Method for analyzing the improvement effect of patients with implanted brain-computer interface controlling spinal cord electrical stimulation

Abstract

The application discloses an implantable brain-computer interface control spinal cord electrical stimulation patient improvement effect analysis method, relates to the technical field of medical rehabilitation, and the method comprises the following steps: multi-dimensional cooperative data acquisition: electrodes are implanted below target brain areas and spinal cord injury segments, detection elements are installed at key positions of exoskeletons, sensors are attached to preset muscle groups of lower limbs, electroencephalogram signals, SCS stimulation parameters, EXS motion data and neuromuscular response data are synchronously collected, and time correlation marks are embedded; by adopting the mode that multi-source signal preprocessing is combined with a multi-modal intention recognition model, the reliability of motion intention decoding is remarkably improved, in the signal preprocessing stage, an adaptive filtering algorithm combining Kalman filtering and wavelet threshold denoising is used, and SCS electrical stimulation interference, EXS motor noise and physiological noise in the electroencephalogram signals are effectively removed.

Description

Technical Field

[0001] This invention relates to the field of medical rehabilitation technology, specifically a method for analyzing the improvement effects of implanted brain-computer interface-controlled spinal cord electrical stimulation on patients. Background Technology

[0002] In the field of medical rehabilitation technology, spinal cord injury is a disease that seriously threatens patients' quality of life. As a key channel for the transmission of nerve signals between the brain and various parts of the body, once the spinal cord is damaged, it will directly disrupt the nerve transmission pathways between the brain, spinal cord, and muscles, leading to varying degrees of motor dysfunction in patients. This seriously affects patients' ability to take care of themselves in daily life and their social participation. In recent years, the SCS-EXS system, which combines spinal cord electrical stimulation with an exoskeleton, has been gradually applied to clinical rehabilitation treatment. This system provides patients with passive training through the synergistic effect of epidural electrical stimulation and exoskeleton, which can activate neuromuscular activity to a certain extent and promote partial recovery of limb motor function.

[0003] However, the existing SCS-EXS system has significant limitations in practical applications. On the one hand, the system mainly focuses on passive training and cannot respond to patients' active movement intentions. In the rehabilitation process, the patient's active participation is crucial for the remodeling of neural function and the recovery of motor function. However, the existing system is unable to construct an active neural cascade control pathway between the brain, spinal cord, and muscles, which limits the further improvement of rehabilitation effects. On the other hand, although brain-computer interface (BCI) EEG acquisition and intention recognition technology can decode the brain's motor intentions, it lacks synergy with spinal cord electrical stimulation and exoskeleton systems when used alone. It cannot fully leverage the advantages of each technology and is difficult to achieve precise and efficient rehabilitation treatment. In addition, the existing methods for analyzing the improvement effect are also insufficient. They only evaluate passive training or single brain control functions and do not cover the multi-dimensional data correlation after the integration of brain control technology and the SCS-EXS system. This one-sided evaluation method cannot comprehensively and accurately reflect the patient's rehabilitation status and cannot provide sufficient technical basis for optimizing personalized brain control rehabilitation programs, resulting in a lack of pertinence and effectiveness in rehabilitation treatment. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a method for analyzing the improvement effects of implantable brain-computer interface-controlled spinal cord electrical stimulation on patients. Through in-depth analysis and processing of multi-source synchronous data, it can accurately identify the patient's motor intentions and establish a multi-dimensional correlation model to quantitatively analyze the conduction and synergistic characteristics of the brain-spinal cord-muscle nerve cascade pathway. Based on this, a comprehensive evaluation system is constructed from multiple important dimensions such as neural pathways, motor function, and long-term recovery. This system can systematically evaluate the reconstruction effect of the brain-spinal cord-muscle nerve cascade control pathway, providing solid technical support for adjusting rehabilitation programs and enabling rehabilitation treatment to more accurately meet the individual needs of patients.

[0005] To address the aforementioned technical problems, this invention provides the following technical solution: a method for analyzing the improvement effects of implanted brain-computer interface-controlled spinal cord electrical stimulation on patients, the method comprising:

[0006] Multi-dimensional collaborative data acquisition: By implanting electrodes in the target brain region and below the spinal cord injury segment, installing detection elements in key parts of the exoskeleton, and attaching sensors to preset muscle groups in the lower limbs, EEG signals, SCS stimulation parameters, EXS motion data and neuromuscular response data are collected simultaneously and time-related markers are embedded.

[0007] Multi-source signal preprocessing and noise suppression: For the acquired EEG, SCS, EXS and neuromuscular signals, interference and deviations are removed and deviations are corrected by dynamic filtering, parameter calibration, baseline correction and threshold detection, respectively, and effective signal features and key moment information are extracted;

[0008] BCI Motion Intent Recognition and Accuracy Verification: Multi-dimensional features are extracted from preprocessed EEG and spinal cord evoked potential signals. A deep learning model including feature fusion and mapping layers is constructed and trained. The model accuracy is verified through accuracy, confusion matrix and robustness tests under different stimulus intensities.

[0009] Brain-spinal cord-muscle cascade pathway association analysis: Calculate the time interval between EEG intention initiation, spinal cord response, and muscle activation; quantify the causal influence of EEG signals on other signals; and compare the overlap between the actual EXS movement trajectory and the standard intention trajectory, as well as the matching rate between SCS stimulation and intention.

