A method for EEG channel selection based on Fisher score

Abstract

The present invention discloses a method for selecting EEG channels based on Fisher scores. The present invention uses a local optimization method based on Fisher scores to achieve channel selection. The present invention first makes a preliminary division of pre-selected channels based on the statistical measurement indicator of Fisher scores, and then uses a local optimization method to complete the selection of the optimal channel subset, analyzes the channel selection rules and accuracy of different bands, and simultaneously provides the position distribution of the selected channel subset in the sensor domain. The method proposed by the present invention significantly improves the classification accuracy of EEG signal motor imagery tasks compared to selecting all channels, and contributes to the portability and precision application of brain-computer interfaces.

Description

Technical Field

[0001] The present invention belongs to the fields of brain-computer interface technology, computer technology, biomedical engineering, artificial intelligence and brain science technology, and specifically relates to an EEG channel selection method based on Fisher scores. Background Art

[0002] A brain-computer interface (BCI) refers to a software and hardware communication system that enables information exchange between the human or animal brain and external devices without relying on conventional neuromuscular physiological systems. The purpose of a brain-computer interface is to control external devices or computers through brain activity, helping people with disabilities to return to normal life as much as possible. Motor imagery (MI) refers to the use of the brain's intention to imagine an action without performing the action, but there will be neuronal activity in the primary sensory motor area. A brain-computer interface system based on motor imagery refers to a system that collects and decodes the brain waves of a person when performing motor imagery tasks, thereby achieving control of external devices (robots, cars, etc.).

[0003] Compared with intracranial EEG (local field potentials and cortical electroencephalograms), scalp signals collected via scalp electrodes have lower spatial resolution. This is because the number of scalp electrodes is fixed and relatively small. Furthermore, due to the volume conduction effect, when neuronal activity in the cerebral cortex is transmitted from brain tissue to different locations on the scalp, the spatial expression of intracranial neuronal signals on the scalp is greatly weakened and interfered. The volume conduction effect also causes EEG signals collected by scalp electrodes to be relatively similar, further weakening the accuracy and effectiveness of scalp EEG decoding.

[0004] Multi-channel EEG signals can improve the resolution of EEG signals. Although a large number of EEG channels can provide richer EEG activity information, it also increases information redundancy and noise interference, and produces high-dimensional data. Since the amplitude of EEG signals is in the microvolt range, it is easily interfered with by noise (artifacts). Therefore, these artifacts need to be removed from the EEG signals to preserve truly valuable information. In order to reduce redundant information in EEG signals and obtain more real signals, reasonable EEG channel selection is very necessary. Channel selection can not only reduce the complexity of the BCI system, but also improve the accuracy of decoding, which is of great significance to the portable application of brain-computer interfaces.

[0005] Another question worth exploring is which metric to use for channel selection. Currently, some channel selection methods use numerical feature ranking to select channels. The problem with these methods is that ranking numerical features does not necessarily lead to a significant improvement in EEG signal classification, or that the best numerical features do not necessarily guarantee the best classification results. However, these methods do improve accuracy when a certain number of channels are selected. Summary of the Invention

[0006] To address the above issues, the present invention provides a locally optimized EEG channel selection method based on Fisher scores. This method uses Fisher scores as numerical features and further studies the impact of locally optimal channel subset combinations on classification accuracy. This method achieves higher classification accuracy than simple numerical feature channel selection. Based on numerical feature sorting, the present invention further studies the impact of locally optimal channel subset combinations on classification accuracy, improving the classification performance of motor imagery EEG signals.

[0007] The present invention is achieved through the following technical solutions.

[0008] A method for selecting EEG channels based on Fisher scores is proposed. The method applies a local optimization method to the EEG channels selected by Fisher scores. The specific steps are as follows:

[0009] (1) Filter the original EEG signal, calculate the Fisher scores of all channels in the corresponding frequency band and sort them;

[0010] (2) Select the top K channels with the highest Fisher scores as the optimal channels, which is the TOP-K-Fisher channel combination.

