Practical few-channel dry electrode-oriented brain-computer interface normal form and decoding method

Abstract

The invention discloses a normal form for a practical small-channel dry electrode brain-computer interface (BCI) and a decoding method of the normal form for the practical small-channel dry electrode BCI. According to the method, multiple excitation forms are introduced, so that the paradigm can induce coupling of electroencephalogram signals in multiple brain regions; according to the method, a Motor Preparation (MP) normal form is designed, instructions in a video are sent out in four modes of voice and character, pure voice, pure character and arrow, a subject focuses on the voice and character instructions broadcasted in the video, and various actions such as leftward and rightward natural turning and forward left foot and right foot moving are executed according to the instructions. The MP normal form can be used as a basis for designing a closed-loop BCI system, and external equipment such as an exoskeleton and a wheelchair can be controlled by decoding a motion preparation signal of a subject. The MP normal form and the corresponding decoding method can effectively assist the rehabilitation process of a dyskinesia patient or enhance the athletic ability of a healthy subject. The method has the characteristics of noninvasiveness, convenience, low cost and low delay, and is expected to be popularized to practical application and clinical rehabilitation scenes.

Description

Technical Field

[0001] This invention relates to the field of brain-computer interface technology, and in particular to a practical low-channel dry electrode brain-computer interface paradigm and a method for decoding electroencephalogram (EEG) signals. Background Technology

[0002] Brain-computer interfaces (BCIs) are technologies that read, process, and decode neural signals generated by brain activity, converting them into specific commands to convey the user's intentions. BCIs can be combined with external devices such as exoskeletons to restore and enhance the motor abilities of patients with severe movement disorders, or to decode the movement intentions of healthy users, building proactive human-computer interaction systems. Electroencephalography (EEG) is one of the BCI methods, widely used due to its low cost, portability, and non-invasiveness.

[0003] Motor imagery (MI) and motor execution (ME) are important paradigms of BCI. MI collects the electroencephalogram (EEG) signals generated when the subject imagines movement of a specific part of their body, and decodes and maps these signals into corresponding control commands. ME, on the other hand, collects the EEG signals generated when the subject actually performs the corresponding actions, enabling control of external devices.

[0004] However, for most untrained subjects, motor imagery (MI), as a relatively abstract paradigm, is often difficult to intuitively understand and accurately execute, thus increasing the cognitive burden of task performance. Furthermore, due to its strong implicitness and subjectivity, it is difficult to objectively assess whether the subject correctly performed the motor imagery. In contrast, while motor imagery (ME) signals are stronger, they exhibit significant motion artifacts, placing higher demands on EEG signal preprocessing; simultaneously, ME signals are typically triggered only after the action occurs, resulting in a time delay that limits its application in high real-time scenarios. Summary of the Invention

[0005] To address the problems of high abstraction, high usage threshold, large response latency, and susceptibility to artifacts in existing MI and ME paradigms, this invention proposes a practical few-channel dry electrode brain-computer interface (BCI) paradigm and decoding method, specifically the Motor Preparation (MP) paradigm. This paradigm is more intuitive and offers low latency and high accuracy. Furthermore, the decoding method employs a corresponding Motor Preparation paradigm dataset and a weighted voting decision method, effectively avoiding the problems of the MI and ME paradigms and improving the decoding accuracy of EEG signals.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] A practical low-channel dry electrode brain-computer interface paradigm is proposed, implemented based on a low-channel dry electrode brain-computer interface device. It uses pure voice, pure text, or voice-text stimulation, and the action commands in the video are issued at a fixed rhythm. The task description is: to focus attention on the commands issued in the video and perform the corresponding actions according to the commands, and to collect the EEG signals of the motor preparation stage between the issuance of the command and the execution of the movement.

