global average pooling) is performed on the node bottleneck representation at each time point.

[0018] Secondly, a FNIRS brain functional state decoding system based on graph information bottleneck is provided, including: The data acquisition and preprocessing module is used to acquire functional near-infrared spectral data and preprocess the functional near-infrared spectral data to obtain an analysis data segment containing multiple channels and multiple time points. A brain network graph construction module is used to construct a brain network graph structure based on the spatial location information of the functional near-infrared spectral data channels, and to define the connection relationships between nodes in the brain network graph structure. The model processing module is configured with a deep learning model trained according to the method of any one of claims 1-9. The deep learning model includes a graph convolutional network module, a graph information bottleneck module, and a time series processing module connected in sequence. The graph convolutional network module processes the node features at each time point in the analysis data segment and, in conjunction with the brain network graph structure, outputs a probability distribution parameter for the instantaneous representation of each node. The graph information bottleneck module samples the instantaneous representation of each node using a reparameterization method based on the probability distribution parameter to obtain the node bottleneck representation. The time series processing module processes the time series composed of the node bottleneck representations at all time points to extract temporal dynamic features. The result output module is used to output the decoded brain functional state based on the temporal dynamic features through a classifier.

[0019] It should be further noted that the technical features corresponding to the above options can be combined or substituted to form new technical solutions if there is no conflict.

[0020] Compared with the prior art, the beneficial effects of the present invention are: 1. Achieved more robust and generalized feature learning, effectively suppressing noise and overfitting: This invention introduces a graph information bottleneck module. This module samples the instantaneous representation of each node using a reparameterization method to obtain the node bottleneck representation. This module forces the model to learn a "compressed" node bottleneck representation. Through the KL divergence regularization term in its loss function, it explicitly constrains the model to retain information relevant to brain functional state labels while helping to remove redundant information and noise (such as physiological noise and micro-motion artifacts) from the fNIRS signal. This is equivalent to embedding a data-driven, dynamic noise filter into the model, improving the robustness of features from an information theory perspective, significantly reducing the risk of model overfitting, and enabling it to exhibit stronger generalization ability on unseen test data.

[0021] 2. Achieved refined and decoupled fusion of spatiotemporal features: This invention employs a unique cascaded architecture of "Spatial Modeling (GCN) - Pointwise Compression (GIB) - Temporal Modeling (LSTM / GRU)". This design decouples the complex spatiotemporal feature learning process. The GCN module is specifically responsible for capturing the spatial topological features at each time point; the GIB module then performs real-time compression and purification of these spatial features; the processed purified spatial features are then fed into the time series module to capture temporal dynamic features. By combining structures such as GCN and LSTM, the spatial topological and temporal dynamic information of fNIRS data can be effectively fused, enabling precise decoding of brain functional states. This step-by-step processing method avoids feature confusion that may be caused by mixed spatiotemporal information input, and can fuse the inherent spatiotemporal two-dimensional information in fNIRS signals more precisely and efficiently than traditional methods.

[0022] 3. An end-to-end automated feature extraction framework is provided: The method of this invention begins with the preprocessing of the original fNIRS signal and continues until the final classification output, forming a complete end-to-end learning framework. This framework eliminates the need for complex and tedious manual feature engineering (such as calculating the mean, variance, and slope of the HbO / HbR signals for each channel); all features are automatically learned by the model through gradient descent. This significantly reduces the reliance of the analysis method on domain expert experience, improves the method's versatility and repeatability, and also avoids information loss due to improper manual feature design.

[0023] 4. Improved the interpretability and reliability of the model: Through the GIB module, the bottleneck representations learned by the model exhibit a clear probability distribution (e.g., a Gaussian distribution). By analyzing the differences in the distribution of bottleneck representations under different brain states (e.g., changes in the mean μ), we can qualitatively understand which brain regions or channels the model considers more important in distinguishing different states. This provides some interpretability to the model's decision-making process, enhancing the reliability and credibility of its results in neuroscience research and clinical applications.