[0011] (3) The K value with the highest EEG information classification accuracy in the TOP-K-Fisher channel combination under the Top-K-Fisher method is selected as the K value of the subsequent local optimization method;

[0012] (4) For the TOP-K-Fisher channel subset selected in step (3), first select a channel and generate a test channel subset of the channel as the basis for subsequent feature extraction and classification, and mark it as the best channel with the highest classification accuracy in the TOP-K-Fisher channel n ;

[0013] (5) Delete the best channel from the TOP-K-Fisher channel subset n The channel, in the current best channel best n Add one of the remaining channels in the TOP-K-Fisher channel subset and generate the next test channel subset best n+1 , according to the EEG information classification accuracy of the channel, mark the best channel with the highest classification accuracy, which is recorded as the local optimal channel

[0014] (6) Repeat step (5) to gradually update the test channel subset best of the newly added channelsn+1 , and gradually mark the local optimal channel Until the number of test channels n c Equal to K;

[0015] (7) Finally, the channel subset with the highest classification accuracy is selected from all test channel subsets.

[0016] In the present invention, in step (1), the original EEG signal is filtered and divided into four frequency bands, namely: α, β, γ, and θ.

[0017] In the present invention, in step (1), the sampling rate of the original EEG signal is 250 Hz.

[0018] In the present invention, in step (2), for the BCI IV IIa data set, it is divided into 4 TOP-K channel combinations with different channel numbers (K = 5, 10, 15, 20); for the self-collected data set, it is divided into 12 TOP-K channel combinations with different channel numbers (K = 10, 20, 30, 40, 50, 60, 70, 80, 90, 100, 110, 120).

[0019] In the present invention, in step (3), the classifier used to classify the EEG signal is SVM, and the classification task is the motor imagery task of the left and right hands.

[0020] Compared with the prior art, the advantages of the present invention are as follows:

[0021] 1) Compared to the TOP-K-Fisher method, which simply selects the top K channels with the highest Fisher scores as the optimal channels, the TOP-K-LOCS method of the present invention selects channels within the TOP-K-Fisher channel set to obtain a series of locally optimal channel subsets, and then selects the subset with the highest accuracy to obtain the final channel set. This method can further select a channel subset effective for motor imagery signal decoding based on the TOP-K-Fisher channels. For the four bands of α, β, γ, and θ, as long as the appropriate K value is selected, the accuracy of the TOP-K-LOCS method is always higher than the accuracy of all channels.

[0022] 1. When the number of channels is small, the accuracy of the TOP-K-Fisher method cannot exceed the accuracy of all channels. However, when TOP-K-LOCS uses K = 5, the average accuracy of each band exceeds the accuracy of all channels. This also shows that the channel combination selected by TOP-K-LOCS is more effective than the channel subset combination of TOP-K-Fisher.

[0023] 2. The channel selection method of the present invention can reduce the number of channels to a certain extent. While ensuring improved accuracy, it also reduces the number of selected channels to a certain extent, which is conducive to the portability and precision application of brain-computer interface systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 This is a flow chart of the local optimization channel selection method of the present invention.

[0025] Figure 2 Calculate the Fisher scores for the dataset in four frequency bands.

[0026] Figure 3 Channel position distribution of nine subjects selected for the TOP-K-LOCS method in the theta band for the public dataset BCI IV IIa.

[0027] Figure 4 Channel position distribution of nine subjects selected for the TOP-K-LOCS method in the alpha band of the public dataset BCI IV IIa.

[0028] Figure 5 Channel position distribution of nine subjects selected for the TOP-K-LOCS method on the public dataset BCI IV IIa in the beta band.

[0029] Figure 6 Channel position distribution of nine subjects selected for the TOP-K-LOCS method on the public dataset BCI IV IIa in the gamma band. DETAILED DESCRIPTION

[0030] The technical solution of the present invention is clearly described below. Obviously, the embodiments described are only some of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making any creative efforts are within the scope of protection of the present invention.

[0031] The present invention proposes a motor imagery EEG decoding method based on channel selection, the flow chart is as follows Figure 1 As shown, the steps are as follows:

[0032] Step 1: Filter the collected EEG signals. Here, we use the BCI IV IIa public dataset with a sampling rate of 250Hz. We divide the EEG signals into four frequency bands, namely θ, α, β, and γ. Then, we calculate the Fisher scores of all channels in the four frequency bands and sort them. The calculated Fisher scores are as follows: Figure 2 shown.