[0010] Comprehensive evaluation of improvement effects across multiple dimensions: A quantitative scoring system based on pathway delay and SCEP amplitude changes is constructed to assess neural pathway reconstruction. The Fugl-Meyer scale is combined with additional brain control-specific assessments and daily life scenario tests to assess motor function. Long-term recovery trends are assessed by monitoring signal changes at multiple follow-up nodes.

[0011] Furthermore, in the multi-source signal preprocessing and noise suppression steps, array-type metal electrodes are implanted in the motor cortex, supplementary motor area, and parietal sensory cortex to collect EEG signals containing rhythmic signals and event-related synchronization / desynchronization signals within a preset frequency band. SCS stimulation trigger markers and EXS motor command markers are embedded in the EEG signals. SCS stimulation parameters and SCS stimulation patterns associated with motor intentions are recorded. SCS stimulation parameters include stimulation frequency, intensity, and pulse width. The motion parameters of the EXS exoskeleton and the time delay between exoskeleton movement and SCS stimulation are collected. The motion parameters include joint rotation angle, driving torque, and movement speed. Spinal cord evoked potentials (SCEPs) are collected through implanted spinal cord electrodes. Electromyography (EMG) sensors are attached to the surface of preset muscle groups in the lower limbs to collect motor unit action potentials (MUAPs).

[0012] Furthermore, in the multi-source signal preprocessing and noise suppression steps, interference suppression and calibration processing are performed on the collected multi-source data. An adaptive filtering algorithm combining Kalman filtering and wavelet threshold denoising is used to denoise the EEG signals, removing SCS electrical stimulation interference, EXS motor noise, and physiological noise, while retaining the EEG signal frequency bands related to motor intention. Amplitude drift correction is performed on the SCS stimulation parameters. Zero-position calibration is performed on the EXS exoskeleton joint angle data, which is performed before each training session. Linear trend elimination is used to remove baseline drift from the SCEP signal. Bandpass filtering is performed on the MUAP signal, and peak detection is performed on the filtered MUAP signal using a preset threshold to extract the muscle activation initiation time.

[0013] Furthermore, in the BCI motion intent recognition and accuracy verification step, a motion intent recognition model is constructed based on the preprocessed data and its accuracy is verified. Time-frequency features, spatial features, and ERD / ERS features are extracted from the EEG signal, and peak features of the SCEP signal are extracted. The extracted features are then fused. A multimodal fusion deep learning model combining EEG-CNN and Transformer is used to train the fused features. Baseline data from a predetermined number of patients are used as the training and validation sets, and the model is optimized using the cross-entropy loss function. The recognition accuracy and confusion matrix of multiple types of motion intent are calculated using the intent recognition confidence formula. The robustness of the model's intent recognition under different SCS stimulus intensities is tested, and the model's accuracy and stability are verified.

[0014] Furthermore, in the BCI motion intent recognition and accuracy verification step, the recognition accuracy and confusion matrix of multiple types of motion intents are calculated using the intent recognition confidence formula, which is as follows: ,in For the first Confidence level for identifying motion intent For the first EEG feature matching values ​​of intention-like, For the first SCEP feature matching value of class intent, The cross-modal correlation coefficient between EEG and SCS. , Match the maximum value for all intent features.

[0015] Furthermore, in the brain-spinal cord-muscle cascade pathway association analysis step, a multi-dimensional association model is established to analyze the conduction and synergistic characteristics of the neural cascade pathway. A time series cross-correlation algorithm is used to locate the EEG intention trigger point, SCEP peak time, and MUAP activation time, calculate the time difference between each time point, and take the average of the time differences of multiple repeated actions. The Granger causality test algorithm is used to analyze the causal influence of the EEG intention signal on SCS stimulation parameters, EXS movement trajectory, and muscle MUAP activation. Based on the BCI intention recognition results, the dynamic time warping algorithm is used to calculate the overlap between the EXS movement trajectory and the intention. The matching rate between the SCS stimulation segments and parameters and the intention is statistically analyzed.

[0016] Furthermore, in the brain-spinal cord-muscle cascade pathway association analysis step, the Granger causality test algorithm is used to analyze the causal influence of EEG intention signals on SCS stimulation parameters, EXS motor trajectory, and muscle MUAP activation. The algorithm formula is as follows: ,in, It is a Granger causality test of EEG intention signal to target signal. The F-statistic is used to quantify the causal driving strength of EEG intention on the target signal. The target signal includes SCS stimulation parameters, EXS motion trajectory, and MUAP activation signal. The larger the F-value, the more significant the active control association between EEG intention and the target signal. It is the sum of squared residuals when constructing a prediction model using the historical data of the target signal itself, reflecting the natural variation pattern of the target signal without external driving force. It is the sum of squared residuals when a prediction model is constructed using both historical data of the target signal itself and historical data of the EEG intention signal. It reflects the change in the prediction error of the target signal after incorporating EEG intention-driven parameters. It is the model lag order, determined based on the frequency of the path signal acquisition. It is the total number of signal samples used for modeling.