[0024] 5. It balances performance and computational efficiency: The GIB module in this invention uses reparameterization techniques for sampling, ensuring that the entire model (including the sampling process) remains differentiable and can be trained end-to-end using the standard backpropagation algorithm without introducing additional training complexity. Furthermore, optional spatial aggregation operations (such as global average pooling) effectively reduce the input dimensionality of subsequent time series processing modules, decreasing the number of model parameters and computational overhead with almost no performance loss, thereby improving training and inference efficiency. Attached Figure Description

[0025] Figure 1 This is a flowchart of a method for decoding FNIRS brain functional states based on graph information bottlenecks according to the present invention. Figure 2 This is a schematic diagram of the network structure of the GIB-GCN-LSTM deep learning model used in this invention; Figure 3 This is an example diagram illustrating the effect of the fNIRS signal preprocessing (filtering) of the present invention. Detailed Implementation

[0026] The technical solution 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, not all embodiments. The components of the embodiments of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. 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.

[0027] It should be noted that the defects in the solutions in the prior art are all the results of the inventors' practice and careful research. Therefore, the discovery process of the above problems and the solutions proposed by the embodiments of this application in the following text should be the inventors' contributions to this application in the process of invention and creation, and should not be understood as technical content known to those skilled in the art.

[0028] Reference Figure 1In one exemplary embodiment, a method for decoding FNIRS brain functional states based on graph information bottlenecks is provided, comprising the following steps: S1. Acquire functional near-infrared spectral data and preprocess the functional near-infrared spectral data to obtain an analysis data segment containing multiple channels and multiple time points; S2. Construct a brain network graph structure based on the spatial location information of the functional near-infrared spectral data channels, and define the connection relationships between nodes in the brain network graph structure; S3. Input the analysis data segment into a deep learning model, the deep learning model including a graph convolutional network module, a graph information bottleneck module and a time series processing module connected in sequence; S4. The node features at each time point in the analysis data segment are processed by the graph convolutional network module, and combined with the brain network graph structure, a probability distribution parameter is output for the instantaneous representation of each node; S5. The graph information bottleneck module samples the instantaneous representation of each node according to the probability distribution parameters through a reparameterization method to obtain the node bottleneck representation; S6. The time series processing module processes the time series composed of the bottleneck representations of all time points to extract the time series dynamic features; S7. Based on the aforementioned temporal dynamic features, the decoded brain functional state is output through a classifier.

[0029] For example, in step S1, fNIRS data are acquired when the subject performs different cognitive tasks (e.g., motor imagery task A and resting state B). The fNIRS device used records raw optical density data from multiple channels (e.g., 36 channels) at a sampling rate of 10 Hz. The raw optical density data is converted into a time series of changes in oxygenated hemoglobin (HbO) and deoxygenated hemoglobin (HbR) concentrations using a modified Lambert-Beer law.

[0030] Preprocessing of HbO and HbR signals: Bandpass filtering: Apply a Butterworth bandpass filter (e.g., 0.01 Hz–0.5 Hz, 4th order) to remove physiological noise (such as heartbeat, breathing) and low-frequency baseline drift.

[0031] Data Segmentation (Epoching): Based on task event markers (e.g., task start signal), extract signal data segments (Epochs) within a fixed time window before and after each event. For example, extract 10-second Epochs starting from the moment the event occurs. The data shape of each Epoch is [number of time points, number of channels, number of features] (e.g., [100, 36, 2], where 2 represents the two signal features HbO and HbR).

[0032] For example, in step S2, the three-dimensional spatial coordinates of the 36 channels in the fNIRS probe (e.g., obtained from the mnt.mat file) are used. Based on these coordinates, the Euclidean distance between the channels is calculated. The adjacency relationships of the graph are constructed using the K Nearest Neighbors (KNN) algorithm. For example, for each channel (node), it is connected to the K spatially nearest other channels (e.g., K=5), generating the edge index of the graph, edge_index. edge_index is a tensor of shape [2, number of edges], representing the connection between nodes.

[0033] For example, in step S3, the deep learning model is named the GIB-GCN-LSTM model, and its structure is as follows: Figure 2 As shown. The model executes steps S4-S7 sequentially, as follows: 1. Model input: Preprocessed fNIRS Epoch data X (shape [batch size, number of time points, number of nodes, number of input features], e.g. [B, 100, 36, 2]) and globally shared edge_index.

