A brain-computer interface algorithm system based on lower limb motor imagery

Abstract

The application discloses a brain-computer interface algorithm system based on lower limb motor imagination, relates to the technical field of brain-computer interface, and has the technical points that the application adopts a high-precision lower limb motor imagination brain-computer interface system based on adaptive time-frequency-Riemann geometry fusion, realizes robust decoding of the motor intention of a knee joint pain patient through optimal frequency band selection and covariance manifold classification, and provides reliable technical support for clinical rehabilitation.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of brain-computer interface technology, and particularly relates to a brain-computer interface algorithm system based on lower limb motor imagery. BACKGROUND

[0002] Chronic knee osteoarthritis pain is one of the most common diseases worldwide, significantly impairing the quality of life of patients and reducing work efficiency, however, its impact on cortical plasticity and motor imagery (MI) is still poorly understood. Although motor imagery-based brain-computer interface (MI-BCI) has made significant progress in upper limb motor function recovery, and it has been proven that MI can induce motor cortex neural activity and achieve high precision in electroencephalogram (EEG) signal decoding, its application in lower limb diseases (especially in patients with pain) has not been fully explored.

[0003] To this end, the present application aims to provide a brain-computer interface algorithm system based on lower limb motor imagery to solve the above problems. SUMMARY

[0004] The purpose of the present application is to solve the above problems, and to provide a brain-computer interface algorithm system based on lower limb motor imagery, which significantly improves the recognition accuracy of motor intention through an innovative adaptive time-frequency-Riemann geometry fusion algorithm, providing reliable technical support for clinical rehabilitation.

[0005] In order to achieve the above purpose, the technical scheme of the present application is as follows:

[0006] The present application provides a brain-computer interface algorithm system based on lower limb motor imagery, which comprises a data preprocessing module, a time-frequency feature extraction module, a feature dimension reduction module, a geometric filtering and classification module, a decision fusion module and a real-time feedback module.

[0007] The data preprocessing module filters the original signal collected from the electroencephalogram device, removes high-frequency noise and artifacts such as eye movement artifacts and power interference, and outputs cleaned EEG data signals for further analysis by applying a 0.5 to 80 Hz band-pass filter and a 50 Hz notch filter. The data preprocessing module obtains clean signal x_f(t) by filtering the original electroencephalogram signal x(t) with a 50 Hz notch filter, and divides x_f(t) into individual trials;

[0008] The time-frequency feature extraction module extracts features in the EEG signal through time-frequency analysis methods. Specifically, continuous wavelet transform and Welch method are used to analyze the EEG signal in different time windows in the frequency domain, and the energy features of each frequency band (such as alpha wave, beta wave) are obtained. These time-frequency features can capture the neural activity related to lower limb motor imagery, providing a basis for subsequent feature classification;

[0009] The time-frequency feature extraction module adopts Welch method and continuous wavelet transform to perform time-frequency analysis on the preprocessed electroencephalogram signals; Welch method calculates power spectral density through periodogram:

[0010]

[0011] where w(n) is a Hamming window, L is a window length, and U is a normalization constant; CWT is defined as:

[0012]

[0013] The analysis frequency band covers [8, 12Hz], [8, 20Hz], [8, 30Hz], [12, 20Hz], [15, 20Hz], [15, 30Hz], [20, 30Hz], and [8, 15Hz] eight intervals, and the optimal classification frequency band is determined through adaptive frequency band selection;

[0014] The feature dimension reduction module uses the local tangent space alignment method to perform feature dimension reduction on the EEG signals, and the high-dimensional features of the EEG signals are compressed to a lower dimension, thereby improving the calculation efficiency, avoiding overfitting of high-dimensional features, and reducing the risk of overfitting;

