A multi-person collaborative rehabilitation training method and system based on motor imagery neurofeedback

Abstract

The application discloses a kind of multi-person collaborative rehabilitation training method and system based on motor imagination neural feedback.The system includes n user host computers, data processing module, neural feedback module, user training process record module;Electroencephalogram acquisition equipment collects the electroencephalogram data of each user;The neural feedback module controls the feedback synchronization of n subjects, and controls the synchronous acquisition of n electroencephalogram acquisition equipment, and realizes audiovisual touch multimodal neural feedback;The data processing module includes preprocessing module, feature extraction and multi-brain fusion decision module;The preprocessing module receives the electroencephalogram data collected by n electroencephalogram acquisition equipment, and pre-processes the collected electroencephalogram data.The feature extraction and multi-brain fusion decision module is used to obtain the ERD result of subject.The user training process record module records the experimental training information of subject each time, which can be used for data offline analysis, and adaptively adjusts user identity information.The application improves the practicability, stability, social interaction of rehabilitation system, improves user experience and immersion, so that the treatment effect of neural feedback rehabilitation training can be significantly improved.

Description

Technical Field

[0001] This invention belongs to the field of EEG signal recognition in the field of biometric recognition, specifically relating to a multi-person collaborative rehabilitation training method and system based on motor imagery neurofeedback. Background Technology

[0002] Cerebral stroke, also known as stroke or cerebrovascular accident (CVA), is characterized by four highs: high incidence, high disability rate, high mortality rate, and high recurrence rate. Early and intensive rehabilitation training in the early stages of stroke is crucial for patients to rebuild bodily function and reduce disability. During rehabilitation, long-term, regular limb and speech rehabilitation training can help induce plasticity changes in the corresponding brain regions, improving patients' independent activity abilities.

[0003] Traditional rehabilitation therapy often involves having patients perform specific limb movements to promote muscle strength recovery and motor nerve reconstruction. However, many patients struggle to perform these movements due to damage to their closed-loop neural circuits. Based on the principle of "neural plasticity," recent studies have demonstrated that using "motor imagery" instead of "motor execution" can also aid in the recovery and reconstruction of the nervous system in motor patients. Rehabilitation training can be conducted by combining external devices (such as walking aids and robotic arms) with motor imagery to restore the patient's damaged nervous system. However, this method cannot capture the patient's brain activity, thus lacking effective and timely feedback, resulting in less than satisfactory training outcomes.

[0004] Motor imagery neurofeedback brain-computer interfaces can acquire patients' neural signals in real time, convert them into control signals, and provide timely feedback in the form of vision, touch, and hearing, thus forming a closed-loop system that can more effectively improve the activation level of the corresponding cerebral cortex and achieve better training results.

[0005] The effectiveness of motor imagery neurofeedback brain-computer interfaces in rehabilitation training depends to some extent on the form of neurofeedback. Previous studies have demonstrated the effectiveness of individual visual, auditory, and tactile feedback. However, relying on a single form of feedback can be monotonous. Currently, more and more researchers are combining multiple feedback forms, integrating them with neural information detection methods and their response mechanisms, to strengthen the coupling effect between each feedback and the triggering process. This aims to maximize the matching efficiency between specific neural functions and peripheral motor functions (i.e., cortex-muscle), thereby achieving optimal motor rehabilitation feedback training results.

[0006] Although multimodal neurofeedback training has achieved relatively good training results, limitations in EEG acquisition and signal processing technologies mean that most experiments currently involve single subjects training independently, lacking social interaction. This leads to problems such as a monotonous training mode and low user participation, manifested in subjects showing noticeable inattention and drowsiness in the latter half of the experiment. Furthermore, patients may be resistant to unfamiliar technology, making it difficult to find willing participants for treatment. In recent years, the field of cognitive science has proposed a novel multi-person synchronous EEG acquisition technology: hyperscanning technology.

[0007] In summary, this invention proposes a multi-person collaborative rehabilitation training system and method based on motor imagery neurofeedback. The multi-person training program can effectively enhance the system's practicality, stability, and social interactivity, improve user experience and immersion, and better restore the subject's motor imagery ability. It can be applied to the collaborative rehabilitation treatment of certain people with physical disabilities or injuries. Summary of the Invention

[0008] The purpose of this invention is to address the shortcomings of existing neurofeedback training systems and methods by providing a multi-person collaborative online rehabilitation training method and system based on motor imagery neurofeedback. This aims to improve the practicality, interactivity, and stability of the rehabilitation system, enhance the user experience and immersion, and significantly improve the therapeutic effect of neurofeedback rehabilitation training.