[0008] Furthermore, the video can be pre-recorded, and the recording method is as follows:

[0009] A random action instruction is displayed on the screen as text, or the audio content of the text is played simultaneously, or the audio content of the text is played while it is displayed as text; the subject executes the action instruction after it is displayed or played; the subject's EEG signals are collected during the exercise preparation phase.

[0010] Furthermore, the execution of the corresponding actions includes various movements such as natural body rotation to the left and right, and stepping forward with the left or right foot.

[0011] Furthermore, the brain-computer interface device used for EEG signal data acquisition is a dry electrode device.

[0012] Furthermore, after the subject correctly completes the corresponding turning movement according to the instructions, it can be ensured that the EEG signals collected during the exercise preparation period are correct and effective.

[0013] Furthermore, the execution of the task includes:

[0014] Step 1: Subjects focus their attention on the instructions given in the video;

[0015] Step 2: The subject performs the corresponding actions according to the instructions given in the video;

[0016] Step 3: If the subject correctly completes the corresponding action according to the instructions, it indicates that the task has been fully understood and executed correctly; after the action is completed, the subject returns to the initial position and prepares to receive the next instruction.

[0017] This invention further provides a decoding method for EEG signals obtained by the practical few-channel dry electrode brain-computer interface paradigm. Based on the two-point observation network DoNet, the method first samples two-point pairs of narrowband signals of low signal-to-noise ratio ultra-short time signals collected during the motor preparation phase, calculates feature vectors based on the sampled values, and finally performs weighted voting decision on the predicted labels output by the neural network to obtain the execution action category.

[0018] Preferably, the present invention uses the dataset corresponding to the paradigm to train the two-point observation network DoNet. The dataset includes the excitation methods of speech text, pure speech and pure text arrows and the corresponding actions. The original collected EEG signals are filtered, cropped and preprocessed, and then spliced ​​to obtain the dataset.

[0019] Preferably, the filtering is performed using a Butterworth filter for bandpass filtering from 0.1 to 48 Hz; the cropping involves cropping the original 1000 data points to 200 points, corresponding to the data from the 2.5th to 4.1st second of the data acquisition process, to ensure that the final data includes information from the movement preparation phase and is not interfered with by electromyographic signals; the splicing involves integrating the EEG data obtained from each subject in 80 experiments into a complete data file to facilitate subsequent training and testing of the neural network; wherein, the original data acquisition is as follows: for each subject, each type of stimulus includes 80 experiments, and each type of stimulus corresponds to two movement commands, left turn and right turn. Each experiment yields two sets of data corresponding to the two movement commands, each set of data lasts for 8 seconds, the sampling frequency is 125 Hz, the data length is 1000 points, and the total experiment duration is 80 minutes.

[0020] Preferably, the weighted voting decision assigns weights to the classification results of the neural network according to their proximity to the motion execution period, with signals closer to the motion execution period being given greater voting weights.

[0021] Furthermore, in the weighted voting decision-making method, the voting weight increases monotonically over time, gradually increasing from 0.5 to 1.

[0022] Furthermore, the weight increment process includes three methods: linear increment, convex function increment, and concave function increment, the specific expressions of which are shown below:

[0023] Linear functions:

[0024] Convex function:

[0025] Concave function:

[0026] Among them, w i The weight of the i-th prediction in the voting decision-making process is given by N, where N is the total number of predictions, and p and k are parameters used to control the steepness of the curve.

[0027] The traditional MI paradigm is relatively abstract, while the ME paradigm suffers from problems such as large motion artifacts and slow response. This invention designs a Motor Preparation (MP) paradigm, which directly uses natural language action instructions. Subjects only need to perform the corresponding actions according to the action instructions, without having to imagine the corresponding motion scenario.

