Classification method for interventional electroencephalogram signal motor imagery tasks

Abstract

The invention relates to the field of biological signal processing, and particularly discloses a classification method for intrusive electroencephalogram signal motor imagery tasks, which comprises the following steps of: acquiring original electroencephalogram signals according to task requirements through an intrusive brain-computer interface system; preprocessing the electroencephalogram signal data, and removing noise in combination with a signal-to-noise ratio; performing independent component analysis to extract an electroencephalogram signal source; signal features are extracted manually, and principal component analysis dimension reduction is carried out; and putting the data into the model for training, and classifying the signals. According to the method, noise and redundant information of the electroencephalogram signals are effectively removed through combination of ICA and PCA, and the influence of artifacts on a classification result is reduced, so that the representative characteristics of the signals are improved, the classification accuracy is improved, and a good classification effect can still be ensured especially in a high-noise environment.

Description

Technical Field

[0001] The present invention relates to the field of biological signal processing, and particularly to electroencephalogram (EEG) signal processing. Background Art

[0002] Brain-computer interface (BCI) technology is a transformative human-computer interaction technology. Its mechanism is to bypass the peripheral nerves and muscles and directly establish a new communication and control channel between the brain and external devices. It captures brain signals and converts them into electrical signals to achieve information transmission and control. BCI technology can be divided into three types according to the invasive method: non-invasive, semi-invasive, and invasive.

[0003] Non-invasive BCI collects brain signals through electrodes or other sensors on the scalp without penetrating the skull. This method causes less harm to users, but the signals collected will attenuate during the transmission from the brain to the scalp, with relatively low quality and greater decoding difficulty. Invasive BCI inserts electrodes directly into brain tissue and can collect higher-quality signals, but this method causes greater harm to users and has risks such as infection and rejection.

[0004] Semi-invasive BCI is achieved through a minimally invasive approach by puncturing a small blood vessel and performing a minimally invasive surgery similar to a cardiac stent to establish a brain-computer connection. Specifically, it uses the neurointerventional technology for treating stroke. After implanting EEG sensors into brain regions such as the motor cortex and visual cortex of the brain through a vein, the nerve stent expands and presses the electrodes against the blood vessel wall close to the brain to obtain signals from the corresponding brain regions. This method avoids the high risk of craniotomy required by traditional invasive BCI and improves the quality of signal acquisition. Currently, it has broad application prospects in many fields such as medicine and scientific research.

[0005] EEG signal classification is to identify the task state of unknown data by analyzing the electrical signals generated by the activities of brain neurons in different task states. These tasks may include cognitive tasks (such as reading and calculating), emotional tasks (such as watching emotional stimuli), motor tasks (such as imagining movement), etc. The current EEG signal classification methods mainly include two types: the first is to extract the basic features of EEG signals and determine the functional relationship between the features and categories through mathematical analysis, with strong interpretability. The second is to use deep learning, input the EEG signals into the model for inference, without the need to manually extract the corresponding features, and it performs well on large datasets.

[0006] Due to risks such as infection in clinical surgeries, currently, non-invasive acquisition methods are mainly used for brain-computer interfaces. All publicly available electroencephalogram (EEG) signal datasets are collected by non-invasive brain-computer interface methods, and most of the EEG signal classifications use deep learning methods. Due to inconsistent datasets, this model is not applicable to invasive EEG classification. Therefore, we propose a method of denoising EEG signals by manually setting parameters and removing artifacts through independent component analysis, which can be applied to invasive EEG signals different from non-invasive ones. By manually extracting features, it can achieve a classification method for a data processor with a relatively fast inference speed without the need for a large-scale dataset.

[0007] The technical problem to be solved by this application is: how to achieve efficient classification of invasive EEG signals in multiple task states. Summary of the Invention

[0008] In order to overcome the deficiencies of the prior art, the purpose of the present invention is to provide a classification method for motor imagery tasks of invasive EEG signals.