[0009] To achieve the above objectives, firstly, a multi-person collaborative rehabilitation training method based on motor imagery neurofeedback is provided, comprising the following steps:

[0010] Step S1: Experimental Preparation

[0011] Each participant was equipped with an EEG acquisition device;

[0012] The pre-defined experimental paradigm consists of one task with four runs, each with ample rest time between runs. Each run is divided into 25 trials, each lasting 10 seconds, for a total of 250 seconds per run. Each trial is further subdivided into the following phases:

[0013] 5) Resting state phase: The virtual scene displayed on the screen is a typical family bedroom. The virtual character stands on the ground, holding a hand gripper in each hand for 2 seconds.

[0014] 6) Task pre-stage: lasts 1 second; an arrow will randomly appear on the screen pointing to the left or right hand, indicating that the task to be performed is a left-hand task or a right-hand task; then, depending on the task type and the number of participants, the number of energy bars and energy slots corresponding to the number of participants will appear simultaneously in the upper left and upper right corners of the screen.

[0015] 7) Task-state feedback phase: The gripper applies continuous grip force for 4 seconds;

[0016] 8) Results presentation and adaptive adjustment phase;

[0017] Step S2: Data Acquisition

[0018] The neurofeedback module controls the synchronization of visual, auditory, and tactile feedback for n subjects, and uses n EEG acquisition devices to synchronously acquire the EEG signals of n subjects, n≥2;

[0019] Step S3: Data Preprocessing

[0020] The data processing module receives EEG data from n EEG acquisition devices and preprocesses the acquired EEG data.

[0021] Step S4: Feature Extraction and Decision Learning

[0022] S41: Extracting AR model parameters based on the Burg algorithm; specifically:

[0023] First, the AR model, also known as the autoregressive model, is an all-pole model, represented by the following difference equation:

[0024]

[0025] Where x(n) is the EEG signal with n sampling points, a i These are the coefficients of the AR model, where p is the order of the AR model, and ω(n) represents the error, which can generally be assumed to have a variance of σ. 2 The noise is zero-mean white noise. The transfer function of the AR model system H(z) is:

[0026]

[0027] Thus, the formula for calculating the power spectrum estimation of the AR model is obtained:

[0028]

[0029] As can be seen from equation (3), to obtain the power spectral density p of a random signal using the AR model... x (ω), the model parameter a must be obtained. k variance of white noise sequence

[0030] Estimating AR model parameters a using the Burg algorithm k and The specific steps are as follows:

[0031] 13) Let f p (n) and b p(n) represent the forward prediction error and backward prediction error of the linear prediction AR model, respectively:

[0032] f p (n)=x(n)+a p1 x(n-1)+...+a pp x(np) Equation (4)

[0033] b p (n)=x(np)+a p1 x(n-p+1)+...+a pp (n) Equation (5)

[0034] Where x(np) represents a term of equation (1), a pi Represents the i-th term of order p;

[0035] 14)P f P b These are the backward prediction error power and the forward prediction error power, respectively.

[0036]

[0037]

[0038] Where H represents the conjugate transpose;

[0039] 15) Let the sum of the forward and backward prediction error powers be:

[0040]

[0041] 16) When the order s increases from 1 to p, f p (n) and b p (n) has the following recurrence relation:

[0042] f s (n)=f s-1 (n)+h s b s-1 (n-1) Equation (7)

[0043] b s (n)=b s-1 (n)+h s b s-1 (n-1) Equation (8)

[0044] And f0(n)=b0(n)=x(n), h s Let h be the reflection coefficient, and h s =a ss .

[0045] 17) The Burg algorithm is such that the sum of the forward and backward prediction error powers Pfb Relative to the reflection coefficient h s Minimum, thus obtaining h s estimate:

[0046]

[0047] 18) h is estimated from equation (9) s Then, at order s, the AR model coefficients are obtained using the Levinson-Durbin recursive algorithm:

[0048] a s，i =a s-1，i +h s a s-1，s-i Equation (10)

[0049] a ss =h s Equation (11)

[0050]

[0051] S42: Substitute the AR model parameters obtained in step S41 into equation (3) to obtain the power spectral density function p. x (ω).

[0052] S43: Calculate the average power of the Mu rhythm (8-12Hz) of the EEG signal at rest for all subjects, and take the mean as the baseline value P. base Specifically, the resting-state power P is obtained by integrating the 8-12Hz EEG data using the power spectral density function. rest 1, P rest 2...P rest (M+N), and calculate their mean as the baseline value P. base :

[0053]

[0054] Where n = M + N, M represents the number of expert subjects, 0 ≤ M ≤ n, and N represents the number of novice subjects, 0 ≤ N ≤ n;

[0055] S44: Calculate the average power of the EEG signals in the task state for all subjects in the 8-12Hz range. In the collaborative mode, obtain the total power P in the task state according to formula (12). end for

[0056]

[0057] Where Ai represents P oldi The power percentage, Bj represents P newi Power percentage; P oldi P newiThese represent the real-time task-state power output by experts and novices, respectively.