[0028] The MP paradigm can serve as the foundation for designing closed-loop BCI systems, controlling external devices such as exoskeletons and wheelchairs by decoding the EEG signals during the subject's motor preparation phase. For example, for stroke or cerebral palsy patients, although these users have significant motor impairments, they can still generate corresponding EEG signals during the motor preparation phase. Using appropriate EEG devices to control exoskeletons can accelerate the rehabilitation process for these patients. For healthy users, the MP paradigm can provide early response and enhance their motor abilities using external devices such as exoskeletons. The MP paradigm for acquiring EEG signals is non-invasive, convenient, low-cost, and low-latency, and holds promise for widespread application in practical and clinical rehabilitation scenarios.

[0029] Compared with the prior art, the present invention has the following obvious substantive features and significant advantages:

[0030] 1. This invention uses voice and text commands as a means of stimulation, which can simultaneously trigger neural activity in the auditory and visual areas of the cerebral cortex, thereby achieving coupling of EEG signals in the semantic processing area; the coupled EEG signals are further coupled with the EEG signals in the motor area, thereby effectively improving the quality of the overall EEG signals.

[0031] 2. Compared with the traditional MI paradigm, this invention is more intuitive. It uses the subject's native language to directly issue voice-text commands. The subject does not need to imagine the movement scene, and the voice-text commands directly correspond to the movement task. There is no command mapping, which is more in line with the subject's intuition.

[0032] 3. Compared with the traditional ME paradigm, this invention has less motion artifacts and less delay, which can improve the quality of EEG signals and the real-time response, and is more user-friendly for subjects with severe motor disorders who have difficulty performing corresponding actions.

[0033] 4. This invention can serve as the basis for designing a closed-loop BCI system, controlling external devices such as exoskeletons and wheelchairs by decoding the user's motion preparation signals. For users with severe movement disorders, the MP paradigm can be combined with external devices to accelerate their rehabilitation process; for healthy users, the MP paradigm can be combined with external devices to enhance their motor abilities.

[0034] 5. The MP paradigm proposed in this invention for acquiring EEG signals has the characteristics of being non-invasive, convenient, low-cost, and low-latency, and is expected to be promoted to practical applications and clinical rehabilitation scenarios. Attached Figure Description

[0035] Figure 1 These are schematic diagrams illustrating four different excitation methods of the present invention;

[0036] Figure 2 This is a schematic diagram of a subject performing an MP task under the guidance of video instructions, according to a preferred embodiment of the present invention;

[0037] Figure 3 This is a schematic diagram of the EEG electrode distribution in a preferred embodiment of the present invention;

[0038] Figure 4 This is a timing diagram of an experimental video of a preferred embodiment of the present invention. Detailed Implementation

[0039] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments and accompanying drawings. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0040] The Motor Preparation (MP) paradigm of this invention uses spoken and written commands as a stimulus, which can simultaneously trigger neural activity in the auditory and visual cortexes of the brain, thereby achieving coupling of EEG signals in the semantic processing area. The coupled EEG signals are then further synergistically coupled with the EEG signals in the motor cortex, effectively improving the overall quality of the EEG signals. Four different stimulus methods were used in the experiment, such as... Figure 1 As shown.

[0041] To verify the superiority of voice-text stimulation, this invention sets up four stimulation methods: voice-text, pure voice, pure text, and arrow. The video of issuing the command can be pre-recorded. During the experiment, the subject needs to maintain a standing posture, receive the command, and perform the corresponding action, such as... Figure 2 As shown. The MP paradigm task can be described as: focusing attention on the spoken text instructions emitted by the video and performing the corresponding actions according to the instructions. Before the experiment, the participants were explained the following three points to help them understand how to perform the MP task:

[0042] (1) You need to focus your attention on the instructions given in the video. (2) The order in which the different videos appear is completely random. (3) During the exercise preparation phase before the exercise is performed, do not move your body or contract your muscles, so as not to affect the acquisition of EEG signals during the exercise preparation phase.

[0043] Specific experiment:

[0044] Data acquisition was carried out using a 16-channel Cyton + Daisy Biosensing Boards system with a sampling frequency of 125 Hz. The distribution of brain electrodes is as Figure 3 shown. A total of 7 healthy subjects were recruited.