[0009] The technical solution adopted by the present invention is: a classification method for motor imagery tasks of invasive EEG signals, including the following steps:

[0010] S1. The invasive brain-computer interface system collects the original EEG signals according to the target task requirements and corrects the baseline of the original EEG signals;

[0011] S2. Perform data preprocessing and cropping on the original EEG signals to form several EEG signal segments from different regions of the brain but with the same length and scale;

[0012] S3. Decompose the EEG signal segments using independent component analysis to obtain several mutually independent source signals, and manually extract the independent components most relevant to the target task from the source signals;

[0013] S4. Based on the selected independent components, manually extract signal features, perform principal component analysis (PCA) for dimensionality reduction, and retain the principal components of the EEG signal data;

[0014] S5. Put the principal components of the EEG signal data into a random forest model for training;

[0015] S6. Put the trained random forest model into application and complete the update of the model through iterative training.

[0016] In some embodiments, regarding S2, the data preprocessing includes:

[0017] S21. The corrected original EEG signals are cropped into EEG signal segments of the same length, and the length of the EEG signal segments is determined according to the task duration. Each EEG signal segment corresponds to one motor imagery movement;

[0018] S22. Remove the low-frequency drift and high-frequency noise in the stopband of the EEG signal segment;

[0019] S23. Combine the signal-to-noise ratio of the vascular intervention stent electrode, estimate the frequency band range of the noise through the overall power spectrum distribution of the EEG signal segment, and remove the noise in this frequency band;

[0020] S24. Standardize or normalize the amplitude of each EEG signal segment so that the EEG signal segments from different brain regions have the same scale.

[0021] In some embodiments, regarding the manual extraction of the independent component most relevant to the target task in the source signal in S3, screening is performed from four directions, including the physiological direction, statistical significance, task relevance, and empirical rules. The physiological direction preferentially selects the energy in the μ / β frequency band, ERD slope, and C3-C4 channel asymmetry directly related to motor imagery; the statistical significance is to screen features with strong separability between categories (p < 0.01) through t-test or ANOVA; the task relevance is to calculate the Pearson correlation coefficient between the feature and the task time window (|r| > 0.3); the empirical rule is to exclude features with an amplitude < 5 μV or variance < 0.1 (the intervention signal threshold can be relaxed to 2 μV).

[0022] In some embodiments, S4 includes:

[0023] S41. Based on the selected independent components, manually extract the independent components with neurophysiological significance of signal features;

[0024] S42. Based on the independent components extracted in S41, extract its frequency domain features, time domain features, time-frequency domain features, and event-related features directly related to the execution of the task;

[0025] S43. Analyze the correlation between the above features, manually screen the extracted features, and remove redundant and noise features.

[0026] In some embodiments, the criteria for manually screening features in S43 include removing highly correlated features, processing overlapping frequency bands, and filtering low-contribution features.

[0027] In some embodiments, the principal components of the EEG signal data are divided into a training set and a test set. The training set data is used to train the random forest model, and the test set data is used to evaluate the trained model, and calculate indicators such as classification accuracy, recall rate, and F1 score.

[0028] The beneficial effects of the present invention are as follows:

[0029] The classification method for the movement imagination task of invasive electroencephalogram (EEG) signals effectively removes noise and redundant information from EEG signals by combining independent component analysis (ICA) and principal component analysis (PCA), reduces the influence of artifacts on the classification results, thereby improving the representative features of the signals and enhancing the classification accuracy. Especially in a high-noise environment, it can still ensure good classification results and is applicable to invasive EEG signals different from non-invasive ones. At the same time, it manually extracts features without the need for a large-scale dataset, and uses a random forest model to achieve classification with small data samples and relatively fast inference speed. PCA dimensionality reduction reduces the dimensionality of the feature space, thereby reducing the computational burden during the training process and also reducing the feature data to be statistically analyzed during the actual inference process, which helps improve the inference speed. This method can achieve efficient classification and a lightweight classification architecture for multi-task invasive EEG signals, can be flexibly adjusted according to different task requirements, and is applicable to a variety of invasive brain-computer interface application scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 is a flowchart of the task acquisition method;