[0017] Furthermore, in the brain-spinal cord-muscle cascade pathway association analysis step, the dynamic time warping algorithm is used to calculate the overlap between the EXS motion trajectory and the intention. The algorithm formula is as follows: ,in, This represents the degree of overlap between the actual movement trajectory of the EXS exoskeleton and the standard trajectory corresponding to the EEG intention. The closer the value is to 1, the higher the matching accuracy between the EXS movement and the EEG intention. This is the actual trajectory data of EXS. With intention standard trajectory data The dynamic time-warped distance is calculated by constructing the distance matrix between the two trajectories and finding the optimal matching path. The smaller the distance, the higher the trajectory similarity. These are the actual motion trajectory data of key joints in the EXS exoskeleton. These are the standard motion trajectory data of key joints corresponding to the brainwave intent; It is the maximum DTW distance between the EXS trajectory and the standard trajectory under this type of intent, and is calibrated offline by simulating extreme scenarios when the EXS trajectory deviates completely from the intent.

[0018] Furthermore, in the comprehensive evaluation step of the multi-dimensional improvement effect, an evaluation system is constructed and quantified from the dimensions of neural pathways, motor function, and long-term recovery. The evaluation indicators of the quantitative scoring system include pathway delay, SCEP amplitude change, and muscle activation synchronicity. The total score of pathway integrity is calculated based on the scores of each evaluation indicator. A brain-controlled specific scoring dimension is added based on the Fugl-Meyer scale, including joint active range of motion, gait stability, and movement completion efficiency. The patient's daily living activities ability under brain-controlled assistance is tested. A follow-up period is set, during which the changes in the energy of the motor cortex rhythmic signal and the changes in SCEP amplitude are monitored to evaluate the neuroplasticity of the brain and spinal cord.

[0019] Compared with existing technologies, this method for analyzing the improvement effects of implanted brain-computer interface-controlled spinal cord electrical stimulation on patients has the following beneficial effects:

[0020] I. This invention significantly improves the reliability of motion intention decoding by combining multi-source signal preprocessing with a multimodal intention recognition model. In the signal preprocessing stage, an adaptive filtering algorithm combining Kalman filtering and wavelet threshold denoising is used to effectively remove SCS electrical stimulation interference, EXS motor noise, and physiological noise from the EEG signal, while retaining the frequency bands related to motion intention. A multimodal fusion deep learning model is constructed and trained by integrating multiple features, which can more accurately identify motion intention. By calculating the recognition accuracy and confusion matrix of multiple types of motion intentions and testing the robustness under different SCS stimulation intensities, the model accuracy is ensured to be stable, providing a reliable basis for subsequent rehabilitation treatment.

[0021] Second, this invention applies Granger causality test to the assessment of neural cascade pathways, realizing quantitative analysis of pathway synergy. It can clearly present the causal influence of EEG intention signals on SCS stimulation parameters, EXS motor trajectory and muscle MUAP activation. Furthermore, it constructs a multi-dimensional assessment system that takes into account both neural repair and functional improvement. It conducts a comprehensive assessment from multiple dimensions such as neural pathways, motor function and long-term recovery, which helps to formulate more scientific and effective rehabilitation plans.

[0022] Other advantages, objectives and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination or study, or may be learned from the practice of the invention. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0024] Figure 1 A flowchart illustrating the method for analyzing the improvement effects of implanted brain-computer interface-controlled spinal cord electrical stimulation in patients;

[0025] Figure 2 A flowchart illustrating the steps of BCI (Brain-Computer Interface) motion intention recognition and accuracy verification in analyzing the improvement effect of implanted brain-computer interface-controlled spinal cord electrical stimulation in patients. Detailed Implementation

[0026] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0027] Example 1

[0028] The patient is a 52-year-old male who suffered a complete T9 segment spinal cord injury due to being struck by a heavy object. Eight months after the injury, he had completely lost his lower limb motor function (Fugl-Meyer lower limb motor score of 0) and showed no signs of voluntary muscle contraction. He had previously received 4 months of SCS-EXS passive training, but there was no significant improvement in the range of motion of his lower limb joints.

[0029] Under general anesthesia, a 4×4 array of platinum-iridium alloy EEG electrodes with a contact diameter of 0.5 mm and a spacing of 1 mm were implanted in the patient's motor cortex (M1 area) and auxiliary motor area via stereotactic surgery. An 8-contact SCS stimulation electrode with a length of 2 cm was implanted in the T11-L5 segment of the spinal cord. A lower limb EXS exoskeleton was customized according to the patient's lower limb dimensions (thigh length 55 cm, calf length 48 cm). Angle sensors with an accuracy of ±0.5° and torque sensors with an accuracy of ±0.1 N·m were installed in the hip, knee, and ankle joints, respectively. The exoskeleton controller was connected to the implanted signal transmission module via Bluetooth 5.0. Surface electromyography (EMG) sensors were attached to the surface of the quadriceps femoris, hamstring, tibialis anterior, and gastrocnemius muscles of the lower limb. A time synchronizer with an accuracy of 1 μs was used to calibrate the EEG acquisition equipment, SCS stimulator, EXS controller, and EMG sensors to ensure that the data timestamp deviation of each device was ≤1 ms.