[0015] The geometric filtering and classification module uses the minimum distance classification (MDRM) method based on Riemann geometry to classify the reduced features. By calculating the covariance matrix of different time-frequency windows and using Riemann distance to measure the similarity between signals, the classifier can efficiently determine the type of motor imagery (left leg or right leg); the geometric filtering and classification module can robustly classify EEG signals from different individuals, handle cross-individual and cross-period differences, and the Riemann distance between the two covariance matrices C_i and C_j is:

[0016] δR(Ci,Cj)=|log(Ci -1 / 2 CjCi -1 / 2 )| F

[0017] where |·| F is the Frobenius norm, and the Riemann mean C is obtained by minimizing the squared Riemann distance sum;

[0018] The decision fusion module fuses the classification results of the geometric filtering and classification module with other classification methods (such as time-space linear discriminant analysis TSLDA); decision fusion improves the classification accuracy by weighted combination of the outputs of multiple classifiers, and can adapt to different electroencephalogram patterns of individuals; the decision fusion module integrates the classification results of the time-space linear discriminant analysis and the discriminant geodesic filter MDRM, and determines the final class through weighted fusion:

[0019]

[0020] Where P_1(c) and P_2(c) are the probability outputs of the two classifiers, and w_1 and w_2 are the corresponding weights;

[0021] The real-time feedback module provides visual or auditory feedback to the user based on the classification results, indicating the execution status of motor imagery; the user can confirm whether the lower limb motor imagery is successful through indicator lights or graphics; the real-time feedback not only helps the user adjust the imagery task, but also promotes the neuroplasticity of the brain through the feedback mechanism, which helps to improve the rehabilitation effect.

[0022] The data preprocessing module includes bandpass filtering (0.5-80Hz) and a bad trial rejection mechanism. When the signal amplitude exceeds ±75μV in the 8-30Hz frequency band, it is automatically marked as a bad trial and excluded from analysis.

[0023] The time-frequency feature extraction module uses a sliding window strategy to dynamically track changes in ERD / ERS patterns. The window length is 1 second and the overlap rate is 50% to capture dynamic neural activity changes during motor imagery.

[0024] When the feature dimensionality reduction module uses the local tangent space permutation algorithm, it retains the first 20 principal components, which can explain more than 90% of the data variance, significantly reducing computational complexity while ensuring classification performance.

[0025] When the geometric filtering and classification module uses the Riemann mean classifier, a specific covariance matrix regularization parameter λ = 0.1 is set for the EEG characteristics of patients with knee pain to improve classification stability.

[0026] In the decision fusion module, the weight w1 of the spatiotemporal linear discriminant analysis is set to 0.6, and the weight w2 of the discriminant geodesic filter (MDRM) is set to 0.4. This weight ratio is determined by cross-validation to obtain the optimal classification performance.

[0027] The system supports cross-session and cross-subject transfer learning. When applying the pre-trained model to new patients using domain adaptation technology, it can achieve a classification accuracy of over 80% with only a small amount of calibration data.

[0028] In this approach, Welch periodograms and continuous wavelet transform (CWT) are used to analyze eight sub-bands (α / β and their combinations) in parallel, and the most discriminative [8-30Hz] joint band is dynamically selected, which can solve the feature extraction problem caused by ERD / ERS pattern variations in pain patients.

[0029] By mapping the EEG covariance matrix to Riemann space and using the minimum Riemann distance classifier (MDRM) to process non-stationary signals, the robustness to electrode offset and individual differences is improved by 37.2% compared to the traditional Euclidean space method.

[0030] By fusing the results of spatiotemporal linear discriminant analysis (TSLDA) and Riemannian geometric classification, and balancing the complementary information of spatiotemporal features and manifold structure through weighted voting (weight ratio 0.6:0.4), the average cross-subject accuracy reached 86.41%.

[0031] A 4-second motor imagery duration and visual-auditory bimodal cues were designed for pain patients, along with a real-time virtual feedback interface, which significantly reduced the invalid test rate (<8%) due to pain-induced distraction.

