Brain-computer interface neural feedback rehabilitation training system

Abstract

The application discloses a brain-computer interface neural feedback rehabilitation training system, which is characterized in that: the system is capable of synchronously collecting electroencephalogram signals, functional near-infrared spectroscopy hemodynamic signals and transcranial magnetic stimulation neural evoked response signals, and performing space-time alignment; the system is capable of extracting event-related potential, power spectral density, blood oxygen saturation, motor evoked potential amplitude and latency characteristics to construct a full-dimensional neural function characteristic sequence; the system is capable of identifying characteristic broken sections and splicing to generate continuity recovery identification points; the system is capable of inputting the characteristic sequence and the identification points into a deep confidence network for deep correlation analysis, performing hierarchical label cutting and splitting on the injury deviation area to dynamically adjust the characteristic weight, matching a rehabilitation training scheme according to the hierarchical result, and collecting neural signals in real time to determine a neural plasticity optimal window; the system is capable of periodically updating the hierarchical result and correcting the rehabilitation training scheme when the change exceeds a threshold value, thereby improving the neural function injury grading precision, diagnosis consistency and rehabilitation scheme matching accuracy.

Description

Technical Field

[0001] This invention relates to the field of brain-computer interface and neurofunctional rehabilitation technology, and in particular to a brain-computer interface neurofeedback rehabilitation training system. Background Technology

[0002] Neurofunctional diagnosis and neurofeedback technology is a key technological direction in the field of human physiological parameter detection and neurofunctional identification and diagnosis. Its core technology system covers key links such as multimodal neural signal synchronous acquisition, neuroelectrophysiology and hemodynamics and evoked conduction feature extraction, quantitative grading of neurofunctional damage, dynamic neurofeedback regulation, and fusion of clinical assessment data. It can realize the accurate detection and dynamic assessment of the electrical activity, metabolic state and conduction pathway integrity of the central nervous system. It has wide application value in the diagnosis, surgical rehabilitation assessment and neurofunctional identification of central nervous system injury diseases such as stroke, traumatic brain injury and spinal cord injury. It has the core capabilities of multi-source signal integration, neural function modeling, damage grading determination and real-time feedback regulation.

[0003] However, traditional methods for grading and diagnosing neurological functional impairment are mostly based on qualitative or semi-quantitative evaluations of central nervous system function based on single-source neural signals, primarily analyzing single indicators such as brain electrical activity, local blood oxygen metabolism, or neural evoked responses. Traditional methods typically rely on physician experience, manual scoring using clinical scales, and independent testing with single-modality devices. Specifically, this includes visual waveform assessment after acquiring a single EEG signal, statistical grading after acquiring local blood oxygen data using a fixed near-infrared probe, and qualitative detection of neural pathways using transcranial magnetic stimulation alone. These methods rely solely on single-modality devices such as EEG, functional near-infrared spectroscopy, or transcranial magnetic stimulation in the neural signal acquisition stage. Signal acquisition is asynchronous, and data cannot be spatiotemporally aligned, allowing only the extraction of single-dimensional neural features. The grading process often employs static models such as expert scoring and fixed-weighting, using methods including waveform comparison, parameter threshold setting, and traditional multi-factor weighted scoring.

[0004] Traditional diagnostic methods often rely on single-modal static parameters and subjective manual assignment to classify damage levels. They lack fine-grained mechanisms for recognizing abnormal brain electrical activity, blood flow metabolic disturbances, and nerve conduction disruptions during dynamic changes in neurological function. This results in the true characteristics of brain damage under multimodal neural signal coupling being masked by static grading labels, making it difficult to present the true state of neurological function corresponding to points of neural feature mutation. Grading labels are prone to mismatch with local brain region damage characteristics. In situations with brain region functional abnormalities, neural pathway blockages, or metabolic disorders, traditional methods cannot effectively break down the grading structure or correct the boundary positions of damage labels. This leads to core problems such as insufficient granularity in neurological function damage grading, blurred diagnostic result boundaries, inaccurate matching of rehabilitation plans, and low diagnostic consistency, severely hindering the development of accurate diagnosis and personalized rehabilitation training for neurological function damage.