[0030] Three data acquisition phases were set up. The baseline period was 1 week before treatment, with data collected once a day for 60 minutes each time. This included 30 minutes of passive SCS-EXS training data (SCS stimulation frequency 40Hz, intensity 3.5mA, pulse width 250μs), 20 minutes of BCI single intention recognition data (patient imagines 8 types of movement intentions such as raising leg, stepping, standing, and sitting), and 10 minutes of neurophysiological baseline data (SCEP, electromyographic latency). The treatment period was 12 weeks, with training 3 times a week for 60 minutes each time. Static neurophysiological data was collected 5 minutes before training, brain-controlled SCS-EXS synergistic data (including EEG signals, SCS stimulation patterns, EXS movement parameters, and electromyographic data) was collected 30 minutes during training, and dynamic neurophysiological test data was collected 25 minutes after training. The follow-up period was 1 month, 3 months, and 6 months after treatment. Data was collected repeatedly at each follow-up node according to the dimensions collected at the baseline period, with each collection lasting 40 minutes.

[0031] Using the MATLAB signal processing toolbox, the 40Hz power frequency interference and harmonics generated by SCS stimulation were first suppressed using the Kalman filter algorithm. Then, the 120-350Hz noise generated by the EXS motor operation was removed using the wavelet threshold denoising algorithm. Simultaneously, physiological artifacts such as ECG and ECG were eliminated. Finally, the μ rhythm, β rhythm, and ERD / ERS signals related to motor intention within the 8-30Hz frequency band were retained. Baseline correction was performed on the processed EEG signals, using the signal mean during the intentionless imagery period as a benchmark to eliminate baseline drift. The SCS stimulation parameter data were calibrated weekly using a high-precision signal acquisition instrument. The stimulator output parameters are compared with the actual measured values ​​to correct for amplitude drift in the stimulation intensity, ensuring that the intensity recording error is ≤0.1mA. Before each training session, the EXS exoskeleton is controlled to move the patient's lower limbs to a zero-position posture of standing and full extension. The output value of the angle sensor at this time is used as a reference to perform zero-position calibration on the joint angle data. The SCEP signal is processed using a linear trend elimination method to remove baseline drift. The electromyographic signal is bandpass filtered from 20 to 500 Hz. The peak detection threshold is set to three times the standard deviation of the baseline noise. The muscle activation initiation time is extracted to ensure that the signal-to-noise ratio of the electromyographic signal is ≥15dB.

[0032] EEG and SCEP data from patients at baseline and during the first four weeks of treatment were collected as training samples, totaling 15,000 samples, which were divided into a training set (10,500 samples) and a validation set (4,500 samples) in a 7:3 ratio. A multimodal fusion deep learning model combining EEG-CNN and Transformer was used as input. The time-frequency features (extracted through short-time Fourier transform), spatial features (electrode array topology correlation), and peak features of SCEP signals were input. The model was optimized and trained using the cross-entropy loss function. During training, the intent recognition confidence formula was used. Assessing the reliability of identification—based on the "leg-raising" intention ( For example, the baseline period , Cross-modal correlation coefficient , , Calculated After 4 weeks of treatment , , The confidence level was significantly improved. Before each training session, patients were asked to complete an imagery test of eight types of motor intentions in sequence, with each type of intention tested three times, for a total of 24 tests. The recognition accuracy of each type of intention was calculated by the model. When the overall accuracy was stable above 88%, the training for that day was started. If the accuracy was below 85%, the impedance of the EEG electrodes was measured using an impedance meter. If the impedance was out of range, the electrode position was adjusted surgically or the electrode contacts were cleaned. The electrodes were then recalibrated and tested again until the accuracy reached the target.

[0033] Using MATLAB time series analysis, the trigger point for motor intention (initiation time of μ-rhythm ERD), the peak time of SCEP signal, and the activation initiation time of EMG signal were located in the EEG signal. Taking leg-raising intention as an example, the time difference between intention triggering and spinal cord response, and between spinal cord response and muscle activation, was calculated, and the average time difference of 30 repeated movements was taken. Before treatment, the total pathway delay was 450 ms, of which the delay between intention triggering and spinal cord response was 280 ms, and the delay between spinal cord response and muscle activation was 170 ms. After 12 weeks of treatment, the total pathway delay decreased to 230 ms, the delay between intention triggering and spinal cord response was 140 ms, and the delay between spinal cord response and muscle activation was 90 ms. Granger causality test was implemented using a Python deep learning framework, and the formula was used. Analyzing the causal influence of EEG intention on target signals, and the effect of pre-treatment EEG on SCS stimulation. After 12 weeks of treatment The causal effect strength increased by 76%; EEG on electromyography activation The score improved from 1.5 to 3.9, a 67% increase, based on the BCI intent recognition results and using the DTW trajectory overlap formula. Calculate the degree of overlap between the EXS trajectory and the intention, and the pre-treatment "leg-lifting" intention. After 12 weeks of treatment The matching rate between SCS stimulation segments and parameters and the intended treatment was statistically analyzed. The matching rate was 48% before treatment and increased to 90% after 12 weeks of treatment.

[0034] Based on the scoring criteria of pathway delay (40 points), SCEP amplitude change (30 points), and muscle activation synchronicity (30 points), a pathway integrity score was completed. Before treatment, the patient's pathway integrity score was 20 points, which increased to 72 points after 12 weeks of treatment. A brain-controlled specific score (total score of 20 points) was added based on the Fugl-Meyer scale. Before treatment, the score was 0 points, which increased to 16 points after 12 weeks of treatment. The patient could complete voluntary leg raising and continuous steps 15 times with brain-controlled SCS-EXS assistance, with a stepping speed of 0.25 m / s. At the 6-month follow-up after treatment, the patient's motor cortex μ rhythm energy increased by 20% and SCEP amplitude increased by 28% compared with the end of treatment. The activities of daily living test showed that the patient could independently climb up and down 8 steps and stand independently for 10 minutes with brain-controlled assistance. The patient could also independently control the lower limbs to complete simple turning movements, and the self-care ability was significantly improved.