[0058] S45: The total power P in the task state... end Compared with baseline value P base For comparison, first compare the attention metric, Focus. task With reference value Focus base If Focus task <Focus base If the focus is on a certain value, the output will be 0; if the focus is on a certain value, the output will be 0. task >=Focus base， Furthermore, if the average power of the EEG channel on the opposite side of the task (C4 channel for the left hand task, and C3 channel for the opposite side) is lower than the baseline value in the 8-12Hz range, it is considered that the event-related desynchronization (ERD) phenomenon has occurred, and the output result is 1; otherwise, the output result is 0.

[0059] Step S5: Multimodal Neural Feedback

[0060] S51: Provide real-time feedback based on the output of step S4: If the output of step S4 is 1, the gripper tightens downwards by an angle θ; if the output of step S4 is 0, the gripper angle remains unchanged. At the same time, based on the average power and attention index output by S4, update the energy bars and energy slots in the upper left and upper right corners in real time to provide real-time visual feedback on the brain activity status of each subject and other collaborating subjects.

[0061] S52: Calculate the comprehensive index of each subject within the time period T, and select the subject with the lowest comprehensive index to receive tactile stimulation.

[0062] Overall Index = (Average Energy over Time T) / (Attention Index over Time T)

[0063] S53: Based on the task completion rate and the comprehensive index of each subject, the subjects are ranked in descending order. The subjects with the highest and lowest rankings are given auditory stimulation, and the user identity information is adaptively adjusted. The user identity information is either a novice or an expert.

[0064] Secondly, a multi-user collaborative online rehabilitation training system based on motor imagery neurofeedback is provided, including n user hosts, a data processing module, a neurofeedback module, and a user training process recording module, where n≥2; each user host is equipped with an EEG acquisition device; the EEG data acquired by the EEG acquisition device is transmitted to the corresponding user host;

[0065] The EEG acquisition device acquires the EEG signals of the corresponding subject;

[0066] The neurofeedback module controls the synchronous audiovisual and tactile feedback of n subjects and the synchronous acquisition of data by n EEG acquisition devices; it also realizes multimodal neurofeedback; the multimodal neurofeedback includes real-time feedback from the grip strength device based on the output results of the feature extraction and decision learning module.

[0067] The data processing module includes a preprocessing module and a feature extraction and multi-brain fusion decision-making module.

[0068] The preprocessing module receives EEG data collected from n EEG acquisition devices and performs preprocessing on the collected EEG data.

[0069] The feature extraction and multi-brain fusion decision module is used to obtain the event-related desynchronization (ERD) results of the subjects.

[0070] Preferably, the neurofeedback module also provides tactile stimulation to the subject with the lowest comprehensive index based on the comprehensive index of each subject within the time period T.

[0071] Preferably, the neurofeedback module also sorts the subjects in descending order based on the completion of the task and the comprehensive indicators of each subject, provides auditory stimulation to the subjects with the highest and lowest rankings respectively, and adaptively adjusts the user identity information.

[0072] The user training process recording module records the subject's training information for each experiment, which can be used for offline data analysis and adaptive adjustment of user identity information.

[0073] Thirdly, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed in a computer, causes the computer to perform the method described thereon.

[0074] Fourthly, a computing device is provided, including a memory and a processor, wherein the memory stores executable code, and the processor executes the executable code to implement the method.

[0075] The present invention has the following advantages and effects compared with the prior art:

[0076] 1) The multi-person collaborative neurofeedback rehabilitation training method provided by the present invention overcomes the mechanical passive mode of traditional clinical rehabilitation training and adds an active rehabilitation mode of user intention decoding.

[0077] 2) Multi-person collaborative neurofeedback rehabilitation training innovatively introduces interactive tasks with multiple participants sharing the power, and introduces a competitive mechanism in which "experts" influence "newbies," which fills the gap in the lack of social interactivity in neurofeedback systems and enhances user participation.

[0078] 3) This invention proposes a multimodal perception neural feedback method, which helps to integrate multiple sensory information such as vision, hearing and touch to influence the user's neural remodeling, manifested as the modulation of event-related synchronization / desynchronization in the α band of the motor area. Attached Figure Description

[0079] Figure 1 This is a system module block diagram of a specific embodiment of the present invention;

[0080] Figure 2 This is a system flowchart illustrating a specific embodiment of the present invention;

[0081] Figure 3 This is a schematic diagram of a multi-person collaborative rehabilitation training scenario based on motor imagery neurofeedback, as described in this invention. Detailed Implementation

[0082] The present invention will be further analyzed below with reference to specific embodiments and accompanying drawings.