[0033] Step 2: Based on the results of the Fisher score calculation, the EEG channels are divided into different channel combinations according to the Fisher score, namely TOP-K-Fisher channels. We divide the BCI IV IIa dataset into four TOP-K channel combinations with different channel numbers (K = 5, 10, 15, 20).

[0034] Step 3: For the channel subset of TOP-K-Fisher, only one channel (best n ,n=1), generate a test subset of channels as the basis for subsequent feature extraction and classification, and mark the best channel with the highest classification accuracy among the TOP-K-Fisher channels. We use SVM for classification.

[0035] Step 4: According to the best channel n ), in the next step, a subset of newly added channels is generated (best n+1 The specific method is to delete all the channels marked as best from the TOP-K-Fisher channel subset. n ) channels, and add the remaining channels in the best channel in turn to generate the next test channel subset (best n+1 ).

[0036] Step 5: Repeat step 4, gradually mark the local optimal channel, and update the test channel subset until the number of test channels n c Equal to K.

[0037] Step 6: Select the best subset and the corresponding accuracy in TOP-K-LOCS.

[0038] In the present invention, the Fisher score is calculated as follows:

[0039]

[0040] Where Fisher(n) is the Fisher score of the nth channel. B (n) is the between-class variance of the nth channel, S W (n) is the intra-class variance of the nth channel. S B The calculation formula for (n) is as follows:

[0041]

[0042] Where C is the number of task types of motor imagery EEG signals, n trails is the total number of motor imagery trials, [n rrails ] i represents the number of trials of the i-th type of motor imagery task, m i(n) is the feature average of the i-type motor imagery task on the n-th channel, m (n) It refers to the characteristic average of all types of motor imagery EEG on the nth channel, W i Indicates that the current experimental task is the i-th type of motor imagery task.

[0043]

[0044] Tables 1 and 2 show the accuracy results of the TOP-K-Fisher and TOP-K-LOCS methods, respectively, using a public dataset. The classifier used a support vector machine (SVM) and the classification task was a motor imagery task for both hands. Table 1 shows that the TOP-K-Fisher classification accuracy improves with increasing K values (K = 5, 10, 15, and 20). Table 2 shows that in the theta band, the highest accuracy reached 82.50%, 8.28% higher than the average accuracy of 74.22% across all channels. In the alpha band, the highest accuracy reached 81.60%, 6.21% higher than the 75.39% across all channels. In the beta band, the highest accuracy reached 85.13%, 6.03% higher than the 79.10% across all channels. In particular, in the gamma band, the average accuracy across all channels reached 72.81%, while the TOP-K-LOCS method with K = 20 achieved an accuracy of 80.66%, a 7.85% improvement.

[0045] Table 1 Accuracy of TOP-K-Fisher on the public dataset at different K values

[0046]

[0047] Table 2 Accuracy of the public dataset TOP-K-LOCS under different K values

[0048]

[0049] Different from the TOP-K-Fisher channel selection method, TOP-K-LOCS achieves the highest accuracy when K=20 in the four bands and the number of selected channels is less than 20, that is, it selects fewer channels than the TOP-K-Fisher method.

[0050] Compared to the TOP-K-Fisher method, the TOP-K-LOCS method can further select a subset of channels effective for decoding motor imagery signals based on the TOP-K-Fisher channels. Results from the four bands show that, as long as the appropriate K value is selected, the accuracy of the TOP-K-LOCS method consistently exceeds that of all channels. This also demonstrates that the channel combination selected by TOP-K-LOCS is more effective than the channel subset combination used by TOP-K-Fisher.