[0031] Figure 2 is a flowchart of the model training process; DETAILED DESCRIPTION OF THE EMBODIMENTS

[0032] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0033] Please refer to Figure 1 , the present invention provides a technical solution: a classification method for the movement imagination task of invasive EEG signals, including the following steps:

[0034] S1. The invasive brain-computer interface system collects the original EEG signals according to the target task requirements and corrects the baseline of the original EEG signals;

[0035] S2. Perform data preprocessing and cropping on the original EEG signals to form several EEG signal segments from different brain regions but with the same length and scale;

[0036] S3. Decompose the EEG signal segments using independent component analysis to obtain several independent source signals, and manually extract the independent components most relevant to the target task from the source signals;

[0037] S4. Based on the selected independent components, manually extract signal features, perform principal component analysis (PCA) dimensionality reduction, and retain the principal components of the EEG signal data;

[0038] S5. The principal components of the electroencephalogram (EEG) signal data are put into a random forest model for training;

[0039] S6. The trained random forest model is put into application and the model is updated through iterative training.

[0040] The detectable EEG signals induced by the motor imagery movement task mainly exist in the sensorimotor cortex, and its EEG signals have the following characteristics: 1. ERD / ERS phenomenon (Event-Related Desynchronization / Synchronization): when imagining movement, the energy of the μ rhythm (8-12 Hz) and β rhythm (13-30 Hz) decays (ERD), and the energy rebounds after the task ends (ERS); 2. Spatiotemporal specificity: contralateral activation of the left and right hemispheres, such as significant signal differences in the C3 / C4 channels of the left brain when imagining right hand movement; 3. Time-frequency non-stationarity: the signal characteristics change rapidly within 0.5-2 seconds, and high time-resolution equipment is required to capture them.

[0041] Regarding the signal collection of these two brain-computer interface invasion methods, non-invasive and invasive, the signal amplitude collected non-invasively is only 5-100 μV, which is easily interfered by electrooculogram / electromyogram (the signal-to-noise ratio <5 dB), and the signal is weak; there is a ±2 Hz offset in the ERD / ERS frequency band among different individuals (personalized calibration is required), and the individual differences are large; it is necessary to synchronously record the real limb movements to align the EEG signal labels, and the annotation cost is high. For invasive collection, since the vascular stent electrode directly contacts the cortex, the signal amplitude is increased to 200-500 μV, and the signal-to-noise ratio >15 dB. The signal frequency band collected invasively is more stable. By synchronously collecting the local field potential (LFP) of the motor cortex through the vascular stent electrode, external sensor-free annotation is achieved, and the annotation cost is low. The signal collected by the invasive method has better quality, a more stable collection process, and a lower later annotation cost.

[0042] Regarding step S1, according to the target task requirements, the original EEG signal is collected through the vascular stent motor of the invasive brain-computer interface system, and during the collection process, the original EEG signal is collected by setting the sampling frequency and amplifier gain, so as to obtain the original EEG signal that has been amplified and baseline corrected.

[0043] Regarding step S2, the data preprocessing includes:

[0044] S21. The corrected original EEG signal is cropped into EEG signal segments of the same length, and the length of the EEG signal segment is determined according to the task duration, and each EEG signal segment corresponds to one motor imagery movement;

[0045] S22. The EEG signal segments are processed by band-pass filtering to remove power frequency interference and by Butterworth filtering to remove low-frequency drift and high-frequency noise in the stopband, ensuring that the frequency response curve in the passband is maximally flat.

[0046] S23. Combining the signal-to-noise ratio of the vascular intervention stent electrodes, estimate the frequency band range of the noise through the overall power spectrum distribution of the EEG signal segments and remove the noise in this frequency band.