[0083] The system block diagram of this invention is attached. Figure 1 As shown in the diagram, the experimental paradigm was first designed, and appropriate parameters were set for each module of the system. Then, the experiment officially began. The user host simultaneously acquired EEG signals from multiple participants. The data processing module preprocessed the EEG signals, then extracted brain features related to motor imagery from the preprocessed data using appropriate methods, calculated attention indices, and further converted the features into corresponding control signals. Finally, the control signals controlled the visual, tactile, and auditory modules in real time and provided feedback to multiple participants. All hosts in the system communicated using socket network connections.

[0084] See attached document Figure 2 With appendix Figure 3 The specific implementation steps of the present invention are as follows:

[0085] Step S1: Experimental Preparation

[0086] First, based on the number of participants, the number of user terminals was determined to receive and store the raw EEG data during the experiment. For this experimental case, 10 participants were recruited, aged 20-25 years, with no history of neurological diseases or limb disabilities. Five participants had no prior experience with BCI experiments based on motor imagery, while the other five had received training in the classic Graz motor imagery task and could perform motor imagery well. Before the experiment began, the experimental procedure was clearly explained to each participant. The 10 participants were divided into 5 groups of two.

[0087] Then, filter parameters were set for the data processing module: a notch filter at 50Hz and a bandpass filter between 0.5-45Hz. Finally, the neurofeedback module set the timing and duration of feedback, and set the required threshold for feedback, based on the pre-designed experimental paradigm. For collaborative tasks, participants were initially designated as either experts or novices; the system automatically adjusted user identities during the task.

[0088] Step S2: Data Acquisition

[0089] Multiple EEG acquisition devices were used to collect the required EEG signals. In this example, two 64-channel Neuroscan devices were used to acquire EEG signals at a sampling frequency of 1000Hz. The electrode caps were placed using the international 10 / 20 system electrode placement method. The 64 electrodes are as follows: FP1, FPz, FP2, AF3, AF4, F7, F5, F3, F1, Fz, F2, F4, F6, F8, FT7, FC5, FC3, FC1, FCz, FC2, FC4, FC6, FT8, T7, C5, C3, C1, Cz, C2, C4, C6, T8, TP7, CP5, CP3, CP1, CPz, CP2, CP4, CP6, TP8, P7, P5, P3, P1, Pz, P2, P4, P6, P8, PO7, PO3, POz, PO4, PO8, O1, CB1, Oz, CB2, O2, M1, M2. An EKG electrode was placed on the left wrist artery to record electrocardiogram data, while a reference electrode was placed on the right earlobe. The two participants in the experimental group sat comfortably in armchairs approximately 1 meter away from the 24-inch LCD screen throughout the experiment, simultaneously undergoing training.

[0090] When the start control signal of the neurofeedback module is received, the user host begins to synchronously forward the collected data to the data processing module for processing, and saves the user's original EEG data locally.

[0091] Step S3: Data Preprocessing

[0092] This experiment focuses on the data processing module. Its network I / O model employs the efficient event-driven Reactor pattern, using I / O multiplexing and thread pool technology to process EEG data from different PCs. The main thread only listens for events, while worker threads handle the data processing. After receiving data from step 2, the acquired EEG data undergoes preprocessing, including downsampling, notch filtering, bandpass filtering, and spatial filtering (removing the average reference).

[0093] 1) Downsampling: The original signal is downsampled at 250Hz during the data preprocessing process.

[0094] 2) Notch filter: Use a 50Hz notch filter to filter out high-frequency noise, baseline drift, electromagnetic interference, etc.

[0095] 3) Bandpass filter: Use a 0.5-45Hz bandpass filter to acquire data in the frequency band related to motion imagery.

[0096] 4) Spatial filter: First, the average value of the EEG data from the electrodes around C3 and C4 is calculated, and then the average value is subtracted from the data of each channel.

[0097] Step S4: Feature Extraction and Classification

[0098] This experiment employs a sensorimotor rhythm paradigm. During motor imagery, ERD (Electroencephalopathy) occurs in the contralateral channel at 8-12 Hz. The frequency domain is the most commonly used method for processing such signals. Currently, various methods can be used to convert EEG time signals into frequency domain signals, such as Fast Fourier Transform (FFT), wavelet transform, and autoregressive parametric models. Since spectral estimation based on autoregressive models (AR) offers better temporal resolution and real-time performance for a given frequency, we use this method to extract the corresponding frequency domain features.