[0051] Figure 3 The results of TOP-K-LOCS channel selection in the theta band using a publicly available dataset were presented. We found that across nine subjects, the average classification accuracy was 82.50%. Furthermore, for Subject 5, the TOP-K-LOCS method discarded several channels located in the parietal lobe, including CZ, C4, C1, and C3. Similarly, for Subjects 6 and 7, TOP-K-LOCS also improved accuracy by discarding several channels located in the parietal lobe, including CZ, C2, C4, and CP2. Therefore, the parietal lobe region may contain less effective motor imagery information in the theta band, leading to the omission of some channels from this region during channel selection based on the TOP-K-LOCS method.

[0052] Table 3 shows the number of channels selected by the TOP-K-LOCS method for nine subjects in the public dataset. The average number of channels in the four bands is 10.56, 8.11, 10.44, and 11.78, respectively. Compared with the original 22 electrode channels, the number of selected channels has been significantly reduced. In specific applications, electrodes can be set based on the average channel count, which not only reduces the number of electrodes in the electrode cap but also improves classification results.

[0053] Table 3 Number of channels selected by the TOP-K-LOCS method for 9 subjects in the public dataset

[0054]

[0055] In summary, this paper utilizes the aforementioned design methods and data to develop a channel-selective motor imagery EEG decoding method. This method significantly reduces the number of channels required for motor imagery BCI classification. Furthermore, the authors analyze the reasons for discarded channels in relation to motor imagery, providing a possible physiological explanation. In summary, this paper holds promise for reducing the number of electrode channels required for collecting EEG signals during motor imagery classification, providing a new direction for the portability of BCI systems.

[0056] The above description of the specific implementation of the present invention is intended to facilitate understanding of the present invention by other persons skilled in the art. It should be noted that the present invention is not limited to the above implementations. As long as various variations are within the scope of the appended claims, all inventions and creations utilizing the present invention are protected.

Claims

1. A method for selecting EEG channels based on Fisher scores, characterized in that: The local optimization method is applied to the EEG channels after Fisher score selection to select channels. The specific steps are as follows: (1) Filter the original EEG signal, calculate the Fisher scores of all channels in the corresponding frequency band and sort them; (2) Based on the results of Fisher score calculation, the EEG channels are divided into different channel combinations according to the Fisher score, and the channels with the highest top K Fisher scores are selected, namely the TOP-K-Fisher channel combination; (3) The K value with the highest EEG information classification accuracy in the TOP-K-Fisher channel combination under the TOP-K-Fisher method is selected as the K value of the subsequent local optimization method; (4) For the TOP-K-Fisher channel subset selected in step (3), first select a channel and generate a test channel subset of the channel, that is, select the EEG data of one channel as the basis for subsequent feature extraction and classification, and mark it as the best channel with the highest classification accuracy in the TOP-K-Fisher channel. n ; (5) Delete the best channel from the TOP-K-Fisher channel subset n The channel, in the current best channel best n Add one of the remaining channels in the TOP-K-Fisher channel subset and generate the next test channel subset best n+1 , according to the EEG information classification accuracy of the channel, mark the best channel with the highest classification accuracy, which is recorded as the local optimal channel (6) Repeat step (5) to gradually update the test channel subset best of the newly added channels n+1 , and gradually mark the local optimal channel Until the number of test channels n c Equal to K; (7) Finally, the channel combination with the highest accuracy is selected from all test channel subsets.

2. The EEG channel selection method according to claim 1, characterized in that: In step (1), the original EEG signal is filtered and divided into four frequency bands: θ (4-8 Hz), α (8-12 Hz), β (12-30 Hz), and γ (30-45 Hz).

3. The EEG channel selection method according to claim 1, characterized in that: In step (1), the sampling rate of the original EEG signal is 250 Hz.

4. The EEG channel selection method according to claim 1, characterized in that: In step (2), the BCI IV IIa dataset is divided into four channel combinations: TOP-K, K = 5, 10, 15, and 20 channels.

5. The EEG channel selection method according to claim 1, characterized in that: In step (2), the self-collected data set is divided into 12 TOP-K channel combinations with different channel numbers, K = 10, 20, 30, 40, 50, 60, 70, 80, 90, 100, 110, 120.

6. The EEG channel selection method according to claim 1, characterized in that: The classifier used to classify EEG signals is SVM, and the classification task is the motor imagery task of the left and right hands.

Images