[0035] Example 2

[0036] The patient is a 45-year-old male who suffered an incomplete L1 segment spinal cord injury due to a fall from a height. He presented to the clinic 6 months after the injury. He had partially lost lower limb motor function after the injury, with grade 1 strength in the left quadriceps femoris and grade 2 strength in the right quadriceps femoris. The strength of the hamstrings and tibialis anterior muscles was grade 1 bilaterally. His Fugl-Meyer lower limb motor score was 18. He was unable to stand independently and needed to move slowly with the help of a walker. He relied on family members for assistance with daily transfers. He had previously received 3 months of traditional SCS-EXS passive training, which only slightly improved the range of motion of the lower limb joints and did not significantly improve active motor control. In addition, he often experienced numbness in his lower limbs during training due to the inappropriate intensity of SCS stimulation.

[0037] Under general anesthesia, stereotactic surgery was performed to implant 3×4 array platinum-iridium alloy EEG electrodes in the patient's motor cortex (M1 area, responsible for lower limb motor control) and supplementary motor area. The electrode contacts were 0.4 mm in diameter and 1.2 mm apart. Intraoperative neurophysiological monitoring confirmed that the electrode contacts could effectively collect motor-related EEG signals (significant β-rhythm inhibition was detected at at least 6 contacts when the patient imagined raising their leg). A 10-contact SCS stimulation electrode with a length of 2.5 cm was implanted in the T12-L4 spinal cord to ensure that the stimulation contacts covered the key motor nerve pathways below the spinal cord injury segment. Based on the patient's lower limb dimensions (thigh length 52 cm, calf length 4 cm), the electrodes were implanted. A 6cm customized lower limb EXS exoskeleton is used, with angle sensors with an accuracy of ±0.3° and torque sensors with an accuracy of ±0.08N・m installed at the hip, knee, and ankle joints, respectively. The contact areas of the exoskeleton at the knee and ankle joints are padded with breathable medical sponge (6mm thick) to avoid skin pressure injuries caused by long-term wear. Disposable surface electromyography (EMG) sensors are attached to the surface of the quadriceps, hamstrings, tibialis anterior, and gastrocnemius muscles on both sides of the lower limbs. A time synchronizer with an accuracy of 0.5μs is used to calibrate the EEG acquisition equipment, SCS stimulator, EXS controller, and EMG sensors. After calibration, the time stamp deviation of the data from each device is controlled within 0.8ms to ensure the time consistency of multi-source data.

[0038] Three data acquisition phases were set up. The baseline period was 10 days before treatment, with a total of 5 data collection sessions, each lasting 55 minutes. Each session included 25 minutes of passive SCS-EXS training data (initial SCS stimulation frequency 35Hz, intensity 3.2mA, pulse width 220μs, fine-tuned according to patient tolerance to avoid significant numbness or discomfort), 20 minutes of BCI-based intention recognition data (patients imagined four basic movement intentions: bilateral leg raising, knee flexion, knee extension, and stepping, with each intention repeated 15 times), and 10 minutes of neurophysiological baseline data (collecting SCEP latency and resting-state electromyographic amplitude). The treatment period was 14 weeks, with training 4 times a week for 55 minutes each time. Static neurophysiological data (SCEP, baseline electromyography values) were collected 8 minutes before training. Brain-controlled SCS-EXS synergistic data (real-time recording of EEG signals, SCS stimulation parameter adjustment process, EXS movement trajectory data, and electromyography activation signals) were collected 30 minutes during training. Dynamic neurophysiological test data (including SCEP response under different SCS stimulation intensities and electromyography contraction strength test) were collected 17 minutes after training. The follow-up period was 2 months, 4 months, and 6 months after treatment. Data was collected repeatedly at each follow-up node according to the dimensions collected at baseline, with each collection lasting 45 minutes. The focus was on supplementing brain-controlled voluntary movement data (data related to active movement after the patient was no longer passively assisted).

[0039] The raw EEG signals were processed using the MATLAB signal processing toolbox. First, a Kalman filter algorithm was used to suppress 35Hz power frequency interference and 2nd and 3rd harmonic interference generated by SCS stimulation. Then, a wavelet threshold denoising algorithm was used to remove 100-300Hz broadband noise generated by the EXS motor. Simultaneously, independent component analysis was used to remove physiological noise such as ECG and EEG signals. Finally, μ rhythms (8-13Hz), β rhythms (13-28Hz), and ERD / ERS signals related to motor intention within the 7-28Hz frequency band were retained. After processing, baseline correction was performed on the EEG signals, using the mean EEG signal of the patient in a resting state (5 minutes of relaxation with eyes closed) as the benchmark to eliminate baseline drift and ensure a signal-to-noise ratio ≥13dB for the processed EEG signals. For the SCS stimulation parameters, the actual parameters of the stimulator output were calibrated every 3 days using a high-precision signal analyzer, and the results were compared and recorded. The deviation between the recorded value and the actual value is corrected for amplitude drift in subsequent recorded stimulus intensity and frequency data using a linear correction formula, ensuring that the parameter recording error is ≤0.08mA (intensity) and ≤1Hz (frequency). Before each training session, the EXS exoskeleton is controlled to bring the patient's lower limbs to a supine position with the lower limbs naturally extended, and the output value of the angle sensor at this time is used as a reference to perform zero-position calibration on the hip, knee, and ankle joint rotation data to eliminate measurement errors caused by mechanical wear of the exoskeleton. The SCEP signal is treated with a linear trend elimination method to remove baseline drift, so that the signal baseline fluctuation range is controlled within ±2μV. The electromyography signal is bandpass filtered from 15 to 550Hz, and the peak detection threshold is set to 2.5 times the standard deviation of the electromyography signal baseline noise. The signal peaks exceeding the threshold are automatically identified and marked, and the moment of peak occurrence is used as the muscle activation start time, ensuring that the electromyography activation moment identification error is ≤5ms.