[0099] We use power spectral estimation based on the AR model and the Burg algorithm to extract the parameters of the AR model. First, the AR model, also known as the autoregressive model, is an all-pole model, which can be represented by the following difference equation:

[0100] First, the AR model, also known as the autoregressive model, is an all-pole model, which can be represented by the following difference equation:

[0101]

[0102] Where x(n) is the EEG signal with n sampling points, a i These are the coefficients of the AR model, where p is the order of the AR model, and ω(n) represents the error, which can generally be assumed to have a variance of σ. 2 The noise is zero-mean white noise. The transfer function of the AR model system H(z) is:

[0103]

[0104] Thus, the formula for calculating the power spectrum estimation of the AR model is obtained:

[0105]

[0106] As can be seen from equation (3), to obtain the power spectral density p of a random signal using the AR model... x (ω), the model parameter a must be obtained. k variance of white noise sequence

[0107] Estimating AR model parameters a using the Burg algorithm k and The specific steps are as follows:

[0108] 19) Let f p (n) and b p (n) represent the forward prediction error and backward prediction error of the linear prediction AR model, respectively:

[0109] f p (n)=x(n)+a p1 x(n-1)+...+a pp x(np) Equation (4)

[0110] b p (n)=x(np)+a p1 x(n-p+1)+...+a pp (n) Equation (5)

[0111] 20)P f P b These are the backward prediction error power and the forward prediction error power, respectively.

[0112]

[0113]

[0114] 21) Let the sum of the forward and backward prediction error powers be:

[0115]

[0116] 22) When the order s changes from 1 to p, f p (n) and b p (n) has the following recurrence relation:

[0117] f s (n)=f s-1 (n)+h s b s-1 (n-1) Equation (9)

[0118] b s (n)=b s-1 (n)+h s b s-1 (n-1) Equation (10)

[0119] And f0(n)=b0(n)=x(n), h s Let h be the reflection coefficient, and h s =a ss .

[0120] 23) The Burg algorithm is to make the sum of the forward and backward prediction error powers P fb Relative to the reflection coefficient h s The minimum value can be obtained from h. s Estimation formula:

[0121]

[0122] 24) h can be estimated from the above formula. s Then, at order s, the AR model coefficients are obtained using the Levinson-Durbin recursive algorithm:

[0123] a s，i =a s-1，i +h s a s-1，s-i Equation (12)

[0124] a ss =h s Equation (13)

[0125]

[0126] S42: Substitute the AR model parameters obtained in step S41 into equation (3) to obtain the power spectral density function p. x (ω).

[0127] S43: Calculate the average power of the C3 and C4 channel EEG signals at rest for all subjects in the 8-12Hz (Mu rhythm), and take the mean as the baseline value P for judging whether motor imagery has occurred. base Specifically, the power P in the resting state is calculated based on the power spectral density function using data from 8-12Hz. rest 1, P rest 2...P rest (M+N), and calculate their mean as the baseline value P. base :

[0128]

[0129] Where n = M + N, M represents the number of expert subjects, 0 ≤ M ≤ n, and N represents the number of novice subjects, 0 ≤ N ≤ n;

[0130] S44: Calculate the average power P of all channels of EEG signals in the resting state for all subjects within the high beta (21-30Hz) frequency range. high-β The average power P in the theta (4-8Hz) frequency range θ The average power P in the low alpha (8-10Hz) frequency range low-α Calculate the baseline attention metric, Focus. base :

[0131]

[0132] S45: Calculate the average power of the C3 and C4 channel EEG signals at 8-12Hz for all subjects in the task state. In the collaborative mode, obtain the total power P in the task state according to formula (12). end for

[0133]

[0134] Where Ai represents P oldi The power percentage, Bj represents P newi The power percentage of P is the preset value. oldi P newi These represent the real-time task-state power output by experts and novices, respectively.

[0135] S46: Using the same method as formula (16), the focus of the EEG signals of all subjects under the task state was also calculated. task .

[0136] S45: The total power P in the task state... end Compared with baseline value P base For comparison, first compare the attention metric, Focus. task With reference value Focus base If Focus task <Focus base If the focus is on a certain value, the output will be 0; if the focus is on a certain value, the output will be 0. task >=Focus base If the average power of the 8-12Hz EEG channel on the opposite side of the task is lower than the baseline value, it is considered that an event-related desynchronization (ERD) has occurred, and the output result is 1; otherwise, the output result is 0. Finally, we synchronously send the output 0, 1 and all feature information to the neurofeedback module for online display and feedback.

[0137] Step S5: Multimodal Neural Feedback

[0138] According to the pre-defined experimental paradigm, a task consists of 4 runs, with ample rest time between each run. Each run is divided into 25 trials, each lasting 10 seconds, for a total of 250 seconds per run.

[0139] A trial is divided into four phases:

[0140] 1) The initial virtual scene displayed on the screen is a typical family bedroom, with the virtual character standing on the ground, holding a hand gripper in each hand for 2 seconds;

[0141] 2) Then, an arrow will randomly appear on the screen pointing to either the left or right hand, indicating whether the task to be performed is a left-hand task or a right-hand task. Based on the task type and the number of users, a number of motion visualization energy bars corresponding to the number of users will appear in the upper left corner of the screen, and the same number of user concentration energy bars will appear in the upper right corner, initialized to 0%. This display phase lasts for 1 second.

