Idle state discrimination method based on state-dependent network dynamic analysis

Abstract

The invention discloses an idle state discrimination method based on dynamic analysis of a state-related network, and belongs to the technical field of brain-computer interfaces. According to the method, aiming at the problem of insufficient idle state detection precision in the asynchronous SSVEP brain-computer interface, high-precision state discrimination is realized by constructing a multi-layer dynamic network model. The method comprises the following steps: acquiring an occipital region electroencephalogram signal containing a task state and an idle state, and constructing an offline data set after baseline removal, downsampling, band-pass filtering and notch preprocessing; overlapping time windows are adopted to reconstruct time sequence signals, a TRCA template is combined to construct a state related network, a Louvain algorithm is utilized to perform module division, and four network features are extracted; and performing real-time classification on the control state and the idle state based on a random forest classifier. According to the method, by dynamically capturing the time-varying topological characteristics of the brain network, the intention state in a complex non-stationary signal is effectively distinguished, the idle state detection specificity and the system anti-interference capability are remarkably improved, and safe and reliable technical support is provided for asynchronous interaction scenes such as a brain-controlled robot.

Description

Technical Field

[0001] The present invention relates to the technical field of brain-computer interfaces, and particularly to a method for discriminating idle states based on dynamic analysis of state-related networks. Background Art

[0002] Brain-Computer Interface (BCI) technology enables human-computer interaction by directly decoding electrophysiological brain activities, providing a new rehabilitation approach for patients with motor disabilities caused by neuromuscular diseases (such as amyotrophic lateral sclerosis, spinal cord injury, etc.). In brain-controlled robot systems, Steady-State Visual Evoked Potential (SSVEP) has become a widely used electroencephalogram (EEG) paradigm due to its advantages such as high signal-to-noise ratio and high information transmission rate. Traditional SSVEP classification systems are divided into synchronous and asynchronous types. Among them, the asynchronous system does not require an external synchronization signal and allows users to trigger commands by gazing at the stimulus source at any time, which is closer to the actual application requirements. However, the core challenge of asynchronous classification lies in real-time discrimination between the control state and the idle state. Existing methods have insufficient detection accuracy for the idle state, resulting in a high false trigger rate of the system and a risk of incorrect robot operation and even potential safety hazards.

[0003] The current bottleneck in asynchronous SSVEP classification mainly stems from two aspects: First, the idle state encompasses all brain activities during the user's non-control period (such as closing eyes to rest, limb movements, observing the environment, etc.), and its EEG patterns are complex and variable, making it difficult to establish a universal model. Second, EEG signals are affected by volume conduction effects and environmental noise interference, with low signal-to-noise ratio and significant inter / intra-individual differences. Especially in real scenarios, motion artifacts and device noise further exacerbate the non-stationarity of the signals. Traditional methods usually rely on fixed-frequency band energy or spatial domain features for state discrimination, but such static features are difficult to capture the dynamic changes of brain networks, resulting in a high misjudgment rate for the idle state. In addition, for brain-controlled systems for patients with motor disabilities, high robustness and low latency need to be considered. Existing technologies are difficult to meet the real-time and safety requirements of clinical applications due to insufficient feature representation capabilities.

[0004] Therefore, there is an urgent need for a new analysis method that can dynamically represent brain state changes and effectively distinguish between the control state and the idle state to improve the practical value and safety of asynchronous SSVEP brain-computer interfaces. Summary of the Invention

[0005] Aiming at the defects of the existing technology, the present invention provides a method for discriminating idle states based on dynamic analysis of state-related networks.

[0006] To achieve the above invention objectives, the technical solutions adopted by the present invention are as follows:

[0007] A method for discriminating the idle state based on the dynamic analysis of state-related networks, comprising the following steps:

[0008] Step 1: Collect the offline electroencephalogram (EEG) signal data of the subject under the SSVEP paradigm, and construct an occipital offline data set including task states and idle states;

[0009] Step 2: Process the offline data set by using the state-related dynamic network analysis method, and extract the network topological features of the task state and the idle state, including:

[0010] Perform temporal reconstruction on the EEG signal through overlapping time windows to construct a multi-layer dynamic network;

[0011] Construct a state-related brain network by using the TRCA template to capture the connection patterns between brain regions in different states;

[0012] Use the Louvain algorithm to partition the network into modules, and extract the network topological features that change over time.