[0032] Compared with existing technologies, the beneficial effects of this solution are:

[0033] This invention validates the EEG dataset by analyzing the proportion of adverse trials, event-related desynchronization / synchronization (ERD / ERS) in motor imagery, and classification performance. After excluding adverse trials, significant contralateral ERD and ipsilateral ERS were observed in the sensorimotor areas (C3 and C4) during the MI task in pain patients, indicating that lower limb MI can significantly enhance the neuroplasticity of the motor areas in pain patients. This lays a solid foundation for future clinical applications, especially for individuals with limited lower limb motor function. Attached Figure Description

[0034] Figure 1 This is a schematic diagram comparing the classification accuracy of four models in 30 subjects in an embodiment of the present invention;

[0035] Figure 2 This is a schematic diagram of the confusion matrix of the four classification models in the embodiments of the present invention. Detailed Implementation

[0036] To enable those skilled in the art to better understand the present invention, the technical solution of the present invention will be described in further detail below with reference to the embodiments and accompanying drawings. Obviously, the described embodiments are merely some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort should fall within the scope of protection of the present invention.

[0037] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other. The present invention will now be described in detail with reference to the embodiments.

[0038] Example 1:

[0039] Classification performance verification:

[0040] To rigorously validate the newly collected dataset of lower limb motor imagery from patients with knee pain, this study employed four classification methods—CSP+LDA, FBCSP+SVM, EEGNet, and the newly developed OTFWRGD algorithm—to analyze EEG signal features from different dimensions, demonstrating the technological evolution from traditional feature engineering methods to advanced machine learning strategies.

[0041] 1. Traditional Spatial Filtering Method (CSP+LDA)

[0042] Common Spatial Pattern (CSP) is a widely used spatial filtering technique in motor imagery brain-computer interface research. It maximizes the variance difference between two types of signals through linear transformation (in this study, it applies to left and right lower limb motor imagery). Combined with Linear Discriminant Analysis (LDA), CSP+LDA has become a benchmark method for offline and online brain-computer interface research due to its simplicity and strong interpretability. This study uses this method to establish a performance baseline, providing a reference for subsequent comparisons of advanced algorithms.

[0043] 2. Extended Filter Bank Method (FBCSP+SVM)

[0044] Filter Bank Common Spatial Patterns (FBCSP) improves upon traditional methods by decomposing EEG signals into multiple frequency bands (α, β, or narrower sub-bands) and applying Common Spatial Patterns (CSP) to each. This technique extracts frequency domain features specific to motor imagery tasks, which are then classified using Support Vector Machines (SVMs). SVMs are robust in brain-computer interface applications, and combined with frequency band-specific spatial filtering, FBCSP+SVM offers unique advantages in evaluating the discriminative power of cross-frequency band features.

[0045] 3. Deep Learning Architecture (EEGNet)

[0046] EEGNet is a compact deep learning architecture designed specifically for EEG signals. It learns spatiotemporal features directly from raw signals through temporal convolution, spatial convolution, and depthwise separable convolutional layers. This study introduces EEGNet to evaluate the ability of an end-to-end learning method to process lower limb motor imagery data from patients with knee osteoarthritis (whose cortical activity patterns may differ from healthy individuals). This model, while maintaining a relatively small number of trainable parameters, can adapt to diverse EEG tasks, making it an ideal tool for validating dataset quality.

[0047] 4. Optimal Time-Frequency Window Riemannian Geometric Classification Algorithm (OTFWRGD)

[0048] The Optimal Time-Frequency Window Riemann Geometric Distance Classification Algorithm (OTFWRGD) innovatively integrates time-frequency feature optimization, covariance feature extraction, and Riemann geometric classification methods, improving the classification performance of motor imagery EEG signals through multi-level processing.

[0049] Data preprocessing stage:

[0050] The raw EEG signal x(t) was filtered by a 50Hz notch filter to obtain a clean signal x_f(t), which was then divided into independent trials.

[0051] Time-frequency feature extraction:

[0052] Time-frequency analysis was performed using the Welch method and continuous wavelet transform (CWT). The Welch method calculates the power spectral density using a periodogram.

[0053]

[0054] Where w(n) is the Hamming window, L is the window length, and U is the normalization constant. CWT is defined as:

[0055]

[0056] The analysis frequency band covers eight intervals: [8,12Hz], [8,20Hz], [8,30Hz], [12,20Hz], [15,20Hz], [15,30Hz], [20,30Hz], and [8,15Hz]. The optimal classification frequency band is determined through adaptive frequency band selection.

[0057] Feature dimensionality reduction:

[0058] The Local Tangent Space Permutation (LTSA) algorithm is used to reduce the dimensionality of the covariance matrix, which reduces computational complexity while preserving local geometric relationships and avoids overfitting of high-dimensional features.

[0059] Geometric filtering and classification:

[0060] After dimensionality reduction, the features are processed by discriminative geometric filtering and classified using the Riemann geometric method. The Riemann distance between the two covariance matrices C_i and C_j is:

[0061] δR(Ci,Cj)=|log(Ci -1 / 2 CjCi -1 / 2 )| F

[0062] Among them |·| F Let C be the Frobenius norm. The Riemann mean C is obtained by minimizing the sum of squared Riemann distances.

[0063] Decision integration:

[0064] The classification results of Spatiotemporal Linear Discriminant Analysis (TSLDA) and Discriminant Geodesic Filtering Method (DGFMDRM) are integrated, and the final category is determined by weighted fusion:

[0065]

[0066] Where P_1(c) and P_2(c) are the probability outputs of the two classes, and w_1 and w_2 are the corresponding weights.

[0067] The OTFWRGD algorithm has the following innovative advantages:

[0068] 1. The time-frequency window adaptive optimization mechanism enhances the ability to capture neural oscillation features related to motor imagery.

[0069] 2. Covariance features are mapped to Riemannian manifolds to enhance robustness against artifacts and individual differences.

[0070] 3. Amplitude scaling invariance improves model generalization performance.

[0071] 4. LTSA dimensionality reduction improves computational efficiency while preserving the intrinsic geometry of EEG.

[0072] 5. Riemann distance measurement is better suited to the characteristics of non-stationary EEG signals than traditional Euclidean methods.

[0073] 6. Decision fusion strategies enhance classification reliability and individualized adaptability.

[0074] This algorithm provides an innovative solution for motor imagery EEG classification by organically combining time-frequency feature optimization, covariance feature extraction, and Riemannian geometric classification, significantly improving the classification accuracy, computational efficiency, and system robustness of brain-computer interface applications.

[0075] Classification performance results:

[0076] To verify the classification performance of the OTFWRGD algorithm, this study conducted a rigorous training and testing process using EEG data from each subject, and employed 10-fold cross-validation to ensure the robustness of the classification results. The detailed steps of this process are as follows:

[0077] (1) Dataset partitioning

[0078] First, all EEG data were divided into training and testing sets. Each subject's EEG dataset was split in a 7:3 ratio, meaning 70% of the data was used for training and 30% for testing. This data partitioning ensured fairness between training and testing, while also allowing the test set to fully reflect the model's performance on unseen data.

[0079] (2) Cross-validation

[0080] To further evaluate the robustness of the OTFWRGD algorithm, 10-fold cross-validation was employed. Specifically, the EEG datasets of all subjects were randomly divided into 10 subsets, with 9 subsets used for training and the remaining subset used for validation. This process was repeated 10 times, selecting a different subset as the validation set each time. Finally, the average accuracy of the cross-validation was calculated to avoid evaluation bias caused by the randomness of data partitioning.

[0081] (3) Model Training

[0082] In each fold cross-validation, the OTFWRGD algorithm is trained using the training set. The training process includes the following steps:

[0083] Data preprocessing: The EEG signals in the training set are subjected to 50Hz notch filtering and bandpass filtering (0.5Hz to 80Hz) to remove artifacts and noise.

[0084] Time-frequency feature extraction: Welch method and continuous wavelet transform (CWT) are used to perform time-frequency analysis on EEG signals to extract time-frequency domain features.

[0085] Feature dimensionality reduction: The Local Tangent Space Permutation (LTSA) algorithm is applied to reduce the dimensionality of time-frequency features, thereby reducing the dimension of the feature space and avoiding overfitting.

[0086] Geometric classification: The Riemannian distance between the dimensionality-reduced covariance matrices is calculated using the Riemannian geometric method to classify the data.

[0087] Decision fusion: The final classification output is determined by weighted fusion of the results of multiple classifiers (such as spatiotemporal linear discriminant analysis (TSLDA) and discriminant geodesic filtering (MDRM).

[0088] (4) Testing and Evaluation

[0089] After each training iteration, the model is evaluated using a test set. During evaluation, the system uses the trained model to classify the test set and calculates the following performance metrics:

[0090] Average accuracy: The percentage of times a model correctly classifies a test set.

[0091] Kappa coefficient: assesses classification consistency and the reliability of the classifier, eliminating the influence of accidental consistency.

[0092] Accuracy and Sensitivity: Accuracy reflects the proportion of positive examples correctly predicted by the classifier, while sensitivity reflects the classifier's ability to correctly identify negative examples.

[0093] (5) Classification results

[0094] In this study, the EEG datasets of each subject were divided into training and testing sets in a 7:3 ratio. To further validate model performance, 10-fold cross-validation was used to ensure robust estimation of performance metrics for all subjects. Table 2 shows the mean results of four models (CSP+LDA, FBCSP+SVM, EEGNet, and OTFWRGD) in terms of average accuracy, classification consistency index Kappa coefficient, and decoding method robustness indices precision and sensitivity. As shown in Table 1, the mean accuracies of the four models were 51.43%, 55.71%, 76.21%, and 86.41%, respectively.

[0095] Table 1 Comparison of average classification performance of different models in 30 subjects

[0096]

[0097] Figure 1 The individual classification accuracy of four models was presented in 30 subjects (a total of 15,000 trials). The results showed that the OTFWRGD method achieved significantly higher accuracy than the other three methods for most subjects. EEGNet performed comparable to or slightly better than traditional methods (CSP+LDA and FBCSP+SVM) in most cases, confirming its adaptive feature learning ability. However, the accuracy of CSP+LDA and FBCSP+SVM showed greater variability among different subjects, suggesting that these methods may be more sensitive to individual differences in knee pain severity or imagination.

[0098] To further evaluate classification performance, Figure 2 The confusion matrices of four models are presented, and the discrimination performance of each classifier on the left and right lower limb motor imagery task is statistically analyzed. The four classification methods exhibit differentiated performance characteristics, confirming the applicability of this lower limb motor imagery EEG dataset to the pain patient population. OTFWRGD demonstrated the highest classification accuracy and stability across all subjects, highlighting its effectiveness in addressing the unique challenges of lower limb pain. EEGNet also showed good performance, demonstrating the potential of deep learning to automatically extract effective features from clinical EEG data. The results of CSP+LDA and FBCSP+SVM indicate that even with traditional machine learning methods, this dataset can still provide reliable motor imagery signals. Overall, these findings confirm the discriminability of neural activity in left and right lower limb motor imagery, providing strong support for the practical value of this dataset in future rehabilitation engineering and brain-computer interface applications.

[0099] Example 2:

[0100] 1. Dataset Overview

[0101] The research team collected a large-scale lower limb motor imagery (MI) electroencephalogram (EEG) dataset, focusing on patients with lower limb knee pain.

[0102] The dataset contains 30 patients, each of whom completed 5 independent trials. Each trial included 100 left and right leg flexion and extension imagery tasks, for a total of 15,000 trials.

[0103] The data is organized according to the EEG-BIDS standard and includes raw data, preprocessed data, and experimental metadata, and is publicly available on the Figshare platform.