[0142] 3) After 1 second, the feedback phase begins. The entire feedback process lasts for 4 seconds, during which the pressure progress of the grip strengthener changes over time. Using the signal from step S4, the grip strengthener's progress is updated every 200 milliseconds, with each update involving a steady decrease of approximately 0% to 5% in pressure.

[0143] Specifically:

[0144] 1. If the control signal is 0, then the pressure is 0%.

[0145] 2. If the control signal is 1, and P base <P end <1.2*P base Then reduce pressure by 2-3%;

[0146] 3. If the control signal is 1, and P end >=1.2*P base Then reduce by 5%;

[0147] During this period, the energy bar in the upper left corner will update every second in real time, allowing each user to see their own energy percentage. The subject number is k, and the output power is P. k And the energy percentage with weight w is

[0148]

[0149] The energy bar for attention concentration in the upper right corner is also updated every second. This continuous visual induction and feedback allows subjects to form long-term memories, thereby enhancing intercortical interactions and effectively promoting the recovery of damaged neural functions.

[0150] In addition, every second the system will count the subject with the lowest comprehensive index during that period. The electrical stimulation device will release a non-painful pulse current to stimulate the subject's task hand, reminding the subject to concentrate. Tactile stimulation will also induce reverse nerve impulses in motor nerve fibers, promoting synaptic repetition and postsynaptic coupling of neurons.

[0151] The specific calculation method for the comprehensive index is as follows: Comprehensive index = (average energy within 1 second) / (average concentration within 1 second).

[0152] Concentration is: The lower the value of this indicator, the more focused the user is when performing a motion visualization task.

[0153] 4) Finally, based on task completion and each user's overall performance, a ranking is displayed. The user with the lowest overall performance is encouraged and motivated, while the user with the highest overall performance is congratulated. The system adaptively adjusts user identities. This positive auditory feedback provides a positive reinforcement effect, stimulating the brain to release dopamine, a neurotransmitter that conveys feelings of excitement and happiness to the participants. This process lasts for 3 seconds.

[0154] Step S6: User training progress recording and system adaptive adjustment

[0155] Throughout the experiment, the user training process recording module will record the complete user information of each subject in this experiment, including: username, password, user training time, user training mode, energy percentage in each round, and concentration.

[0156] The system's adaptive adjustment module first loads the user's previous training information before each experiment begins, and then uses the user's most recent training information to comprehensively compare information from multiple users to determine whether the user is initially an expert or a novice in this round of experiment.

[0157] The specific determination method is as follows:

[0158] 1. If this is the user's first time training, they will be automatically identified as a novice player.

[0159] 2. If a user has participated in training before, the system will automatically compare the past training information of multiple users.

[0160] 1) Compare the number of training sessions; the one with more training sessions is the expert.

[0161] 2) If the number of training sessions is the same, compare the average energy ratio; the one with the higher ratio is an expert, otherwise a novice.

[0162] Secondly, during the system experiment, after every 5 trials, the rankings are recalculated based on comprehensive indicators, automatically determining the top M as experts and the top N as novices, thus maintaining a cooperative yet competitive relationship among users.