[0040] Five baseline EEG data points (12,000 samples in total) and SCEP data (8,000 samples in total) were collected from patients as initial training data, divided into a training set (16,000 samples) and a validation set (4,000 samples) at an 8:2 ratio. A multimodal fusion deep learning model combining EEG-CNN and Transformer was used as input, incorporating the time-frequency characteristics (frequency distribution of different time periods extracted through short-time Fourier transform), spatial characteristics (signal correlation between different EEG electrode contacts), and ERD / ERS features (β). The model inputs the amplitude and duration of rhythm inhibition, along with the peak amplitude and latency characteristics of the SCEP signal. These features are then fused and input into the model for training. The model is optimized using the cross-entropy loss function, with a training cycle of 60 rounds. After each round of training, the model performance is evaluated using a validation set. Initial training is stopped when the validation set accuracy consistently exceeds 90%. During the treatment period, the latest training data (approximately 3000 samples) from patients is collected every 3 weeks to incrementally fine-tune the model parameters, adapting to the dynamic changes in the patient's EEG signals as the treatment progresses (such as increased amplitude of motor intention-related EEG features).

[0041] Before each training session, patients completed four types of motor intention imagery tests sequentially, 12 times for each type, for a total of 48 tests. The recognition accuracy for each type of intention and the overall accuracy were calculated using a model. Training was initiated for the day when the overall accuracy was ≥88%. If the accuracy was <85%, the EEG electrode impedance was checked first (normal range 4-14kΩ). If the impedance was outside the range, the electrode contacts and scalp contact area were cleaned with saline solution, reconnected, and tested again. If the impedance was normal, the baseline intensity of SCS stimulation was adjusted (±0.2mA) to reduce the interference of stimulation on the EEG signal until the accuracy reached the target. For example, before a training session in the 5th week of treatment, the patient's step intention recognition accuracy was 82%. The test revealed that the impedance of two EEG electrodes reached 18kΩ. After cleaning, the impedance dropped to 10kΩ, and the accuracy improved to 91% upon retesting, meeting the training initiation criteria.

[0042] Using MATLAB time series analysis, cross-correlation algorithms were employed to locate the trigger point of motor intention (beta rhythm ERD initiation), the peak time of SCEP signal (the moment of maximum amplitude of spinal cord response), and the activation initiation time of electromyography (EMG) signal (the moment when the EMG peak first exceeds the threshold) in EEG signals. Taking bilateral leg raising intention as an example, the time difference between intention triggering and spinal cord response, and between spinal cord response and muscle activation were calculated. Each type of intention was tested 25 times, and the average time difference was taken as the pathway delay index. Before treatment, the delay between intention triggering and spinal cord response was 260 ms, the delay between spinal cord response and muscle activation was 150 ms, and the total pathway delay was 410 ms. After 14 weeks of treatment, the delay between intention triggering and spinal cord response was 130 ms, the delay between spinal cord response and muscle activation was 80 ms, and the total pathway delay decreased to 210 ms, indicating a significant improvement in pathway conduction efficiency.

[0043] Granger causality tests were implemented using a deep learning framework to analyze the causal influence of EEG intention signals on SCS stimulation parameter adjustments, EXS motion trajectory changes, and electromyographic activation. Before treatment, the causal influence of EEG intention on SCS stimulation parameters was weak, and its causal influence on electromyographic activation could only explain 15%. After 14 weeks of treatment, the causal influence of EEG intention on SCS stimulation parameters could explain 52% of the parameter changes, and its causal influence on electromyographic activation could explain 48%, indicating that EEG intention could effectively drive downstream pathway responses. Based on BCI intention recognition results, the overlap between EXS motion trajectory and intention standard trajectory was calculated using the Dynamic Time Warping (DTW) algorithm. Before treatment, the overlap of the knee flexion intention trajectory was only 58%, which increased to 90% after 14 weeks of treatment. The matching rate between SCS stimulation segments and parameters and intention was statistically analyzed. Before treatment, the matching rate was 65%, which increased to 93% after 14 weeks of treatment, ensuring that SCS stimulation could accurately match the EEG intention requirements.

[0044] Based on a quantitative scoring system of pathway delay (35 points), SCEP amplitude change (30 points), and muscle activation synchronicity (35 points), a pathway integrity score was completed. Before treatment, the patient's pathway integrity score was 22 points, which improved to 86 points after 14 weeks of treatment. A brain-controlled specific scoring dimension (total score of 25 points) was added based on the Fugl-Meyer scale, including joint active range of motion (8 points), gait stability (9 points), and movement completion efficiency (8 points). Before treatment, the brain-controlled specific score was 3 points, which improved to 21 points after 14 weeks of treatment. At this time, the patient could independently complete bilateral leg raises and continuous steps 20 times with brain-controlled SCS-EXS assistance, with a stepping speed of 0.3 m / s. The intensity of SCS stimulation could be automatically adjusted according to the EEG intention, and the patient had no obvious numbness or discomfort.