[0045] The acquisition experiment of the dataset included in the present invention consisted of two types of tasks: natural body rotations to the left and to the right. The timing diagram of one experiment is as Figure 4 shown, and the specific process is described as follows:

[0046] (1) At t = 0 s, a white cross appeared on the screen and lasted for 2 seconds.

[0047] (2) Subsequently, the white cross disappeared, and a body rotation instruction appeared randomly: for voice and text stimulation, the words "turn", "left" (or "right"), and "rotate" would appear on the screen in sequence, and a corresponding Mandarin voice would be emitted when each word appeared. The word "turn" lasted for 0.5 seconds, the word "left" (or "right") lasted for 2 seconds, and the word "rotate" lasted for 0.5 seconds. The words would continue to be displayed on the screen after they appeared; for pure voice stimulation, the screen remained black, and the body rotation instruction was emitted at the same time as that for voice and text stimulation; for pure text stimulation, the video did not emit sound, and the words "turn", "left" (or "right"), and "rotate" would also appear on the screen in sequence, with the appearance time being the same as that for voice and text stimulation; for arrow stimulation, the video would first go black for 0.5 seconds, then an arrow pointing to the left (or right) would appear and last for 2 seconds, and then the arrow would fade out to the left (or right) of the screen.

[0048] (3) When the "rotate" instruction was issued (or the arrow faded out of the screen), the subject performed a natural body rotation to the left (or right).

[0049] (4) Subsequently, all voices, texts, and arrows disappeared, and the screen remained black for 2 seconds for the subject to reset and rest, preparing to receive the instruction issued by the next video.

[0050] The production process of the dataset was as follows: for each subject, a total of 640 groups of data were collected in the experiment (each of the 4 stimulation types included 80 experiments, and each stimulation type corresponded to two movement instructions of turning left and turning right. For each subject, the collection of 640 groups of data was completed一次性完成). The duration of each group of data was 8 seconds, the sampling frequency was 125 Hz, the data length was 1000 points, and the total duration of the experiment was approximately 80 minutes. Before the formal experiment, each subject practiced MP for about 3 minutes to familiarize themselves with the experimental process. Subsequently, the originally collected electroencephalogram signals were preprocessed by filtering and cropping, and the dataset was obtained by splicing.

[0051] The filtering is performed using a Butterworth filter for bandpass filtering from 0.1 to 48 Hz; the cropping involves reducing the original 1000 data points to 200 points, corresponding to the data from the 2.5th to 4.1st second of the data acquisition process, to ensure that the final data includes information from the exercise preparation phase and is not interfered with by electromyographic signals; the stitching involves integrating the EEG data obtained from each subject in 80 experiments into a complete data file to facilitate subsequent training and testing of the neural network.

[0052] The motion preparation paradigm decoding method of this invention is based on the DoNet dual-point observation network proposed by Niu Xu et al. (Model and Data Dual-Driven Double-Point Observation Network for Ultra-Short MIEEEG Classification, IEEE J. Biomed. Health Inform., 28(6), 3434–3445). First, dual-point pairs of narrowband signals with low signal-to-noise ratio ultra-short time are sampled. Feature vectors are calculated based on the sampled values. Finally, the predicted labels output by the neural network are weighted and voted on. The weighted voting process is as follows:

[0053] For each predicted label output by the neural network, there is a corresponding weight. This weight is allocated according to its proximity to the motion execution period; signals closer to the execution period are assigned a larger voting weight. The weight increases monotonically over time, gradually increasing from 0.5 to 1. The weight increase process includes three methods: linear increase, convex function increase, and concave function increase, the specific expressions of which are shown below:

[0054] Linear functions:

[0055] Convex function:

[0056] Concave function:

[0057] Among them, w i The weight of the i-th prediction in the voting decision-making process is given by N, where N is the total number of predictions, and p and k are parameters used to control the steepness of the curve.