[0013] Step 3: Train a random forest classifier based on the network topological features to achieve real-time discrimination of the SSVEP control state and the idle state.

[0014] Further, the construction of the offline data set in Step 1 specifically includes: arranging multiple electrodes in the occipital region to collect EEG signals, with a sampling rate not lower than 250 Hz;

[0015] Perform baseline removal, band-pass filtering and 50 Hz notch filtering on the original data, and downsample it to 250 Hz;

[0016] Intercept a continuous 11-second data segment including the SSVEP task execution period and the idle period, and organize it into a multi-dimensional data set according to trials.

[0017] Further, the setting of the overlapping time windows in Step 2 satisfies: the window length is 1 - 2.5 seconds, and the overlapping time is 0.1 - 1 second;

[0018] Generate a personalized temporal segmentation scheme by combining different window lengths and overlapping times.

[0019] Further, the construction of the state-related brain network by using the TRCA template specifically includes:

[0020] Select the data of the first time window extracted from the task period of each trial in the template data set as the TRCA template, and use task-related component analysis (TRCA) to calculate the TRCA model corresponding to the target.

[0021] Calculate the network connection matrix of the multi-channel signals in each time window under different TRCA templates and models.

[0022] Furthermore, the network topology features extracted include: global participation coefficient, global gateway coefficient, global diversity coefficient, and global efficiency. For different states, the four types of features within the time window of each state are combined to obtain the topology feature change trend of each state across the time window.

[0023] Furthermore, the classifier in step 3 is a random forest classifier, and the training process includes:

[0024] Ten-fold cross validation was used to divide the training set and test set;

[0025] Optimize classifier hyperparameters through grid search;

[0026] Nine categories of modeling are performed for eight types of control instruction states and idle states.

[0027] Furthermore, the construction of the multi-layer dynamic network includes:

[0028] The network connection matrices of consecutive time windows are stacked in time order;

[0029] Analyze the time-varying patterns of network properties and capture the critical characteristics of state transitions.

[0030] Furthermore, the idle state determination criteria include:

[0031] The network topology characteristics are lower than the task state.

[0032] The fluctuation of topological features across time windows exceeds the task state range.

[0033] Furthermore, the method is implemented in an asynchronous SSVEP brain-computer interface system, comprising:

[0034] Real-time collection of EEG signals and pre-processing;

[0035] Update the status judgment result every 0.5 seconds;

[0036] When the idle state is determined for three consecutive times, the control instruction output is shielded.

[0037] Compared with the prior art, the advantages of the present invention are:

[0038] 1. By dynamically capturing the time-varying connectivity patterns of brain networks, this method can effectively distinguish complex idle states (such as resting with eyes closed and limb activity) from SSVEP control states, significantly reducing the risk of system false triggering and thus improving the safety of brain-controlled robot operations.

[0039] 2. Adopting multi-layer dynamic network analysis technology, we can overcome the limitations of traditional static features on the non-stationarity of EEG signals, maintain stable classification performance in scenarios with noise interference or large individual differences, and improve the system's adaptability in real environments.

[0040] 3. By adaptively selecting time window parameters, customized analysis solutions are provided for different users, significantly improving the generalization ability of the cross-subject model and making it more applicable to diverse user groups such as patients with motor disorders.

[0041] 4. Through accurate idle state discrimination, asynchronous control without external synchronization signals is achieved, making the human-computer interaction closer to natural operation habits and providing more reliable technical support for fields such as rehabilitation medicine and intelligent prosthetics.

[0042] 5. The classification model based on dynamic network topology features reduces the dependence on a large amount of training data or complex preprocessing processes, reduces the system computing burden, and improves the real-time response efficiency. Description of the Drawings

[0043] Figure 1 is the flowchart of the offline experiment of the embodiment of the present invention.

[0044] Figure 2 is the schematic diagram of the stimulus target coding of the embodiment of the present invention. Detailed Embodiments

[0045] To make the objectives, technical solutions and advantages of the present invention clearer and more understandable, the following further describes the present invention in detail with reference to the drawings and by way of examples.