[0047] S24. Standardize or normalize the amplitude of each EEG signal segment so that the EEG signal segments from different brain regions have the same scale, eliminating the errors caused by differences in different channel electrodes, ensuring that the EEG signal segments from each channel have zero mean and unit variance, which helps the stable convergence of the subsequent independent component analysis algorithm.

[0048] Regarding step S3, the independent component analysis algorithm is used to decompose the EEG signal segments. This step aims to decompose the mixed EEG signal segments into several mutually independent source signals, and each source signal represents the neural activities in different regions of the brain. According to the spatio-temporal characteristics, spectral characteristics, and correlation with the task of the independent components, manually select the independent components that are most relevant to the target task in the source signals to ensure that the selected independent components can accurately reflect the activity state of the brain when performing a specific task. At the same time, optimize the selected components using the information maximization criterion to further improve the quality of the source signals.

[0049] When manually selecting the independent components that are most relevant to the target task, screen from four directions of criteria, including the physiological direction, statistical significance, task relevance, and empirical rules. Regarding the physiological direction, preferentially select the μ / β band energy, ERD slope, and C3-C4 channel asymmetry that are directly related to motor imagery; regarding statistical significance, screen the features with strong separability between categories (p < 0.01) through t-test or ANOVA; regarding task relevance: calculate the Pearson correlation coefficient (|r| > 0.3) between the features and the task time window; empirical rule: exclude the features with amplitude < 5 μV or variance < 0.1 (the intervention signal threshold can be relaxed to 2 μV).

[0050] Regarding step S4, it includes the following steps: S41. Based on the selected independent components, manually extract the independent components with neurophysiological significance signal features. The signal features with neurophysiological significance include amplitude, frequency, phase, and coherence, and these signal features can reflect the neuroactivity characteristics of the brain in different states; S42. Based on the independent components extracted in S41, extract their frequency domain features, time domain features, time-frequency domain features, etc. At the same time, combined with the prior knowledge related to the task, extract the event-related features directly related to task execution from the independent components, such as event-related desynchronization (ERD) and event-related synchronization (ERS) and other features; S43. Analyze the correlation between the above features, and manually screen the extracted features to remove redundant and noise features; S44. Perform PCA processing on the screened feature subset to reduce the dimension of the feature space; S45. PCA projects the high-dimensional data onto a low-dimensional space through linear transformation while retaining the main variation information of the data. This step helps to reduce the amount of calculation, improve the running efficiency of the classifier, and may improve the classification performance.

[0051] The criteria for manual screening in step S43 include the following:

[0052] Eliminating highly correlated features: If the correlation coefficient between two features > 0.85, retain the feature with the larger variance (assuming it has more information);

[0053] Processing overlapping frequency bands: For multiple features within the same frequency band (such as the β-wave energy of channel C3, the wavelet entropy of C3 β-wave), retain the feature with a higher correlation with the classification label;

[0054] Filtering low-contribution features: Manually exclude features with an impact on the classification accuracy < 1% (through feature permutation importance testing).

[0055] Regarding step S5, the data processor learns a random forest classifier. Divide the principal components of the EEG signal data that have completed steps S1 - S4 into a training set and a test set. Use the training set data to train the random forest model, and minimize the classification error by adjusting the model parameters and optimizing the algorithm. At the same time, adopt strategies such as cross-validation to evaluate the performance of the model to ensure the stability and reliability of the model.

[0056] Use the test set data to evaluate the trained model, and calculate metrics such as classification accuracy, recall rate, F1 score, etc. Optimize the model according to the evaluation results, such as adjusting feature selection, improving algorithm parameters and other strategies to improve the classification performance of the model. Since the principal component dimension of the EEG signal data obtained after PCA is moderate and the data volume is small, it is suitable for the tree model splitting rule learning. The random forest model has strong interpretability, is suitable for small data volumes, and has a fast real-time inference speed.