[0058] The training and performance evaluation experiments were conducted 10 times on each subject, with 600 generations per training session. In the weighted increasing function, both p and k in the convex and concave functions were set to 2. The average classification accuracy is shown in Table 1, in percentage.

[0059] Table 1

[0060]

[0061] As can be seen from Table 1, the MP paradigm proposed in this invention achieves significantly higher classification accuracy in the vast majority of subjects (except S7) compared with traditional arrow excitation.

Claims

1. A practical few-channel dry electrode brain-computer interface paradigm, implemented based on a few-channel dry electrode brain-computer interface device, characterized in that, The video is pre-recorded, and the recording method is as follows:

2. The practical few-channel dry electrode brain-computer interface paradigm of claim 1, wherein, A kind of action instruction appears randomly, and the action instruction is displayed in text on the screen, or the voice content of the text is played synchronously, or the voice content of the text is played while the text is displayed; the subject executes the action instruction after the instruction display or play is completed; the electroencephalogram signal of the subject in the movement preparation stage is collected. The execution of the corresponding action includes natural rotation to the left and right, and stepping forward with the left foot and the right foot.

3. The practical few-channel dry electrode brain-computer interface paradigm according to claim 1 or 2, wherein, The execution of the task includes:

4. The practical few-channel dry electrode brain-computer interface paradigm of claim 1, wherein, Step 1: the subject focuses on the instruction issued in the video; Step 2: the subject executes the corresponding action according to the instruction issued in the video; Step 3: if the subject correctly completes the corresponding action according to the instruction, it indicates that the task has been completely understood and correctly executed; after the action is completed, the subject returns to the initial position and prepares to receive the next instruction. Based on the double-point observation network DoNet, the double-point pair of the narrow-band signal of the low signal-to-noise ratio ultra-short time signal collected in the movement preparation stage is sampled, the feature vector is calculated according to the sampling value, and finally the prediction label output by the neural network is weighted and voted to obtain the action category.

5. The practical few-channel dry electrode brain-computer interface paradigm derived electroencephalography signal decoding method of claim 1, wherein, The double-point observation network DoNet is trained using the data set corresponding to the paradigm, and the data set contains the excitation mode of voice and text, pure voice and pure text arrow, and the corresponding action; the original collected electroencephalogram signal is filtered, cropped and pretreated, and spliced to obtain the data set.

6. The decoding method of claim 5, wherein, The filtering is 0.1-48Hz band-pass filtering using a Butterworth filter; the cropping is to crop the original 1000-point data to 200 points, corresponding to the data from 2.5 seconds to 4.1 seconds in the data collection process, to ensure that the final data contains information of the movement preparation stage and is not disturbed by the electromyogram signal; 7. The decoding method of claim 6, wherein, The splicing is to integrate the electroencephalogram data obtained by each subject in 80 experiments into a complete data file; wherein the original collection of electroencephalogram signal data is: for each subject, each excitation type contains 80 experiments, and each excitation type corresponds to two movement instructions of left turn and right turn, each experiment obtains two groups of data corresponding to the two movement instructions, each group of data has a duration of 8 seconds, a sampling frequency of 125Hz, and a data length of 1000 points, and the total duration of the experiment is 80 minutes. The weighted voting decision assigns weights to the classification results of the neural network according to the proximity to the movement execution period, and the closer the signal to the movement execution period, the greater the voting weight.

8. The decoding method of claim 5, wherein, The voting weight monotonically increases with time, gradually increasing from 0.5 to 1.

9. The decoding method of claim 8, wherein, The weight increasing process includes linear increasing, convex function increasing and concave function increasing, and the specific expressions are as follows:

10. The decoding method of claim 8 or 9, characterized by, ​ Linear function: Convex function: Concave function: where w i where w is the weight of the i-th prediction in the voting decision process, N is the total number of predictions, and p and k are parameters used to control the steepness of the curve.

Images