[0045] Three follow-up periods were set up at 2 months, 4 months, and 6 months after treatment. In each period, motor cortex rhythm signals and SCEP amplitude data were collected. At the 6-month follow-up, the β rhythm energy of the patient's motor cortex increased by 28% compared with the end of treatment, and the SCEP amplitude induced by SCS stimulation increased by 32% compared with the end of treatment, indicating that the plasticity of the brain and spinal cord nerves continued to improve. In the activities of daily living test, the patient could independently climb up and down 12 steps, walk 80 meters independently, and put on and take off shoes and socks independently. The Fugl-Meyer lower limb motor score improved to 42 points, and the patient was completely weaned off the dependence on the walking aid. The self-care ability was greatly improved, the patient's satisfaction with the rehabilitation effect was significantly improved, and the patient took the initiative to express his willingness to participate in community rehabilitation exchange activities.

[0046] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A method for analyzing the improvement effect of implanted brain-computer interface-controlled spinal cord electrical stimulation on patients, characterized in that, The method includes: Multi-dimensional collaborative data acquisition: EEG signals, SCS spinal cord stimulation parameters, EXS exoskeleton motion data and neuromuscular response data are acquired. The EEG signals are acquired through implanted electrodes in the target brain region and embedded with time-related markers. The EXS exoskeleton motion data are acquired through detection elements installed in key parts of the exoskeleton. The neuromuscular response data are acquired through sensors attached to preset muscle groups in the lower limbs. Multi-source signal preprocessing and noise suppression: The acquired EEG signals, SCS spinal cord stimulation parameters, EXS exoskeleton motion data and neuromuscular response data are processed by dynamic filtering, parameter calibration, baseline correction and threshold detection to remove interference and correct deviations, and extract effective signal features and key moment information. BCI Motor Intent Recognition and Accuracy Verification: Multi-dimensional features are extracted from preprocessed EEG and spinal cord evoked potential signals (SCEP). A deep learning model including feature fusion and mapping layers is constructed and trained. The model accuracy is verified through accuracy, confusion matrix and robustness tests under different stimulus intensities. Brain-spinal cord-muscle cascade pathway association analysis: Calculate the time interval between EEG intention initiation, spinal cord response, and muscle activation; quantify the causal influence of EEG signals on SCS spinal cord electrical stimulation parameters, EXS motor trajectory, and muscle MUAP activation; and compare the overlap between the actual EXS motor trajectory and the standard intention trajectory, as well as the matching rate between SCS stimulation and intention. Comprehensive evaluation of improvement effects across multiple dimensions: A quantitative scoring system based on pathway delay and SCEP amplitude changes is constructed to assess neural pathway reconstruction. The Fugl-Meyer scale is combined with additional brain control-specific assessments and daily life scenario tests to assess motor function. Long-term recovery trends are assessed by monitoring signal changes at multiple follow-up nodes.

2. The method for analyzing the improvement effect of implantable brain-computer interface-controlled spinal cord electrical stimulation on patients according to claim 1, characterized in that, In the multi-source signal preprocessing and noise suppression steps, array-type metal electrodes implanted in the motor cortex, supplementary motor area, and parietal sensory cortex are used to collect EEG signals containing rhythmic signals and event-related synchronization / desynchronization signals within a preset frequency band. SCS stimulation trigger markers and EXS motor command markers are embedded in the EEG signals. SCS spinal cord stimulation parameters and SCS stimulation patterns associated with motor intentions are recorded. SCS spinal cord stimulation parameters include stimulation frequency, intensity, and pulse width. The motion parameters of the EXS exoskeleton and the time delay between exoskeleton movement and SCS stimulation are collected. The motion parameters include joint rotation angle, driving torque, and movement speed. Spinal cord evoked potentials (SCEPs) are collected through implanted spinal cord electrodes. Electromyography (EMG) sensors are attached to the surface of preset muscle groups in the lower limbs to collect motor unit action potentials (MUAPs).

3. The method for analyzing the improvement effect of implantable brain-computer interface-controlled spinal cord electrical stimulation on patients according to claim 1, characterized in that, In the multi-source signal preprocessing and noise suppression steps, the acquired multi-source data undergoes interference suppression and calibration processing. An adaptive filtering algorithm combining Kalman filtering and wavelet threshold denoising is used to denoise the EEG signals, removing SCS electrical stimulation interference, EXS motor noise, and physiological noise, while retaining the EEG signal frequency bands related to motor intention. Amplitude drift correction is performed on the SCS spinal cord electrical stimulation parameters. Zero-position calibration is performed on the EXS exoskeleton joint rotation data, which is performed before each training session. Linear trend elimination method is used to remove baseline drift from SCEP signals; The MUAP signal is bandpass filtered, and the peak value of the filtered MUAP signal is detected by a preset threshold to extract the muscle activation initiation time.