[0005] Therefore, there is an urgent need to provide a technical solution to address the above problems. Summary of the Invention

[0006] To address the aforementioned technical problems, this invention provides a brain-computer interface neurofeedback rehabilitation training system. The technical solution of this brain-computer interface neurofeedback rehabilitation training system is as follows: The multimodal neural signal synchronous acquisition module is used to synchronously acquire electroencephalogram (EEG) signals, functional near-infrared spectroscopy hemodynamic signals, and transcranial magnetic stimulation (TMS) evoked neural response signals and perform spatiotemporal alignment. The multi-source neural feature extraction and preprocessing module is used to perform baseline correction, power frequency interference filtering, and motion artifact removal on the spatiotemporally aligned signals, extract event-related potential features, power spectral density features, blood oxygen saturation features, and motion evoked potential amplitude and latency features, construct a full-dimensional neural function feature sequence, identify feature break segments based on the feature gradient changes of the full-dimensional neural function feature sequence, and perform node splicing on the feature break segments that meet the isomorphic conditions to generate neural function continuity recovery markers; The neurological function damage grading module is used to input the full-dimensional neurological function feature sequence and the neurological function continuity recovery markers into a deep belief network for deep association analysis, output the neurological function damage grading results, and perform grading label pruning and splitting operations on the neurological damage offset region where the feature fluctuation amplitude exceeds a preset threshold in continuous time sequence. Based on the grading label pruning and splitting operations, the feature weights are dynamically adjusted so that the grading labels match the actual damage state of the brain region. A graded and adaptive dynamic neurofeedback training module is used to match rehabilitation training programs based on the graded results of neurological function impairment, and to collect changes in neural signals in real time during the rehabilitation training process to determine the optimal window for neural plasticity. The injury score dynamic update and visualization output module is used to re-trigger neural signal acquisition, feature extraction and grading calculation according to a preset cycle, update the neural function injury grading results, and correct the rehabilitation training plan when the grading change exceeds a preset level change threshold, generating a visualized diagnostic report that includes injury grading, functional area localization and rehabilitation potential prediction.

[0007] The technical solution of this invention simultaneously acquires and spatiotemporally aligns electroencephalogram (EEG) signals, functional near-infrared spectroscopy hemodynamic signals, and transcranial magnetic stimulation (TMS) evoked neural response signals. It extracts event-related potentials (ERPs), power spectral density, blood oxygen saturation, and the amplitude and latency of motor evoked potentials to construct a full-dimensional neural function feature sequence. It identifies characteristic fracture segments and splices them to generate continuous recovery markers. The feature sequence and markers are input into a deep belief network for deep association analysis, and graded label trimming and splitting are performed on the injury offset area to dynamically adjust feature weights. Based on the grading results, a rehabilitation training program is matched, and neural signals are acquired in real time to determine the optimal window for neural plasticity. The grading results are periodically updated, and the rehabilitation training program is corrected when changes exceed a threshold. This solves the problems of insufficient grading granularity, blurred diagnostic boundaries, inaccurate rehabilitation program matching, and low diagnostic consistency caused by traditional diagnostic methods that rely on single-modal static parameters and subjective manual assignment. It improves the accuracy of neural function injury grading, diagnostic consistency, and the accuracy of rehabilitation program matching.

[0008] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, and in order to make the above and other objects, features and advantages of the present invention more apparent and understandable, specific embodiments of the present invention are described below. Attached Figure Description

[0009] 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.

[0010] The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings: Figure 1 This is a schematic diagram of an embodiment of a brain-computer interface neurofeedback rehabilitation training system according to the present invention. Detailed Implementation

[0011] Exemplary embodiments of the invention will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the invention are shown in the drawings, it should be understood that the invention can be implemented in various forms and should not be limited to the embodiments set forth herein.

[0012] Figure 1 A schematic diagram of an embodiment of a brain-computer interface neurofeedback rehabilitation training system provided by the present invention is shown. Figure 1 As shown, the brain-computer interface neurofeedback rehabilitation training system includes: The multimodal neural signal synchronous acquisition module 110 is used to synchronously acquire electroencephalogram (EEG) signals, functional near-infrared spectroscopy hemodynamic signals, and transcranial magnetic stimulation (TMS) evoked neural response signals and perform spatiotemporal alignment.

[0013] