Stroke circle gait correction system based on BCI intention decoding

Abstract

The application relates to a stroke circle gait correction system based on BCI intention decoding, which comprises an online consultation platform; a patient wears an electroencephalogram cap, VR glasses, and tDCS and rTMS electrodes, and wears a power-assisted robot; the patient performs flat ground walking, step climbing and obstacle crossing training based on a VR training scene; an intention intensity signal is decoded by using a BCI intention recognition module; the power-assisted robot converts the intention intensity signal into a servo control instruction for driving; a knee joint assists hip and knee flexion; an adjustable damping abduction limiting joint limits leg abduction and circle drawing of the patient; and combined with multi-scene VR feedback and time-sequenced tDCS and rTMS nerve regulation, the stroke circle gait correction is realized from the aspects of "idea-action-feedback-remodeling".

Description

Technical Field

[0001] This invention belongs to the field of medical device technology, and in particular relates to a gait correction system for stroke patients based on BCI intent decoding. Background Technology

[0002] After a stroke, hemiplegic patients often exhibit a circling gait when walking. This gait is essentially due to four factors: First, impaired transmission of motor intentions, with damage to the motor cortex on the affected side, preventing the effective transmission of nerve signals indicating the intention to flex the knee / hip to the lower limb muscles. Second, muscle weakness, particularly in the hip flexors (iliopsoas) and knee flexors (biceps femoris), prevents the completion of normal leg lifting movements, leading to compensatory outward circling using the hip joint as a fulcrum. Third, sensory and feedback deficiencies, including impaired proprioception and visual feedback, making it difficult for the brain to perceive and self-correct movement deviations. Fourth, insufficient neuroplasticity, rendering traditional passive rehabilitation ineffective in activating active motor intentions, and causing chronic-phase patients to enter a plateau in recovery.

[0003] The existing stroke correction systems have the following problems: traditional gait correctors only provide passive kinetic assistance without connecting to the patient's active movement intentions, which easily leads to "patients passively following without brain participation" and extremely low efficiency of neural remodeling; the use of a single BCI or neuromodulation technology lacks temporal synergy with mechanical assistance and contextual feedback, making it impossible to form a closed-loop intervention; the application of neuromodulation is chaotic, and tDCS and rTMS do not combine with the temporal logic of repeated facilitation therapy (RIP), resulting in unstable stimulation effects and poor correction of the patient's circling gait.

[0004] Therefore, designing a stroke gait correction system based on BCI intention decoding that can more accurately identify patient intentions and assist in precise walking training has become an urgent technical problem to be solved. Summary of the Invention

[0005] To address the aforementioned technical issues, a stroke gait correction system based on BCI intent decoding is provided, comprising the following modules and a central control unit.

[0006] The BCI intention recognition module is equipped with a non-invasive 64-channel EEG cap that covers the bilateral motor cortex M1 area and the supplementary motor area SMA. It is used to capture and decode the patient's active motor intention of knee / hip flexion, and to identify and output the intensity of the patient's BCI intention in gait.

[0007] The lower limb robot assist module is equipped with a split dual-segment assist robot, including an adjustable damped abduction limiting joint set in the thigh segment, which dampens and blocks hip abduction >15°, and a motor-driven knee flexion joint set in the lower leg segment, which adjusts the output assist of the knee flexion joint according to the intensity of BCI intent, and is equipped with sensors to detect kinematic data including hip and knee flexion angle, stride length, stride frequency, and abduction angle.

[0008] The multi-scene VR feedback module is equipped with VR glasses and a differential drive chassis. The VR glasses show patients VR training scenarios including walking on flat ground, climbing stairs, and crossing obstacles. The differential drive chassis is equipped with a center of gravity pressure sensor, which follows the patient's walking movements based on the sensor signal feedback.

[0009] The sequential neuromodulation module is equipped with tDCS and rTMS electrodes. With tDCS as the core, the stimulation target points are the M1 area of ​​the primary motor cortex on the affected side and the lower limb representative area. It is combined with rTMS assistance and repetitive facilitation therapy for assistance. The stimulation target points of rTMS are the M1 area on the healthy side and the SMA area on the affected side. The repetitive facilitation therapy includes providing tactile stimulation to the skin after assisting in the completion of the movement and giving verbal prompts to the patient.

[0010] The multimodal data acquisition and evaluation module, based on EEG cap detection and feedback from the assistive robot, collects kinematic and EEG data, which are then used for manual input of FMA scores and 10-meter walking speed data to evaluate training effectiveness and iterate the program.

[0011] The central control unit (CCU) connects all modules for unified scheduling, enabling time-series coordination of intent detection, robot assistance, data feedback, and evaluation and control.

[0012] As a further improvement to this method, the motion intention analysis process includes,

[0013] s1, EEG is collected by covering the M1 and SMA areas with an EEG cap. The core channels of the EEG cap include Cz, C3, C4, FCz, CPz, CP3, and CP4. The reference electrodes are bilateral mastoid processes (A1 / A2) or the average reference scheme. The sampling rate is 1000Hz, and DC coupling is used for acquisition.

[0014] s2, preprocesses the collected data, including filtering, online artifact removal, and baseline calibration;

[0015] s3, perform individual calibration, collect EEG data from the patient during multiple real gait movements, extract features, determine individual thresholds based on the features, and record them in the patient configuration file for storage;

[0016] s4. Perform ERD+MRCP dual-modal feature fusion. Extract feature vectors from preprocessed EEG data, input the feature vectors into the classifier, output three-dimensional intent probability, use the three-dimensional intent probability to calculate the continuous value of intent intensity IntentScore, and substitute it into the hierarchical decision to obtain the intent level.

[0017] S5 outputs the intention level to the assistive robot through the central control unit, and has a decision verification consisting of a confirmation window mechanism and false positive suppression. The assistive robot is linearly mapped to the assistance percentage based on the intention level.

[0018] As a further improvement to this method, the filtering in step s2 includes band-stop filtering using a second-order IIR filter and band-pass filtering using parallel dual-branch processing; the band-stop filtering uses a 50Hz power frequency notch filter with a bandwidth of 48-52Hz; the band-pass filtering is divided into an ERD branch and an MRCP branch, the ERD branch uses 8–30Hz band-pass filtering to obtain μ-band 8–13Hz and β-band 14–30Hz, and the MRCP branch uses 0.05–3Hz zero-phase low-pass filtering; the online artifact removal is based on an offline training model of independent component analysis (ICA), online projection removes EOG and EMG artifacts, and sets an adaptive threshold of ±100μV to remove large transient artifacts; the baseline calibration uses the average potential of the resting window 500ms before the triggering of motor imagery as the baseline, and subtracts the baseline in real time for offset.

[0019] As a further improvement to this method, step s3 involves having the patient perform 10-15 attempts at knee / hip flexion on the affected side and recording EEG. Feature extraction includes calculating the ERD percentage and MRCP negative peak amplitude for each action and combining them to obtain the IntentScore for each action. Individual thresholds are set by sorting the fusion scores from low to high, with the 30th percentile used as the "weak / moderate" cutoff value and the 70th percentile used as the "moderate / strong" cutoff value. The patient profile includes individual thresholds, ICA projection matrix, and filtering parameters.

[0020] As a further improvement to this method, the formula for calculating the ERD percentage is as follows: ,in, The energy of the μ+β band within the resting state window of -500 to 0 ms is obtained by calculating the total energy in μV². The total energy of the μ+β band within a 0–800ms window after the initiation of motor imagery is obtained by calculating the total energy of the μ+β band. The MRCP negative peak amplitude is obtained by setting an extraction window from 1000ms before the motor imagery trigger point to 200ms after the trigger point, acquiring the EEG signals of the three channels Cz, FCz, and CPz within the window, and calculating the average value to obtain the MRCP negative peak amplitude, in μV.

[0021] As a further improvement of this method, the feature vector in step s4 includes 1-dimensional ERD percentage, 2-dimensional relative power change in the μ band, 2-dimensional relative power change in the β band, 1-dimensional negative peak amplitude of MRCP, 1-dimensional latency of the negative peak of MRCP, 1-dimensional contralateral / ipsilateral ERD ratio, 1-dimensional variance of energy fluctuation within the time window, and 1-dimensional DC drift rate after baseline correction; the 2-dimensional relative power change in the μ band includes the average values of EEG signals from Cz and C3 channels; the 2-dimensional relative power change in the β band includes the average values of EEG signals from Cz and C4 channels; the latency of the negative peak of MRCP is detected by using a sliding window with a length of 500 ms and a step size of 50 ms to detect the negative trend within the extraction window of the negative peak amplitude of MRCP. When the amplitudes in three consecutive windows decrease and the minimum value is less than -2 μV, it is determined as the onset of MRCP, and the latency of the negative peak of MRCP is determined; the contralateral / ipsilateral ERD ratio is the ratio of C3 and C4 channels; the variance of energy fluctuation within the time window is obtained based on calculating the variance of the instantaneous energy within the ERD branch window; the DC drift rate after baseline correction is obtained by performing linear fitting within the window after baseline calibration and taking the slope of the fitting line.

[0022] As a further improvement of this method, the classifier in step s4 processes the feature vector based on the LightGBM (Light Gradient Boosting Machine) model or the online adaptive linear discriminant analysis (LDA) module. The training strategy of the classifier is to use historical data with real movement labels of 50 - 100 stroke patients to pre-train the basic model offline. After each training, the actual joint angle changes feedback by the robot are used as weak labels to fine-tune the classifier in a semi-supervised manner, and it is updated every 5 iterations; the three-dimensional intention probabilities are from weak to strong as , , , and the calculation formula for the continuous value of intention intensity is , and the value range is 0 - 100; the determination conditions for the intention level in the hierarchical decision include

[0023] a), > 0.6 or IntentScore ≤ 33 is a weak intention, and high assistance is output, with the corresponding assistance percentage range being 80% - 100%;

[0024] b), > 0.5 and 33 < IntentScore ≤ 66 is a medium intention, and medium assistance is output, with the corresponding assistance percentage range being 40% - 70%;

[0025] c), > 0.5 or IntentScore > 66 is a strong intention, and low assistance is output, with the corresponding assistance percentage range being 10% - 30%.

[0026] As a further improvement to this method, the confirmation window mechanism in step s5 slides a window every 100ms from the moment the patient begins motor imagery. If the intent level classification results of two consecutive sliding windows are consistent, the output is a valid intent. If two consecutive windows are inconsistent, the previous valid result is output, and a 50ms delay is set for re-detection. The maximum tolerance delay is set to 300ms. False positive suppression is achieved by immediately canceling the current assistance and triggering a VR prompt when the sensor of the assistive robot detects a joint angle change that does not match the intent type. For every 3 false positives, the system automatically increases the filtering order of the MRCP branch and recalibrates the baseline. The default filtering order is 4 and increases incrementally.

[0027] As a further improvement to this method, the individual threshold boundary value in step s3 is equipped with an individual calibration mechanism. By collecting the IntentScore of 10-15 real exercise training sessions of the patient, sorting them from smallest to largest, the 30th percentile is taken as the "weak / medium" boundary LowThreshold, and the 70th percentile is taken as the "medium / strong" boundary HighThreshold. The calibration is performed once every 5 training sessions. If the highest IntentScore is <30, LowThreshold is set to 1 / 3 of the maximum value, and HighThreshold is set to 2 / 3 of the maximum value. If the lowest IntentScore is >70, LowThreshold is set to IntentScore (minimum score) - 10 and not less than 0, and HighThreshold is set to IntentScore (maximum score) + 10 and not higher than 100.

[0028] As a further improvement to this method, the training effect evaluation includes calculating the circle correction rate, intent recognition accuracy, and assist matching degree.

[0029] The circle correction rate is based on the total number of steps taken by the patient, counted by the sensors of the assistive robot. The standard is whether the abduction angle of the affected lower limb is <15° and the hip and knee flexion angle meets the training requirements. The effective normal gait steps are identified and the calculation formula is: Circle correction rate = (effective normal gait steps / total steps) × 100%;

[0030] The intent recognition accuracy and assist matching degree are based on the single-modal classification of the ERD branch and the MRCP branch. The single-modal classification of the ERD branch is based on the ERD percentage ranging from 5% to 85%, from low to high, corresponding to the weak to strong motion intent and the corresponding percentage assist output of the robot from strong to weak. The single-modal classification of the MRCP branch is based on the MRCP negative peak amplitude ranging from -1 to -15μV, from large to small, corresponding to the weak to strong motion intent and the corresponding percentage assist output of the robot from strong to weak. The results of the single-modal classification are combined and compared with the intent level and the assist percentage in step s5 to obtain the intent recognition accuracy and assist matching degree.

[0031] The iterative scheme includes,

[0032] The robot-assisted adjustment is based on the following criteria: a circle correction rate of ≥80%, an intent recognition accuracy of ≥85%, an assistance matching degree of ≥90%, and an improvement of ≥6 points in the lower limb FMA score compared to the baseline. When the results of three consecutive training sessions meet any two of the criteria, the percentage of robot assistance output is gradually reduced within the range of 0-5.

[0033] The intensity of tDCS and the number of RIP facilitation steps were adjusted based on the following criteria: a circle correction rate of <60%, no improvement after two consecutive training sessions, low electromyographic activation on the affected side, and the patient being in the chronic phase of stroke. Once any one of these criteria was met, the intensity of tDCS was gradually increased, starting at 2 mA, increasing by +0.5 mA, with a maximum of 2.5 mA. The pre-activation sequence was maintained at 200 ms, the stimulation duration was 5–10 minutes, and the number of RIP facilitation steps was increased from 3 per step to 4–5.

[0034] By adopting the above method, the split-type dual-segment assistive robot mechanically restricts abduction and guides normal hip and knee flexion, significantly reducing the range of circling during training. The BCI intention recognition module identifies and outputs the intensity of the patient's BCI intention in gait. With the patient's movement intention as the core, appropriate robot-assisted output is formed, avoiding passive training, promoting neural remodeling, and combining temporal neural modulation with multimodal feedback, thus prolonging the plasticity window and helping patients in the chronic phase to overcome the recovery plateau.

[0035] By creating patient profiles through individual calibration, the training effect is evaluated and the program is iterated after training. Based on the dynamic adjustment of multimodal data, precise rehabilitation with "one policy for one person" is achieved, and the effect can be quantified and traced. The multi-scenario VR feedback module provides VR training scenarios and simulated outdoor walking support at the same time, which solves the problems of boring and poor compliance in traditional rehabilitation. Attached Figure Description

[0036] Fig. 1 The diagram shown is a schematic of the operation process framework of the correction system of this patent.

[0037] Fig. 2 The diagram shown is a schematic of the assistive robot structure.

[0038] 1- Thigh segment, 2- Adjustable damped abduction limiting joint, 3- Lower leg segment, 4- Knee flexion joint. Detailed Implementation

[0039] To solve the above technical problems, such as Figs. 1-2 The system shown is a stroke gait correction system based on BCI intent decoding, including the following modules and a central control unit.

[0040] The BCI intention recognition module is equipped with a non-invasive 64-channel EEG cap that covers the bilateral motor cortex M1 area and the supplementary motor area SMA. It is used to capture and decode the patient's active motor intention of knee / hip flexion, and to identify and output the intensity of the patient's BCI intention in gait.

[0041] The lower limb robot assist module is equipped with a split dual-segment assist robot, including an adjustable damped abduction limiting joint set in the thigh segment, which dampens and blocks hip abduction >15°, and a motor-driven knee flexion joint set in the lower leg segment, which adjusts the output assist of the knee flexion joint according to the intensity of BCI intent, and is equipped with sensors to detect kinematic data including hip and knee flexion angle, stride length, stride frequency, and abduction angle.

[0042] The multi-scene VR feedback module is equipped with VR glasses and a differential drive chassis. The VR glasses show patients VR training scenarios including walking on flat ground, climbing stairs, and crossing obstacles. The differential drive chassis is equipped with a center of gravity pressure sensor, which follows the patient's walking movements based on the sensor signal feedback.

[0043] The sequential neuromodulation module is equipped with tDCS and rTMS electrodes. With tDCS as the core, the stimulation target points are the M1 area of ​​the primary motor cortex on the affected side and the lower limb representative area. It is combined with rTMS assistance and repetitive facilitation therapy for assistance. The stimulation target points of rTMS are the M1 area on the healthy side and the SMA area on the affected side. The repetitive facilitation therapy includes providing tactile stimulation to the skin after assisting in the completion of the movement and giving verbal prompts to the patient.

[0044] The multimodal data acquisition and evaluation module, based on EEG cap detection and feedback from the assistive robot, collects kinematic and EEG data, which are then used for manual input of FMA scores and 10-meter walking speed data to evaluate training effectiveness and iterate the program.

[0045] The central control unit (CCU) connects all modules for unified scheduling, enabling time-series coordination of intent detection, robot assistance, data feedback, and evaluation and control.

[0046] The process of analyzing movement intent includes,

[0047] s1, EEG is collected by covering the M1 and SMA areas with an EEG cap. The core channels of the EEG cap include Cz, C3, C4, FCz, CPz, CP3, and CP4. The reference electrodes are bilateral mastoid processes (A1 / A2) or the average reference scheme. The sampling rate is 1000Hz, and DC coupling is used for acquisition.

[0048] s2, preprocesses the collected data, including filtering, online artifact removal, and baseline calibration;

[0049] s3, perform individual calibration, collect EEG data from the patient during multiple real gait movements, extract features, determine individual thresholds based on the features, and record them in the patient configuration file for storage;

[0050] s4. Perform ERD+MRCP dual-modal feature fusion. Extract feature vectors from preprocessed EEG data, input the feature vectors into the classifier, output three-dimensional intent probability, use the three-dimensional intent probability to calculate the continuous value of intent intensity IntentScore, and substitute it into the hierarchical decision to obtain the intent level.

[0051] S5 outputs the intention level to the assistive robot through the central control unit, and has a decision verification consisting of a confirmation window mechanism and false positive suppression. The assistive robot is linearly mapped to the assistance percentage based on the intention level.

[0052] The filtering in step s2 includes band-stop filtering using a second-order IIR filter and band-pass filtering using parallel dual-branch processing. The band-stop filtering uses a 50Hz power frequency notch filter with a bandwidth of 48-52Hz. The band-pass filtering is divided into an ERD branch and an MRCP branch. The ERD branch performs band-pass filtering at 8–30Hz to obtain μ-band 8–13Hz and β-band 14–30Hz. The MRCP branch uses a 0.05–3Hz zero-phase low-pass filter. The online artifact removal is based on an offline training model of independent component analysis (ICA). Online projection removes EOG and EMG artifacts and sets an adaptive threshold of ±100μV to remove large transient artifacts. The baseline calibration uses the average potential of the resting window 500ms before the trigger of motor imagery as the baseline and subtracts the baseline in real time for offset.

[0053] Step s3 involves having the patient perform 10-15 attempts at knee / hip flexion on the affected side and recording EEG. Feature extraction includes calculating the ERD percentage and MRCP negative peak amplitude for each movement and combining them to obtain the IntentScore for each movement. Individual thresholds are set by sorting the fusion scores from low to high, with the 30th percentile used as the "weak / moderate" cutoff value and the 70th percentile used as the "moderate / strong" cutoff value. The patient profile includes individual thresholds, ICA projection matrix, and filtering parameters.

[0054] The formula for calculating the ERD percentage is as follows: ,in, The energy of the μ+β band within the resting state window of -500 to 0 ms is obtained by calculating the total energy in μV². The total energy of the μ+β band within a 0–800ms window after the initiation of motor imagery is obtained by calculating the total energy of the μ+β band. The MRCP negative peak amplitude is obtained by setting an extraction window from 1000ms before the motor imagery trigger point to 200ms after the trigger point, acquiring the EEG signals of the three channels Cz, FCz, and CPz within the window, and calculating the average value to obtain the MRCP negative peak amplitude, in μV.

[0055] The feature vector in step s4 includes a 1D ERD percentage, a 2D μ-band relative power change, a 2D β-band relative power change, a 1D MRCP negative peak amplitude, a 1D MRCP negative peak latency, a 1D contralateral / ipsilateral ERD ratio, a 1D energy fluctuation variance within a time window, and a 1D baseline-corrected DC drift rate. The 2D μ-band relative power change includes the average values ​​of the Cz and C3 channels of EEG signal. The 2D β-band relative power change includes the average values ​​of the Cz and C4 channels of EEG signal. The MRCP negative peak latency is determined by using a sliding window with a length of 500ms and a step size of 50ms, detecting a negative trend within the MRCP negative peak amplitude extraction window. When the amplitude decreases for three consecutive windows and the minimum value is less than -2μV, it is determined to be MRCP. The onset function determines the MRCP negative peak latency; the opposite / same-side ERD ratio is the ratio of channels C3 and C4; the energy fluctuation variance within the time window is obtained based on the variance of instantaneous energy within the ERD branch window; the baseline-corrected DC drift rate is obtained by linear fitting within the window after baseline calibration, and the slope of the fitted straight line is taken.

[0056] The classifier in step s4 is based on a Lightweight Gradient Boosting Machine (LightGBM) model or an Online Adaptive Linear Discriminant Analysis (LDA) module to process feature vectors. The training strategy involves using historical data from 50-100 stroke patients with real motion labels to pre-train the base model offline. After each training iteration, the actual joint angle changes fed back by the robot are used as weak labels to fine-tune the classifier in a semi-supervised manner, updating it every 5 iterations. The three-dimensional intent probabilities, from weakest to strongest, are... , , The formula for calculating the continuous value of intent intensity is: The value range is 0-100; the determination criteria for the intention level in the hierarchical decision-making process include,

[0057] a) >0.6 or IntentScore ≤ 33 represents a weak intention, and high assistance is output, with the corresponding assistance percentage range being 80% - 100%;

[0058] b), >0.5 and 33 < IntentScore ≤ 66 represents a medium intention, and medium assistance is output, with the corresponding assistance percentage range being 40% - 70%;

[0059] c), >0.5 or IntentScore > 66 represents a strong intention, and low assistance is output, with the corresponding assistance percentage range being 10% - 30%.

[0060] The confirmation window mechanism in step s5 is a sliding window that slides once every 100 ms from the moment the patient starts motor imagery. If the intention level classification results of two consecutive sliding windows are the same, it outputs a valid intention. If they are inconsistent for two consecutive times, it outputs the previous valid result and delays 50 ms for re - detection, with a maximum tolerance delay of 300 ms. When false positives are suppressed and the joint angle change detected by the sensor of the assistance robot does not match the intention type, the current assistance is immediately revoked and a VR prompt is triggered. Every time 3 false positives occur, the system automatically increases the filtering order of the MRCP branch and re - performs baseline calibration. The default filtering order is 4 and it increases gradually by one order each time.

[0061] The cut - off value of the individual threshold in step s3 has an individual calibration mechanism. By collecting the IntentScore of the patient's 10 - 15 real - motion training sessions, sorting them from smallest to largest, the 30th percentile is taken as the LowThreshold for the "weak / medium" boundary, and the 70th percentile is taken as the HighThreshold for the "medium / strong" boundary. It is recalibrated every 5 training sessions. If the highest IntentScore < 30, set LowThreshold to 1 / 3 of the maximum value and HighThreshold to 2 / 3 of the maximum value. If the lowest IntentScore > 70, set LowThreshold = IntentScore (minimum score) - 10 and not less than 0, and set HighThreshold = IntentScore (maximum score) + 10 and not higher than 100.

[0062] The training effect evaluation includes calculating the circle - drawing correction rate, intention recognition accuracy, and assistance matching degree;

[0063] The circle correction rate is based on the total number of steps taken by the patient, counted by the sensors of the assistive robot. The standard is whether the abduction angle of the affected lower limb is <15° and the hip and knee flexion angle meets the training requirements. The effective normal gait steps are identified and the calculation formula is: Circle correction rate = (effective normal gait steps / total steps) × 100%;

[0064] The intent recognition accuracy and assist matching degree are based on the single-modal classification of the ERD branch and the MRCP branch. The single-modal classification of the ERD branch is based on the ERD percentage ranging from 5% to 85%, from low to high, corresponding to the weak to strong motion intent and the corresponding percentage assist output of the robot from strong to weak. The single-modal classification of the MRCP branch is based on the MRCP negative peak amplitude ranging from -1 to -15μV, from large to small, corresponding to the weak to strong motion intent and the corresponding percentage assist output of the robot from strong to weak. The results of the single-modal classification are combined and compared with the intent level and the assist percentage in step s5 to obtain the intent recognition accuracy and assist matching degree.

[0065] The iterative scheme includes,

[0066] The robot-assisted adjustment is based on the following criteria: a circle correction rate of ≥80%, an intent recognition accuracy of ≥85%, an assistance matching degree of ≥90%, and an improvement of ≥6 points in the lower limb FMA score compared to the baseline. When the results of three consecutive training sessions meet any two of the criteria, the percentage of robot assistance output is gradually reduced within the range of 0-5.

[0067] The intensity of tDCS and the number of RIP facilitation steps were adjusted based on the following criteria: a circle correction rate of <60%, no improvement after two consecutive training sessions, low electromyographic activation on the affected side, and the patient being in the chronic phase of stroke. Once any one of these criteria was met, the intensity of tDCS was gradually increased, starting at 2 mA, increasing by +0.5 mA, with a maximum of 2.5 mA. The pre-activation sequence was maintained at 200 ms, the stimulation duration was 5–10 minutes, and the number of RIP facilitation steps was increased from 3 per step to 4–5.

[0068] As an example of this solution, during the preparation phase, a non-invasive 64-channel EEG cap and VR glasses are fitted to the patient's head, along with tDCS and rTMS electrodes. A assistive robot is worn on the patient's legs. The system is then activated for individual and baseline calibration. During training, the patient operates a controller to activate the VR training scene and actively imagines the "affected side knee flexion / hip flexion stepping" movement. The BCI intent recognition module decodes the intent intensity signal, and the assistive robot converts the intent intensity signal into servo control commands to drive the knee flexion joint to assist in hip and knee flexion, while the adjustable damped abduction limiting joint restricts the patient's leg abduction and circumduction. During the process, the VR scene displays the gait effect in real time based on feedback from the motion sensors on the assistive robot.

[0069] The temporal logic of the temporalized neural modulation module is as follows:

[0070] 1) Detect and identify BCI intent;

[0071] 2) tDCS is pre-activated;

[0072] 3) Assistive robots help patients complete actions;

[0073] 4) Repeated facilitation therapy to facilitate the flow of fluids;

[0074] 5) Reinforcement using rTMS;

[0075] Among them, the stimulation parameters of tDCS are 2mA at the anode for 20 minutes, and it is initiated 200ms before BCI recognizes the intention to move; the stimulation of rTMS is to inhibit the healthy side with a low frequency of 1Hz and excite the affected side with a high frequency of 10Hz, and to provide 5 minutes of stimulation after each training session to strengthen the neural connection of the movement; the repetitive facilitation therapy guides the patient to actively correct by facilitating tactile and verbal cues simultaneously.

[0076] Finally, the system collects multimodal data to evaluate the corrective effect and adjust the assist intensity and stimulation parameters for the next training session. In the single-modal grading of the ERD branch, a percentage of 5%–25% indicates weak cortical activation, representing weak intent; 26%–55% indicates clear cortical activation, representing moderate intent; and 56%–85% indicates strong cortical activation, representing strong intent. A percentage less than 5% is considered as no effective intent, and a percentage greater than 85% is cross-validated with the MRCP negative peak amplitude. If the MRCP negative peak amplitude is <-10μV, it is treated as "strong intent." In the single-modal grading of the MRCP branch, -1 to -3μV indicates weak intent, suggesting the signal may originate from the patient's sensory expectation rather than motor preparation; -4 to -7 indicates moderate intent, consistent with typical motor preparation potentials; and -8 to -15 indicates strong intent, suggesting the patient is in an active effort or compensatory state on the healthy side. In adjusting the intensity of tDCS and the number of RIP facilitations, when a patient enters a plateau phase after two consecutive training sessions without improvement, and when the patient's disease course exceeds 6 months, it is determined that the patient is in the chronic phase of stroke. The intensity of tDCS and the number of RIP facilitations are then gradually increased, coupled with motor output.

[0077] The split-type, two-segment assistive robot mechanically restricts abduction and guides normal hip and knee flexion, significantly reducing the range of circumduction during training. It uses a BCI intent recognition module to identify and output the intensity of the patient's BCI intent in gait, forming appropriate robot-assisted output based on the patient's movement intent. This avoids passive training, promotes neural remodeling, and combines sequential neural modulation with multimodal feedback to extend the plasticity window and help patients in the chronic phase overcome the recovery plateau.

[0078] By creating patient profiles through individual calibration, the training effect is evaluated and the program is iterated after training. Based on the dynamic adjustment of multimodal data, precise rehabilitation with "one policy for one person" is achieved, and the effect can be quantified and traced. The multi-scenario VR feedback module provides VR training scenarios and simulated outdoor walking support at the same time, which solves the problems of boring and poor compliance in traditional rehabilitation.

Claims

1. A gait correction system for stroke based on BCI intent decoding, characterized in that: Includes the following modules and central control sheet, The BCI intention recognition module is equipped with a non-invasive 64-channel EEG cap that covers the bilateral motor cortex M1 area and the supplementary motor area SMA. It is used to capture and decode the patient's active motor intention of knee / hip flexion, and to identify and output the intensity of the patient's BCI intention in gait. The lower limb robot assist module is equipped with a split dual-segment assist robot, including an adjustable damped abduction limiting joint set in the thigh segment, which dampens and blocks hip abduction >15°, and a motor-driven knee flexion joint set in the lower leg segment, which adjusts the output assist of the knee flexion joint according to the intensity of BCI intent, and is equipped with sensors to detect kinematic data including hip and knee flexion angle, stride length, stride frequency, and abduction angle. The multi-scene VR feedback module is equipped with VR glasses and a differential drive chassis. The VR glasses show patients VR training scenarios including walking on flat ground, climbing stairs, and crossing obstacles. The differential drive chassis is equipped with a center of gravity pressure sensor, which follows the patient's walking movements based on the sensor signal feedback. The sequential neuromodulation module is equipped with tDCS and rTMS electrodes. With tDCS as the core, the stimulation target points are the M1 area of ​​the primary motor cortex on the affected side and the lower limb representative area. It is combined with rTMS assistance and repetitive facilitation therapy for assistance. The stimulation target points of rTMS are the M1 area on the healthy side and the SMA area on the affected side. The repetitive facilitation therapy includes providing tactile stimulation to the skin after assisting in the completion of the movement and giving verbal prompts to the patient. The multimodal data acquisition and evaluation module, based on EEG cap detection and feedback from the assistive robot, collects kinematic and EEG data, which are then used for manual input of FMA scores and 10-meter walking speed data to evaluate training effectiveness and iterate the program. The central control unit (CCU) connects all modules for unified scheduling, enabling time-series coordination of intent detection, robot assistance, data feedback, and evaluation and control.

2. The stroke gait correction system based on BCI intent decoding according to claim 1, characterized in that: The process of analyzing movement intent includes, s1, EEG is collected by covering the M1 and SMA areas with an EEG cap. The core channels of the EEG cap include Cz, C3, C4, FCz, CPz, CP3, and CP4. The reference electrodes are bilateral mastoid processes (A1 / A2) or the average reference scheme. The sampling rate is 1000Hz, and DC coupling is used for acquisition. s2, preprocesses the collected data, including filtering, online artifact removal, and baseline calibration; s3, perform individual calibration, collect EEG data during multiple real gait movements of the patient, extract features, determine individual thresholds based on the features, and record them in the patient configuration file for storage; s4. Perform ERD+MRCP dual-modal feature fusion. Extract feature vectors from preprocessed EEG data, input the feature vectors into the classifier, output three-dimensional intent probability, use the three-dimensional intent probability to calculate the continuous value of intent intensity IntentScore, and substitute it into the hierarchical decision to obtain the intent level. S5 outputs the intention level to the assistive robot through the central control unit, and has a decision verification consisting of a confirmation window mechanism and false positive suppression. The assistive robot is linearly mapped to the assistance percentage based on the intention level.

3. The stroke gait correction system based on BCI intent decoding according to claim 2, characterized in that: The filtering in step s2 includes band-stop filtering using a second-order IIR filter and band-pass filtering using parallel dual-branch processing. The band-stop filtering uses a 50Hz power frequency notch filter with a bandwidth of 48-52Hz. The band-pass filtering is divided into an ERD branch and an MRCP branch. The ERD branch performs band-pass filtering at 8–30Hz to obtain μ-band 8–13Hz and β-band 14–30Hz. The MRCP branch uses a 0.05–3Hz zero-phase low-pass filter. The online artifact removal is based on an offline training model of independent component analysis (ICA). Online projection removes EOG and EMG artifacts and sets an adaptive threshold of ±100μV to remove large transient artifacts. The baseline calibration uses the average potential of the resting window 500ms before the trigger of motor imagery as the baseline and subtracts the baseline in real time for offset.

4. The stroke gait correction system based on BCI intent decoding according to claim 3, characterized in that: Step s3 involves having the patient perform 10-15 attempts at knee / hip flexion on the affected side and recording EEG. Feature extraction includes calculating the ERD percentage and MRCP negative peak amplitude for each movement and combining them to obtain the IntentScore for each movement. Individual thresholds are set by sorting the fusion scores from low to high, with the 30th percentile used as the "weak / moderate" cutoff value and the 70th percentile used as the "moderate / strong" cutoff value. The patient profile includes individual thresholds, ICA projection matrix, and filtering parameters.

5. The stroke gait correction system based on BCI intent decoding according to claim 4, characterized in that: The formula for calculating the ERD percentage is as follows: ,in, The energy of the μ+β band within the resting state window of -500 to 0 ms is obtained by calculating the total energy in μV². The total energy of the μ+β band within a 0–800ms window after the initiation of motor imagery is obtained by calculating the total energy of the μ+β band. The MRCP negative peak amplitude is obtained by setting an extraction window from 1000ms before the motor imagery trigger point to 200ms after the trigger point, acquiring the EEG signals of the three channels Cz, FCz, and CPz within the window, and calculating the average value to obtain the MRCP negative peak amplitude, in μV.

6. The stroke gait correction system based on BCI intent decoding according to claim 5, characterized in that: The feature vector in step s4 includes a 1D ERD percentage, a 2D μ-band relative power change, a 2D β-band relative power change, a 1D MRCP negative peak amplitude, a 1D MRCP negative peak latency, a 1D contralateral / ipsilateral ERD ratio, a 1D energy fluctuation variance within a time window, and a 1D baseline-corrected DC drift rate. The 2D μ-band relative power change includes the average values ​​of the Cz and C3 channels of EEG signal. The 2D β-band relative power change includes the average values ​​of the Cz and C4 channels of EEG signal. The MRCP negative peak latency is determined by using a sliding window with a length of 500ms and a step size of 50ms, detecting a negative trend within the MRCP negative peak amplitude extraction window. When the amplitude decreases for three consecutive windows and the minimum value is less than -2μV, it is determined to be MRCP. The onset function determines the MRCP negative peak latency; the opposite / same-side ERD ratio is the ratio of channels C3 and C4; the energy fluctuation variance within the time window is obtained based on the variance of instantaneous energy within the ERD branch window; the baseline-corrected DC drift rate is obtained by linear fitting within the window after baseline calibration, and the slope of the fitted straight line is taken.

7. The stroke gait correction system based on BCI intent decoding according to claim 6, characterized in that: The classifier in step s4 is based on a Lightweight Gradient Boosting Machine (LightGBM) model or an Online Adaptive Linear Discriminant Analysis (LDA) module to process feature vectors. The training strategy involves using historical data from 50-100 stroke patients with real motion labels to pre-train the base model offline. After each training iteration, the actual joint angle changes fed back by the robot are used as weak labels to fine-tune the classifier in a semi-supervised manner, updating it every 5 iterations. The three-dimensional intent probabilities, from weakest to strongest, are... , , The formula for calculating the continuous value of intent intensity is: The value range is 0-100; the determination criteria for the intention level in the hierarchical decision-making process include, a) A score >0.6 or IntentScore ≤33 indicates a weak intent, which will be output with a high boost, with the corresponding boost percentage ranging from 80% to 100%. b), > 0.5 and 33 < IntentScore ≤ 66 represents a medium intention, output medium assistance, and the corresponding assistance percentage range is 40% - 70%; c) A score >0.5 or IntentScore >66 indicates a strong intent, resulting in a low boost output, with the corresponding boost percentage ranging from 10% to 30%.

8. The stroke gait correction system based on BCI intent decoding according to claim 7, characterized in that: The confirmation window mechanism in step s5 starts from the moment the patient begins motor imagery and slides a window every 100ms. If the intent level classification results of two consecutive sliding windows are consistent, the output is a valid intent. If two consecutive windows are inconsistent, the previous valid result is output and a 50ms delay is set for re-detection. The maximum tolerance delay is set to 300ms. False positive suppression is achieved by immediately canceling the current assistance and triggering a VR prompt when the sensor of the assistive robot detects a joint angle change that does not match the intent type. For every 3 false positives, the system automatically increases the filtering order of the MRCP branch and recalibrates the baseline. The default filtering order is 4 and increases incrementally.

9. The stroke gait correction system based on BCI intent decoding according to claim 4, characterized in that: The individual threshold boundary value in step s3 has an individual calibration mechanism. By collecting the IntentScore of 10-15 real exercise training sessions of the patient, sorting them from smallest to largest, the 30th percentile is taken as the "weak / medium" boundary LowThreshold, and the 70th percentile is taken as the "medium / strong" boundary HighThreshold. Recalibration is performed every 5 training sessions. If the highest IntentScore is <30, LowThreshold is set to 1 / 3 of the maximum value, and HighThreshold is set to 2 / 3 of the maximum value. If the lowest IntentScore is >70, LowThreshold is set to IntentScore (minimum score) - 10 and not less than 0, and HighThreshold is set to IntentScore (maximum score) + 10 and not higher than 100.

10. The stroke gait correction system based on BCI intent decoding according to claim 7, characterized in that: Training effectiveness evaluation includes calculating the circle correction rate, intent recognition accuracy, and assist matching degree; The circle correction rate is based on the total number of steps taken by the patient, counted by the sensors of the assistive robot. The standard is whether the abduction angle of the affected lower limb is <15° and the hip and knee flexion angle meets the training requirements. The effective normal gait steps are identified and the calculation formula is: Circle correction rate = (effective normal gait steps / total steps) × 100%; The intent recognition accuracy and assist matching degree are based on the single-modal classification of the ERD branch and the MRCP branch. The single-modal classification of the ERD branch is based on the ERD percentage ranging from 5% to 85%, from low to high, corresponding to the weak to strong motion intent and the corresponding percentage assist output of the robot from strong to weak. The single-modal classification of the MRCP branch is based on the MRCP negative peak amplitude ranging from -1 to -15μV, from large to small, corresponding to the weak to strong motion intent and the corresponding percentage assist output of the robot from strong to weak. The results of the single-modal classification are combined and compared with the intent level and the assist percentage in step s5 to obtain the intent recognition accuracy and assist matching degree. The iterative scheme includes, The robot-assisted adjustment is based on the following criteria: a circle correction rate of ≥80%, an intent recognition accuracy of ≥85%, an assistance matching degree of ≥90%, and an improvement of ≥6 points in the lower limb FMA score compared to the baseline. When the results of three consecutive training sessions meet any two of the criteria, the percentage of robot assistance output is gradually reduced within the range of 0-5. The intensity of tDCS and the number of RIP facilitation steps were adjusted based on the following criteria: a circle correction rate of <60%, no improvement after two consecutive training sessions, low electromyographic activation on the affected side, and the patient being in the chronic phase of stroke. Once any one of these criteria was met, the intensity of tDCS was gradually increased, starting at 2 mA, increasing by +0.5 mA, with a maximum of 2.5 mA. The pre-activation sequence was maintained at 200 ms, the stimulation duration was 5–10 minutes, and the number of RIP facilitation steps was increased from 3 per step to 4–5.

Images