[0034] 2. GCN encoder and information bottleneck parameter output: For each time point t in the input epoch, the node features Xt (shape [batch size, number of nodes, number of input features]) are first processed through one or more graph convolutional layers (GCNConv). For example, the first GCN (gcn1) maps the input features to the hidden feature space: H(1)t = ReLU(BN(GCNConv1(Xt,edge_index))). Where BN is batch normalization.

[0035] Then, the mean µt (shape [batch size, number of nodes, bottleneck_dim]) and log-standard deviation logσt (shape [batch size, number of nodes, bottleneck_dim]) of the bottleneck representation of each node are calculated from H(1)t (or features processed by more GCN layers) using two sets of parallel GCN layers (gcn_mu, gcn_logstd). bottleneck_dim is the preset bottleneck dimension.

[0036] Information bottleneck sampling (reparameterization): The bottleneck representation Zt is obtained by sampling from the learned Gaussian distribution N(µt,σ2t) using the reparameterization technique: Zt =µt+ϵ·exp(0.5·logσt), where ϵ is noise sampled from the standard normal distribution N(0,I). The shape of Zt is [batch size, number of nodes, bottleneck_dim].

[0037] 3. Spatial aggregation (optional): To reduce the input dimensionality of subsequent LSTMs, the bottleneck representation Zt at each time point can be spatially aggregated. For example, global average pooling can be used: Z̄t=mean(Zt,dim=1), resulting in a graph-level bottleneck representation with shape [batch size, bottleneck_dim].

[0038] 4. LSTM processes timing information: All time-level bottleneck representations Z̄t (or unaggregated Zt after appropriate transformation) are formed into a sequence [Z̄1, Z̄2, ..., Z̄T] (shape [batch size, number of time points, bottleneck_dim]), and input into an LSTM network.

[0039] The LSTM network learns the temporal dependencies of the sequence and outputs its hidden states. Typically, the hidden state hT(shape[batch size, rnn_hidden_channels]) of the last time step of the LSTM is used as the spatiotemporal feature representation of the entire epoch.

[0040] 5. Classifier: The output hT of the LSTM is passed through one or more fully connected layers (FC) and a softmax activation function (usually included in the loss function) to obtain the classification probabilities of different brain functional states.

[0041] Furthermore, the model's total loss function Ltotal consists of two parts: Ltotal = Ltask + β · LGIB Here, Ltask is the task-related loss, typically the cross-entropy loss function for classification tasks, calculating the difference between the model's prediction and the true label. LGIB is the graph information bottleneck loss, usually the sum or average of the KL divergences between the posterior distribution q(Z|X) learned by all nodes at all time points (defined by µ and logσ output by the GCN) and the pre-defined prior distribution p(Z) (usually a standard normal distribution N(0,I)). Its calculation formula is: LGIB = KL(q(Z|X)||p(Z)) = 0.5∑(µ² +σ² −log(σ²)−1). β is a hyperparameter used to balance task performance and the degree of information compression.

[0042] Use gradient descent optimization algorithms (such as Adam) to minimize the total loss function Ltotal and iteratively update the model parameters.

[0043] After training, the model's decoding accuracy, precision, recall, F1 score, and other metrics are evaluated using an independent test dataset. The trained model can be used to decode brain functional states corresponding to new fNIRS data in real time or offline.

[0044] For example, on the fNIRS dataset that distinguishes between two different motion imagery tasks, the method based on the GIB-GCN-LSTM model proposed in this embodiment shows advantages in classification accuracy and model generalization compared to traditional machine learning methods (such as SVM) and the GCN-LSTM model without GIB. Figure 3 The effect of a typical fNIRS signal after being filtered by a 0.01-0.5Hz bandpass filter is shown, demonstrating that baseline drift and high-frequency noise are effectively removed.

[0045] The above detailed embodiments are a description of the present invention. It should not be considered that the specific embodiments of the present invention are limited to these descriptions. For those skilled in the art, several simple deductions and substitutions can be made without departing from the concept of the present invention, and all of these should be considered to fall within the protection scope of the present invention.