Claims

1. A multi-person collaborative online rehabilitation training system based on motor imagery neurofeedback, characterized in that... It includes n user hosts, a data processing module, a neurofeedback module, and a user training process recording module, where n≥2; each user host is equipped with one EEG acquisition device; The EEG data collected by the EEG acquisition device is transmitted to the corresponding user's host computer; The EEG acquisition device acquires the EEG signals of the corresponding subject; The neurofeedback module controls the simultaneous visual, auditory, and tactile feedback of n subjects and controls the simultaneous acquisition of EEG data by n EEG acquisition devices; and realizes multimodal neurofeedback. The data processing module includes a preprocessing module and a feature extraction and multi-brain fusion decision-making module. The multimodal neural feedback includes real-time feedback from the grip strength device based on the output results of the feature extraction and multi-brain fusion decision-making module. The preprocessing module receives EEG data collected from n EEG acquisition devices and preprocesses the collected EEG data. The feature extraction and multi-brain fusion decision-making module is used to obtain the event-related desynchronization (ERD) results of the participants; specifically: S41: Extracting AR model parameters based on the Burg algorithm; specifically: First, the AR model, also known as the autoregressive model, is an all-pole model, which can be represented by the following difference equation: Equation (1) in For EEG signals with n sampling points, These are the coefficients of the AR model, where p is the order of the AR model. To represent the error, let's assume the variance is... Zero-mean white noise; AR model system The transfer function is: Equation (2) Thus, the formula for calculating the power spectrum estimation of the AR model is obtained: Equation (3) As can be seen from equation (3), to obtain the power spectral density of a random signal using the AR model... Model parameters must be obtained. variance of white noise sequence ; Estimating AR model parameters using the Burg algorithm The specific steps are as follows: 1) Let f p (n) and b p (n) represent the forward prediction error and backward prediction error of the linear prediction AR model, respectively: Equation (4) ) (n) Equation (5) in This represents a term in expression one. Represents the i-th term of order p; 2) , These are the backward prediction error power and the forward prediction error power, respectively. Equation (6) Equation (7) Where H represents the conjugate transpose; 3) Let the sum of the forward and backward prediction error powers be: Equation (8) 4) When the order s changes from 1 to p, and (n) has the following recurrence relation: Equation (9) Equation (10) and , Let be the reflection coefficient, and = ; 5) The Burg algorithm is to sum the forward and backward prediction error powers. Relative to the reflection coefficient Minimum, from which we can obtain Estimation formula: Equation (11) 6) The result estimated from the above formula is... Then, at order s, the AR model coefficients are obtained using the Levinson-Durbin recursive algorithm: Equation (12) Equation (13) Equation (14) S42: Substitute the AR model parameters obtained in step S41 into equation (3) to obtain the power spectral density function. ; S43: Calculate the average power of the resting-state EEG signals (8-12Hz, Mu rhythm) of all subjects, and take the mean as the baseline value P. base Specifically: The resting power P is obtained by integrating the 8-12Hz (Mu rhythm) EEG data using the power spectral density function. rest 1,P rest 2...P rest (M+N), and calculate their mean as the baseline value P. base : Equation (15) Where n = M + N, M represents the number of expert subjects, 0 ≤ M ≤ n, and N represents the number of novice subjects, 0 ≤ N ≤ n; S44: Calculate the average power of the EEG signals in the task state for all subjects in the 8-12Hz range. In the collaborative mode, obtain the total power P in the task state according to formula (16). end for Equation (16) in express The power percentage, express The power percentage mentioned above is a preset value; , These represent the real-time task-state power output by experts and novices, respectively. S45: The total power P in the task state... end Compared with baseline value P base For comparison, first compare attention metrics. Compared with the benchmark value ,like < If the output is 0, then the output will be 0; if If the average power of the 8-12Hz EEG channel on the opposite side of the task is lower than the baseline value, it is considered that the event-related desynchronization (ERD) phenomenon has occurred, and the output result is 1; otherwise, the output result is 0.

2. The system according to claim 1, characterized in that... The neurofeedback module also provides tactile stimulation to the subject with the lowest comprehensive index within time period T.

3. The system according to claim 1, characterized in that... The neurofeedback module also sorts participants in descending order based on task completion and overall performance, provides auditory stimulation to the highest and lowest ranked participants, and adaptively adjusts user identity information; the user identity information is either a novice or an expert.

4. The system according to claim 1, characterized in that... The user training process recording module records the subject's training information for each experiment, analyzes the experimental data, and adaptively adjusts the user identity information; the user identity information is either a novice or an expert.

5. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed in the computer, the computer performs the following methods: Step S1: Experimental Preparation Each participant was equipped with an EEG acquisition device; The pre-defined experimental paradigm consists of one task with four runs, each with ample rest time between runs. Each run is divided into 25 trials, each lasting 10 seconds, for a total run duration of 250 seconds. Each trial is further subdivided into the following phases: 1) Resting state phase: The virtual scene displayed on the screen is a typical family bedroom. The virtual character stands on the ground, holding a hand gripper in each hand for 2 seconds. 2) Task pre-processing phase: lasts 1 second; An arrow will randomly appear on the screen pointing to the left or right hand, indicating whether the task to be performed is a left-hand task or a right-hand task; at the same time, according to the task type and the number of participants, the number of energy bars and energy slots corresponding to the number of participants will appear on the screen. 3) Task-state feedback phase: The gripper applies continuous gripping force, and the energy bar and energy gauge are updated in real time for 4 seconds; 4) Results presentation and adaptive adjustment phase; Step S2: Data Acquisition The neurofeedback module controls the synchronization of visual, auditory, and tactile feedback for n subjects, and uses n EEG acquisition devices to synchronously acquire the EEG signals of n subjects, n≥2; Step S3: Data Preprocessing The data processing module receives EEG data collected from n EEG acquisition devices and preprocesses the collected EEG data. Step S4: Feature Extraction and Multi-Brain Fusion Decision Making S41: Extracting AR model parameters based on the Burg algorithm; specifically: First, the AR model, also known as the autoregressive model, is an all-pole model, represented by the following difference equation: Equation (1) in For EEG signals with n sampling points, These are the coefficients of the AR model, where p is the order of the AR model. To represent the error, let's assume the variance is... Zero-mean white noise; AR model system The transfer function is: Equation (2) Thus, the formula for calculating the power spectrum estimation of the AR model is obtained: Equation (3) As can be seen from equation (3), to obtain the power spectral density of a random signal using the AR model... Model parameters must be obtained. variance of white noise sequence ; Estimating AR model parameters using the Burg algorithm The specific steps are as follows: 7) Let f p (n) and b p (n) represent the forward prediction error and backward prediction error of the linear prediction AR model, respectively: Equation (4) ) (n) Equation (5) in A term of expression (1) Represents the i-th term of order p; 8) , These are the backward prediction error power and the forward prediction error power, respectively. Equation (6) Equation (7) Where H represents the conjugate transpose; 9) Let the sum of the forward and backward prediction error powers be: Equation (8) 10) When the order s changes from 1 to p, and (n) has the following recurrence relation: Equation (9) Equation (10) and , Let be the reflection coefficient, and = ; 11) The Burg algorithm is the one that sums the forward and backward prediction error powers. Relative to the reflection coefficient Minimum, thus obtained estimate: Equation (11) 12) The result is estimated from equation (9). Then, at order s, the AR model coefficients are obtained using the Levinson-Durbin recursive algorithm: Equation (12) Equation (13) Equation (14) S42: Substitute the AR model parameters obtained in step S41 into equation (3) to obtain the power spectral density function. ; S43: Calculate the average power of the Mu rhythm (8-12Hz) of the EEG signal at rest for all subjects, and take the mean as the baseline value P. base Specifically: The resting-state power P is obtained by integrating the 8-12Hz EEG data using the power spectral density function. rest 1,P rest 2...P rest (M+N), and calculate their mean as the baseline value P. base : Equation (15) Where n = M + N, M represents the number of expert subjects, 0 ≤ M ≤ n, and N represents the number of novice subjects, 0 ≤ N ≤ n; S44: Calculate the average power of the EEG signals in the task state for all subjects in the 8-12Hz range. In the collaborative mode, obtain the total power P in the task state according to formula (16). end for Equation (16) in express The power percentage, express The percentage of power; , These represent the real-time task-state power output by experts and novices, respectively. S45: First compare attention metrics Compared with the benchmark value ,like < If the output is 0, then the output will be 0; if The total power P in the task state end Compared with baseline value P base By comparison, if the average power of the 8-12Hz EEG channel on the opposite side of the task is lower than the baseline value, it is considered that the event-related desynchronization (ERD) phenomenon has occurred, and the output result is 1; otherwise, the output result is 0. Step S5: Multimodal Neural Feedback S51: Provide real-time feedback based on the output of step S4: If the output of step S4 is 1, the gripper tightens downwards by an angle θ; if the output of step S4 is 0, the gripper angle remains unchanged. At the same time, based on the average power and attention index output by S4, update the energy bar and energy slot on the screen in real time to provide each subject with real-time visual feedback on the brain activity status of themselves and other collaborating subjects. S52: Calculate the comprehensive index of each subject within the time period T, and select the subject with the lowest comprehensive index to receive tactile stimulation. Overall index = (Average energy over time T) / (Attention index over time T) S53: Based on the task completion rate and the comprehensive index of each subject, the subjects are ranked in descending order. The subjects with the highest and lowest rankings are given auditory stimulation, and the user identity information is adaptively adjusted. The user identity information is either a novice or an expert.

6. The computer-readable storage medium according to claim 5, characterized in that... The EEG acquisition device uses a 64-channel Neuroscan with a sampling frequency of 1000Hz. The electrode caps adopt the international 10 / 20 system electrode placement method. The 64 electrodes are designated as FP1, FP2, AF3, AF4, F7, F5, F3, F1, Fz, F2, F4, F6, F8, FT7, FC5, FC3, FC1, FCz, FC2, FC4, FC6, FT8, T7, C5, C6, C7, C8, C9 ...

3. C1, Cz, C2, C4, C6, T8, TP7, CP5, CP3, CP1, CPz, CP2, CP4, CP6, TP8, P7, P5, P3, P1, Pz, P2, P4, P6, P8, PO7, PO3, POz, PO4, PO8, O1, CB1, Oz, CB2, O2, M1, M2; EKG electrodes are placed on the left hand artery to record ECG data, and the reference electrode is placed on the right earlobe.

7. The computer-readable storage medium according to claim 5, characterized in that... The network I / O model of the data processing module adopts the efficient event processing pattern Reactor, and uses I / O multiplexing + thread pool technology to process EEG data from different user hosts. The main thread only listens for events, while the worker threads are responsible for processing the data.

8. The computer-readable storage medium according to claim 5, characterized in that... The preprocessing includes downsampling, notch filtering, bandpass filtering, and spatial filtering; (1) Downsampling: The original signal is downsampled at 250Hz during the data preprocessing process; (2) Notch filter: Use a 50Hz notch filter to filter out high-frequency noise, baseline drift, electromagnetic interference, etc. (3) Bandpass filter: Use a 0.5-45Hz bandpass filter to acquire data in the frequency bands related to motor imagery and attention; (4) Spatial filter: First, calculate the average value of the EEG data of the electrodes around C3 and C4, and then subtract the average value from the data of each channel.

Images