[0057] The specific implementation operations are as follows:

[0058] S1. The invasive brain-computer interface system collects the original EEG signals according to the task requirements and corrects the baseline.

[0059] As Figure 1 shown in the task acquisition method, the subject performs five motor imagery tasks of clenching the left fist, clenching the right fist, kicking the left leg, kicking the right leg, and running according to the prompts, and collects the original EEG signals through the invasive brain-computer interface. Each task lasts for 5 seconds, and the interval between tasks is 3 seconds. The experiment is divided into 5 sessions, and 30 task cycles of data are collected in each session.

[0060] The original EEG signals are obtained through the vascular stent electrodes of the invasive brain-computer interface. The sampling frequency is set to 2000Hz respectively, the amplifier gain is set to 500 times, and the baseline of the original EEG signals is corrected to reduce the baseline drift of the sampling device and the environment.

[0061] S2. Preprocess and clip the original EEG signals.

[0062] According to the task duration, the original EEG signals are clipped into equal-length segments of 5s, and each segment corresponds to one motor imagery task. A band-stop filter (49 - 51Hz) is used to remove the power frequency interference; the Butterworth band-pass filter is further used to eliminate the stop-band noise and ensure that the frequency response curve of the pass-band (0.5 - 200Hz) is flat.

[0063] Combined with the signal-to-noise ratio of 30 of the vascular intervention stent electrodes, the overall power spectral density of each EEG signal segment is calculated, the average power spectral density of all EEG signal segments is calculated, the frequency bands of the pure signal and the noise signal are calculated according to the ratio of 30:1, and the signal is filtered according to the noise estimation value to ensure the purity of the signal.

[0064] S3. Perform independent component analysis on the EEG signal segments and manually extract the EEG signal sources.

[0065] Normalize each EEG signal segment so that the signals from all channels have zero mean and unit variance, and eliminate the uneven signal amplitude caused by electrode differences.

[0066] Using the fast independent component analysis algorithm, the mixed EEG signals are decomposed into several independent components, and each component represents the neural activity source of a specific brain region. According to the spatio-temporal characteristics, spectral characteristics and correlation with the task, manually select the independent components most relevant to the motor imagery task. Optimize the independent components using the information maximization criterion to ensure that the extracted signals can accurately reflect the neural activity patterns related to motor imagery.

[0067] S4. Manually extract the signal features, perform principal component analysis (PCA) for dimensionality reduction, and retain the principal components of the EEG signal data.

[0068] Extract the features below the signal:

[0069] Frequency-domain features: Extract the power spectral density of alpha waves (8–13 Hz) and beta waves (13–30 Hz), which are closely related to the motor imagery task.

[0070] Time-domain features: Calculate features such as the average amplitude, variance, and peak value of the signal.

[0071] Time-frequency domain features: Perform wavelet transform on the signal to extract the features of the electroencephalogram signal.

[0072] Event-related features: Extract features such as event-related desynchronization (ERD) and event-related synchronization (ERS), which can reflect the changes in brain region activity.

[0073] Analyze the correlation between features, remove redundant features, and use principal component analysis (PCA) to reduce the dimension of high-dimensional features, reducing the dimension of the feature space while retaining 95% of the information of the electroencephalogram signal data.

[0074] S5. Put the principal components of the electroencephalogram signal data into the random forest model for training;

[0075] As Figure 2 shown in the model training process, randomly divide the principal components of the electroencephalogram signal data into a training set (80%) and a test set (20%).

[0076] Select the random forest classifier as the learning model. The random forest can effectively process high-dimensional and complex data by integrating multiple decision trees. Use the training set to train the model, adopt five-fold cross-validation to find the best hyperparameters (such as the number of trees, maximum depth, etc.), and save the model locally.

[0077] Use the test set to evaluate the model performance, and calculate metrics such as classification accuracy (Accuracy), recall rate (Recall), and F1 score (F1 Score) through the following formulas.