[0046] The present invention provides an idle state discrimination method based on dynamic analysis of state-related networks, including the following steps:

[0047] Step 1: Build an offline test system for the SSVEP paradigm, conduct offline tests and record the EEG data of the occipital region.

[0048] A. Use the Brainconn NeuroHub EEG acquisition system. According to the 10-20 system, the EEG data is collected by nine electrodes located in the occipital region ("Oz", "O1", "O2", "Pz", "PO3", "PO4", "PO5", "PO6", "POz"), the ground electrode is located at AFz, the reference electrode is located at CPz, the sampling rate is 1000 Hz, and the electrode impedance is kept below 10 kΩ. During the test, the subject sits on a comfortable chair about 60 cm away from the monitor.

[0049] B. Design and configure the user interface to display the flashing SSVEP stimuli and the robot visual feedback. The interface is implemented using Psychtoolbox on a 23.6-inch LCD monitor (1920×1080 pixels, 60 frames per second). The sine wave stimulation method is adopted, and the joint frequency-phase modulation (JFPM) method is used to encode eight stimulus targets (as Figure 2 shown).

[0050] C. Conduct offline experiments. The experimental process is as follows Figure 1 As shown, a total of 4 groups of experiments are conducted, and the EEG data of each channel during the entire experimental process are collected and recorded.

[0051] Each group of experiments contains 24 trials, with a total of 4 groups (n = 4).

[0052] After each trial starts, the system randomly prompts the target stimulus source; after 4 seconds, all stimulus sources flash simultaneously. The subject needs to fixate on the target and remain motionless within 5 seconds, and then rest for 2 seconds.

[0053] There is a 1 - 3 - minute break between groups, allowing the subjects to move freely.

[0054] D. Preprocess the collected EEG signals. The collected data is first baseline - removed, then down - sampled to 250 Hz, band - pass filtered using a 5 - th order Butterworth filter with a frequency range of 7 - 90 Hz, and notch filtered at 50 Hz. An 11 - second - long data segment of the entire trial is extracted. For subsequent offline cross - validation, only the first 10 trials of all subjects are selected to obtain the offline dataset where N f represents the stimulus targets of different frequencies, N c represents the number of channels used in the occipital region, N s represents the number of sampling points, N t represents the number of offline trials, and N t = n × 24.

[0055] Step 2. Use the offline dataset to train and test the state - related dynamic network analysis method for SSVEP asynchronous classification performance, specifically as follows:

[0056] A. Data partitioning and label definition

[0057] 1. Template extraction: Extract the first time - window data from the task period of each trial in the template dataset as the TRCA template, and use task - related component analysis (TRCA) to calculate the TRCA model corresponding to the target.

[0058] 2. Dataset partitioning: Adopt ten - fold cross - validation to divide the data into a training set (90%) and a test set (10%), and select a template set (50%) from the training set.

[0059] 3. State labels: The task state is defined as the SSVEP stimulus response period (5 seconds), and the idle state is the preparation period (4 seconds).

[0060] B. Overlapping time - series reconstruction

[0061] 1. Window parameter setting:

[0062] Window length: 1s, 1.5s, 2s, 2.5s;

[0063] Overlap time: 0.1s, 0.5s, 1s;

[0064] 2. Personalized combination: After cross - validating each combination of window length and overlap time window, the optimal window length (2s) and overlap time (0.5s) are selected to generate a continuous overlapping time series.

[0065] C. Construction of state - related brain networks

[0066] 1. Network construction:

[0067] Calculate the network connection matrix of multi - channel signals in each time window under different TRCA templates and models.

[0068] 2. Generation of multi - layer dynamic network: Stack the brain networks of continuous time windows in chronological order to form a dynamic network sequence.

[0069] D. Extraction of network topological features

[0070] 1. Modularity analysis: Use the Louvain algorithm to partition the community of the network in each time window.

[0071] 2. Extraction of node topological features:

[0072] According to the result of module partition, calculate the participation coefficient, network relationship coefficient, diversity coefficient, and global efficiency of the nodes.

[0073] And take the average according to the network layer to obtain the global participation coefficient, global network management coefficient, global diversity coefficient, and global efficiency.

[0074] 3. Combine the features of each layer according to the time window to obtain continuous time - varying network topological features.