Claims

1. A method for decoding FNIRS brain functional states based on graph information bottleneck, characterized in that, Includes the following steps: S1. Acquire functional near-infrared spectral data and preprocess the functional near-infrared spectral data to obtain an analysis data segment containing multiple channels and multiple time points; S2. Construct a brain network graph structure based on the spatial location information of the functional near-infrared spectral data channels, and define the connection relationships between nodes in the brain network graph structure; S3. Input the analysis data segment into a deep learning model, the deep learning model including a graph convolutional network module, a graph information bottleneck module and a time series processing module connected in sequence; S4. The node features at each time point in the analysis data segment are processed by the graph convolutional network module, and combined with the brain network graph structure, a probability distribution parameter is output for the instantaneous representation of each node; S5. The graph information bottleneck module samples the instantaneous representation of each node according to the probability distribution parameters through a reparameterization method to obtain the node bottleneck representation; S6. The time series processing module processes the time series composed of the bottleneck representations of all time points to extract the time series dynamic features; S7. Based on the aforementioned temporal dynamic features, the decoded brain functional state is output through a classifier.

2. The FNIRS brain functional state decoding method based on graph information bottleneck as described in claim 1, characterized in that, The functional near-infrared spectral data includes signals of changes in oxygenated hemoglobin concentration and / or changes in deoxygenated hemoglobin concentration.

3. The FNIRS brain functional state decoding method based on graph information bottleneck as described in claim 1, characterized in that, The preprocessing steps include bandpass filtering and data segmentation.

4. The FNIRS brain functional state decoding method based on graph information bottleneck as described in claim 1, characterized in that, The construction of the brain network graph structure includes: Based on the three-dimensional spatial coordinates between functional near-infrared spectral data channels, the K-nearest neighbor algorithm is used to construct the coordinates.

5. The FNIRS brain functional state decoding method based on graph information bottleneck as described in claim 1, characterized in that, The probability distribution parameters include the mean and logarithmic standard deviation of the Gaussian distribution.

6. The FNIRS brain functional state decoding method based on graph information bottleneck as described in claim 1, characterized in that, The loss function of the graph information bottleneck module includes: The task loss term measures the difference between the classifier output and the true state. The regularization term is used to minimize the mutual information between the node bottleneck representation and the input data, while maximizing the mutual information between it and the brain functional state label.

7. The FNIRS brain functional state decoding method based on graph information bottleneck as described in claim 6, characterized in that, The regularization term is implemented by introducing a KL divergence term into the loss function, which calculates the difference between the posterior distribution of the node bottleneck representation and a preset prior distribution.

8. The FNIRS brain functional state decoding method based on graph information bottleneck as described in claim 1, characterized in that, The time series processing module is a long short-term memory network or a gated recurrent unit.

9. The FNIRS brain functional state decoding method based on graph information bottleneck as described in claim 1, characterized in that, Before inputting the time series representation of the node bottleneck into the time series processing module, a spatial aggregation operation is performed on the node bottleneck representation at each time point.

10. A FNIRS brain functional state decoding system based on graph information bottleneck, characterized in that, include: The data acquisition and preprocessing module is used to acquire functional near-infrared spectral data and preprocess the functional near-infrared spectral data to obtain an analysis data segment containing multiple channels and multiple time points. A brain network graph construction module is used to construct a brain network graph structure based on the spatial location information of the functional near-infrared spectral data channels, and to define the connection relationships between nodes in the brain network graph structure. The model processing module is configured with a deep learning model trained according to the method of any one of claims 1-9. The deep learning model includes a graph convolutional network module, a graph information bottleneck module, and a time series processing module connected in sequence. The graph convolutional network module processes the node features at each time point in the analysis data segment and, in conjunction with the brain network graph structure, outputs a probability distribution parameter for the instantaneous representation of each node. The graph information bottleneck module samples the instantaneous representation of each node using a reparameterization method based on the probability distribution parameter to obtain the node bottleneck representation. The time series processing module processes the time series composed of the node bottleneck representations at all time points to extract temporal dynamic features. The result output module is used to output the decoded brain functional state based on the temporal dynamic features through a classifier.

Images