[0078]

[0079] Among them, TP (True Positive) is the number of samples correctly predicted as the positive class; TN (True Negative) is the number of samples correctly predicted as the negative class; FP (False Positive) is the number of samples incorrectly predicted as the positive class (actually the negative class); FN (False Negative) is the number of samples incorrectly predicted as the negative class (actually the positive class).

[0080] S6. Put the trained random forest model into application and complete the update of the prediction model through iterative training.

[0081] Finally, it should be noted that the above are only preferred examples of the present invention and are not used to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or perform equivalent replacements for some of the technical features. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A classification method for motor imagery tasks of invasive electroencephalogram signals, characterized in that It includes the following steps: S1. The invasive brain-computer interface system collects the original EEG signals according to the target task requirements and corrects the baseline of the original EEG signals; S2. Perform data preprocessing and cropping on the original EEG signals to form several EEG signal segments from different brain regions but with the same length and scale; S3. Use independent component analysis to decompose the EEG signal segments to obtain several independent source signals, and manually extract the independent components most relevant to the target task in the source signals; S4. Based on the selected independent components, manually extract signal features, perform principal component analysis (PCA) dimensionality reduction, and retain the principal components of the EEG signal data; S5. Put the principal components of the EEG signal data into a random forest model for training; S6. Put the trained random forest model into application and complete the update of the model through iterative training.

2. The classification method for the imagined movement task of invasive electroencephalogram signals according to claim 1, wherein Regarding S2, the data preprocessing includes: S21. The corrected original EEG signals are cropped into EEG signal segments of the same length, and the length of the EEG signal segments is determined according to the task duration. Each EEG signal segment corresponds to a single motor imagery movement; S22. Remove the low-frequency drift and high-frequency noise in the stopband from the EEG signal segments; S23. Combine the signal-to-noise ratio of the vascular intervention stent electrodes, estimate the frequency band range of the noise through the overall power spectrum distribution of the EEG signal segments, and remove the noise in this frequency band; S24. Standardize or normalize the amplitude of each EEG signal segment so that the EEG signal segments from different brain regions have the same scale.

3. A classification method for an invasive electroencephalogram signal motor imagery task according to claim 1, characterized in that, Regarding manually extracting the independent components most relevant to the target task in S3, screening is performed from four directions of criteria, including the physiological direction, statistical significance, task relevance, and empirical rules. The physiological direction gives priority to selecting the μ / β band energy, ERD slope, and C3-C4 channel asymmetry directly related to motor imagery; the statistical significance is to screen features with strong separability between categories (p<0.01) through t-test or ANOVA; the task relevance is to calculate the Pearson correlation coefficient (|r|>0.3) between the feature and the task time window; the empirical rule is to exclude features with an amplitude <5 μV or a variance <0.1 (the invasive signal threshold can be relaxed to 2 μV).

4. A classification method for an intervention-based electroencephalogram signal motor imagery task according to claim 1, characterized in that, S4 includes: S41. Based on the selected independent components, manually extract the independent components of the signal features with neurophysiological significance; S42. Based on the independent components extracted in S41, extract its frequency domain features, time domain features, time-frequency domain features, and event-related features directly related to the execution of the task; S43. Analyze the correlation between the above features, manually screen the extracted features, and remove redundant and noise features.

5. A classification method for an invasive electroencephalogram signal motor imagery task according to claim 4, characterized in that, The criteria for manually screening features in S43 include high-correlation feature elimination, frequency band overlap processing, and low-contribution feature filtering.

6. A classification method for an invasive electroencephalogram signal motor imagery task according to claim 1, characterized in that, The principal components of the EEG signal data are divided into a training set and a test set. The training set data is used to train the random forest model, and the test set data is used to evaluate the trained model, and calculate indicators such as classification accuracy, recall rate, and F1 score.

Images