4. The method for analyzing the improvement effect of implantable brain-computer interface-controlled spinal cord electrical stimulation on patients according to claim 1, characterized in that, In the BCI motion intent recognition and accuracy verification step, a motion intent recognition model is constructed based on the preprocessed data and its accuracy is verified. Time-frequency features, spatial features, and ERD / ERS features are extracted from the EEG signal, and peak features of the SCEP signal are extracted. The extracted features are fused. A multimodal fusion deep learning model combining EEG-CNN and Transformer is used to train the fused features. Baseline data of a preset number of patients are used as the training set and validation set. The cross-entropy loss function is used to optimize the training of the model. The recognition accuracy and confusion matrix of multiple types of motion intentions were calculated using the intention recognition confidence formula. The robustness of the model in intention recognition under different SCS stimulus intensities was tested, and the accuracy and stability of the model were verified.

5. The method for analyzing the improvement effect of implantable brain-computer interface-controlled spinal cord electrical stimulation on patients according to claim 4, characterized in that, In the BCI motion intent recognition and accuracy verification step, the recognition accuracy and confusion matrix of multiple types of motion intents are calculated using the intent recognition confidence formula, which is as follows: ,in For the first Confidence level for identifying motion intent For the first EEG feature matching values ​​of intention-like, For the first SCEP feature matching value of class intent, The cross-modal correlation coefficient between EEG and SCS. , Match the maximum value for all intent features.

6. The method for analyzing the improvement effect of implantable brain-computer interface-controlled spinal cord electrical stimulation on patients according to claim 1, characterized in that, In the brain-spinal cord-muscle cascade pathway association analysis step, a multi-dimensional association model is established to analyze the conduction and synergistic characteristics of the neural cascade pathway. A time series cross-correlation algorithm is used to locate the EEG intention trigger point, SCEP peak time, and MUAP activation time, calculate the time difference between each time point, and take the average of the time differences of multiple repeated actions. The Granger causality test algorithm is used to analyze the causal influence of the EEG intention signal on SCS spinal cord electrical stimulation parameters, EXS movement trajectory, and muscle MUAP activation. Based on the BCI intention recognition results, the dynamic time warping algorithm is used to calculate the overlap between the EXS movement trajectory and the intention. The matching rate between the SCS stimulation segment and parameters and the intention is statistically analyzed.

7. The method for analyzing the improvement effect of implantable brain-computer interface-controlled spinal cord electrical stimulation on patients according to claim 6, characterized in that, In the brain-spinal cord-muscle cascade pathway association analysis step, the Granger causality test algorithm is used to analyze the causal influence of EEG intention signals on SCS spinal cord electrical stimulation parameters, EXS motor trajectory, and muscle MUAP activation. The algorithm formula is as follows: ,in, It is a Granger causality test of EEG intention signal to target signal. The F-statistic is used to quantify the causal driving strength of EEG intention on the target signal. The target signal includes SCS spinal cord stimulation parameters, EXS motor trajectory, and MUAP activation signal. The larger the F-value, the more significant the active control association between EEG intention and the target signal. It is the sum of squared residuals when constructing a prediction model using the historical data of the target signal itself, reflecting the natural variation pattern of the target signal without external driving force. It is the sum of squared residuals when a prediction model is constructed using both historical data of the target signal itself and historical data of the EEG intention signal. It reflects the change in the prediction error of the target signal after incorporating EEG intention-driven parameters. It is the model lag order, determined based on the frequency of the path signal acquisition. It is the total number of signal samples used for modeling.

8. The method for analyzing the improvement effect of implantable brain-computer interface-controlled spinal cord electrical stimulation on patients according to claim 6, characterized in that, In the brain-spinal cord-muscle cascade pathway association analysis step, the dynamic time warping algorithm is used to calculate the overlap between the EXS motion trajectory and the intention. The algorithm formula is as follows: ,in, This represents the degree of overlap between the actual movement trajectory of the EXS exoskeleton and the standard trajectory corresponding to the EEG intention. The closer the value is to 1, the higher the matching accuracy between the EXS movement and the EEG intention. This is the actual trajectory data of EXS. With intention standard trajectory data The dynamic time-warped distance is calculated by constructing the distance matrix between the two trajectories and finding the optimal matching path. The smaller the distance, the higher the trajectory similarity. These are the actual motion trajectory data of key joints in the EXS exoskeleton. These are the standard motion trajectory data of key joints corresponding to the brainwave intent; It is the maximum DTW distance between the EXS trajectory and the standard trajectory under this type of intent, and is calibrated offline by simulating extreme scenarios when the EXS trajectory deviates completely from the intent.

9. The method for analyzing the improvement effect of implantable brain-computer interface-controlled spinal cord electrical stimulation on patients according to claim 1, characterized in that, In the comprehensive evaluation step of the multi-dimensional improvement effect, an evaluation system is constructed and quantified from the dimensions of neural pathways, motor function and long-term recovery. The evaluation indicators of the quantitative scoring system include pathway delay, SCEP amplitude change and muscle activation synchronicity. The total score of pathway integrity is calculated based on the scores of each evaluation indicator. Based on the Fugl-Meyer scale, a brain-controlled specific scoring dimension was added, including joint active range of motion, gait stability, and movement completion efficiency; the ability of patients to perform daily living activities with brain-controlled assistance was tested; a follow-up period was set to monitor changes in the energy of motor cortex rhythmic signals and changes in SCEP amplitude during the follow-up period, and to assess the neuroplasticity of the brain and spinal cord.

Images