[0104] 2. Experimental Design

[0105] The experiment used a 32-channel semi-dry electrode EEG device (ZhenTec NT1) with a sampling rate of 500Hz.

[0106] Task design: Each trial includes 2 seconds of preparation, 4 seconds of motor imagery, and 4 seconds of rest. Visual cues guide patients to complete the motor imagery of the left or right leg.

[0107] Data preprocessing included 0.5-80Hz bandpass filtering and 50Hz notch filtering, and segmentation and labeling were performed using the EEGLAB toolbox.

[0108] 3. Data Analysis Methods

[0109] ERD / ERS analysis: Significant event-related desynchronization (ERD) and ipsilateral synchronization (ERS) were observed in the contralateral sensorimotor areas (C3 / C4 channels) during motor imagery, especially in the 8-30 Hz frequency band.

[0110] Classification Algorithm:

[0111] Traditional methods: CSP+LDA (51.43% accuracy), FBCSP+SVM (55.71% accuracy).

[0112] Deep learning method: EEGNet (76.21% accuracy).

[0113] The new algorithm, OTFWRGD (Optimal Time-Frequency Window in Riemannian Geometric Distance), achieves an average accuracy of 86.41%, significantly outperforming other methods.

[0114] The OTFWRGD algorithm combines time-frequency feature selection, covariance matrix dimensionality reduction (LTSA), and Riemannian geometric classification (MDRM) to improve performance through decision fusion.

[0115] 4. Key Findings

[0116] Motor imagery in patients with pain can still elicit detectable neural activity, but the intensity of ERD (-14.6%) is lower than that in healthy individuals (-20% to -30%).

[0117] The OTFWRGD algorithm performed well in classification tasks, verifying the feasibility of decoding lower limb motor imagery EEG signals.

[0118] The dataset provides an important foundation for the application of BCI in rehabilitation medicine, especially for patients with lower limb movement disorders.

[0119] 5. Limitations and Future Directions

[0120] The sample is currently limited to patients with KOA, and future studies need to expand it to other lower limb diseases (such as spinal cord injury).

[0121] The experimental task design is static; dynamic scenarios need to be added to improve its practicality.

[0122] The real-time performance and computational efficiency of the OTFWRGD algorithm need further optimization.

[0123] It is recommended to combine multimodal signals (such as EMG, fNIRS) to improve system robustness.

[0124] 6. Clinical significance

[0125] This study confirms that patients with pain retain their motor imagery abilities, supporting the application of BCI technology in rehabilitation.

[0126] High-precision classification algorithms provide technical support for the development of personalized rehabilitation strategies.

[0127] 7. Abbreviations and Terms

[0128] EEG: Electroencephalography; MI: Motor Imagery; BCI: Brain-Computer Interface; ERD / ERS: Event-Related Desynchronization / Synchronization; CSP: Common Spatial Pattern; LDA: Linear Discriminant Analysis; MDRM: Riemann Mean Minimum Distance.

[0129] 8. Data availability

[0130] The dataset and code are publicly available on GigaDB, following the EEG-BIDS standard, to facilitate reuse in subsequent research.

[0131] The above specific embodiments are merely explanations of the present invention and are not intended to limit the present invention. After reading this specification, those skilled in the art can make modifications to these embodiments without contributing any inventive step, but as long as they are within the scope of the claims of the present invention, they are protected by patent law.

Claims

1. A lower limb motor imagery based brain-computer interface algorithm system, characterized in that: The algorithm system comprises a data preprocessing module, a time-frequency feature extraction module, a feature dimension reduction module, a geometric filtering and classification module, a decision fusion module and a real-time feedback module. The data preprocessing module filters the original signal collected from the electroencephalogram device, removes high-frequency noise and artifacts, applies a band-pass filter of 0.5 to 80 Hz and a 50 Hz notch filter, and outputs the cleaned EEG data signal for further analysis. The time-frequency feature extraction module performs frequency domain analysis on the EEG signal in different time windows through time-frequency analysis method to obtain the energy features of each frequency band. The time-frequency feature extraction module uses Welch method and continuous wavelet transform to perform time-frequency analysis on the preprocessed electroencephalogram signal. Where w(n) is the Hamming window, L is the window length, and U is the normalization constant; CWT is defined as: The analysis frequency band covers [8, 12Hz], [8, 20Hz], [8, 30Hz], [12, 20Hz], [15, 20Hz], [15, 30Hz], [20, 30Hz], [8, 15Hz] eight intervals, and the optimal classification frequency band is determined by adaptive frequency band selection. The feature dimension reduction module uses the local tangent space alignment method to reduce the dimension of the EEG signal, which compresses the high-dimensional features of the EEG signal to a lower dimension, thereby improving the calculation efficiency, avoiding overfitting of high-dimensional features, and reducing the risk of overfitting. The geometric filtering and classification module uses the minimum distance classification method based on Riemann geometry to classify the reduced features. By calculating the covariance matrix of different time-frequency windows and using Riemann distance to measure the similarity between signals, the classifier can efficiently determine the type of motor imagery; The geometric filtering and classification module can robustly classify EEG signals from different individuals, handle cross-individual and cross-period differences, and the Riemann distance between the two covariance matrices C_i and C_j is: δR(Ci,Cj) = |log(Ci -1 / 2 CjCi -1 / 2 )| F Among them |·| F The Frobenius norm and the Riemann mean C are obtained by minimizing the sum of squared Riemann distances; The decision fusion module fuses the classification results of the geometric filtering and classification module with the time-space linear discriminant analysis TSLDA method, and the decision fusion combines the outputs of multiple classifiers through weighted combination, improves the classification accuracy, and can adapt to different electroencephalogram patterns of different individuals; The decision fusion module integrates the classification results of the time-space linear discriminant analysis and the discriminant geodesic filter MDRM, and determines the final class through weighted fusion: Where P_1(c) and P_2(c) are the probability outputs of the two classifiers, and w_1 and w_2 are the corresponding weights; The real-time feedback module provides visual or auditory feedback to the user according to the classification result, indicating the execution of motor imagery; The user confirms whether the lower limb motor imagery is successful through the indicator light or the figure; The real-time feedback not only helps the user to adjust the imagination task.

2. The motor imagery-based brain-computer interface algorithm system according to claim 1, characterized in that: The data preprocessing module can perform band-pass filtering and bad trial rejection mechanism. When the signal amplitude exceeds ±75 μV in the 8-30 Hz frequency band, it is automatically marked as a bad trial and excluded from analysis.

3. The lower limb motor imagery based brain-computer interface algorithm system according to claim 1, characterized in that: The time-frequency feature extraction module adopts a sliding window strategy to dynamically track the ERD / ERS mode changes. The window length is 1 second, and the overlap rate is 50%, to capture the dynamic changes of neural activity during motor imagery.

4. The lower limb motor imagery based brain-computer interface algorithm system according to claim 1, characterized in that: The feature dimension reduction module adopts the local tangent space alignment algorithm, retains the first 20 principal components in the local tangent space alignment algorithm, ensures the classification performance, and reduces the computational complexity.

5. The lower limb motor imagery based brain-computer interface algorithm system according to claim 1, characterized in that: When the geometric filtering and classification module adopts the Riemann mean classifier, for the EEG features of patients with knee pain, a specific covariance matrix regularization parameter λ=0.1 is set to improve the classification stability.

6. The lower limb motor imagery based brain-computer interface algorithm system according to claim 1, characterized in that: In the decision fusion module, the weight w1 of the spatiotemporal linear discriminant analysis is set to 0.6, and the weight w2 of the discriminant geodesic filter MDRM is set to 0.

4. The weight ratio is determined by cross-validation to obtain the optimal classification performance.

7. The lower limb motor imagery based brain-computer interface algorithm system according to claim 1, characterized in that: The system can transfer learning across sessions and across subjects.

Images