[0075] E. Pattern classification and performance verification

[0076] 1. Classifier training: Adopt the one - against - many method of the random forest (RF) classifier to construct a classifier set. The input features are continuous time - varying network topological features, and the output is the nine - class classification result of 8 control states and 1 idle state.

[0077] 2. Parameter optimization: Determine the optimal hyperparameters of RF (number of trees = 200, maximum depth = 10) through grid search.

[0078] 3. Performance metrics: The average classification accuracy of the offline test set reaches 92.87%, and the false trigger rate drops to 2.2%.

[0079] Those of ordinary skill in the art will realize that the embodiments described herein are provided to assist the reader in understanding the implementation methods of the present invention, and it should be understood that the protection scope of the present invention is not limited to such specific statements and embodiments. Those of ordinary skill in the art can make various other specific deformations and combinations that do not depart from the essence of the present invention based on these technical revelations disclosed in the present invention, and these deformations and combinations are still within the protection scope of the present invention.

Claims

1. A method for determining the idle state based on dynamic analysis of a state-related network, characterized in that It includes the following steps: Step 1: Collect the offline electroencephalogram (EEG) signal data of the subject under the SSVEP paradigm, and construct an occipital offline dataset including task states and idle states; Step 2: Process the offline dataset by using a state-related dynamic network analysis method, and extract the network topological features of the task state and the idle state, including: Perform temporal reconstruction on the EEG signals through overlapping time windows to construct a multi-layer dynamic network; Construct a state-related brain network by using a TRCA template to capture the connection patterns between brain regions in different states; Use the Louvain algorithm to partition the network into modules, and extract four time-varying network topological features; Step 3: Train a random forest classifier based on the network modular features to realize real-time discrimination between the SSVEP control state and the idle state.

2. The idle state discrimination method according to claim 1, wherein: The construction of the offline dataset in Step 1 specifically includes: arranging multiple electrodes in the occipital region to collect EEG signals, and the sampling rate is not less than 250 Hz; Perform baseline removal, band-pass filtering and 50 Hz notch filtering on the original data, and downsample it to 250 Hz; Intercept a continuous 11-second data segment including the SSVEP task execution period and the idle period, and organize it into a multi-dimensional dataset according to trials.

3. The idle state discrimination method according to claim 1, wherein: The setting of the overlapping time windows in Step 2 satisfies: the window length is 1 - 2.5 seconds, and the overlapping time is 0.1 - 1 second; Generate a personalized temporal segmentation scheme by combining different window lengths and overlapping times.

4. The idle state determination method according to claim 1, characterized in that: The construction of the state-related brain network by using a TRCA template specifically includes: Select the data of the first time window extracted from the task period of each trial in the template dataset as the TRCA template, and use task-related component analysis (TRCA) to calculate the TRCA model corresponding to the target. Calculate the network connection matrix of the multi-channel signals in each time window under different TRCA templates and models.

5. The idle state discrimination method according to claim 1, wherein: The extraction of the network topological features includes: global participation coefficient, global network relationship coefficient, global diversity coefficient, global efficiency. For different states, combine the four types of features within the time windows of their respective states to obtain the topological feature variation trends of each state across time windows.

6. The idle state discrimination method according to claim 1, wherein: The classifier in Step 3 is a random forest classifier, and the training process includes: Adopt ten-fold cross-validation to divide the training set and the test set; Optimize the hyperparameters of the classifier through grid search; Perform nine-class modeling on eight types of control instruction states and the idle state.

7. The idle state discrimination method according to claim 1, wherein: The construction of the multi-layer dynamic network includes: Stack the network connection matrices of consecutive time windows in chronological order; Analyze the time-varying law of the network properties to capture the critical features of state transitions.

8. The idle state discrimination method according to claim 1, wherein: The discrimination criteria for the idle state include: The network topological features are lower than those in the task state; The fluctuations of the topological features across time windows exceed the range of the task state.

9. The idle state discrimination method according to claim 1, characterized in that: The idle state discrimination method is implemented in an asynchronous SSVEP brain-computer interface system, including: Real-time collection of EEG signals and preprocessing; Updating the state discrimination result every 0.5 seconds; Blocking the output of control instructions when it is determined to be in the idle state three times in a row.

Images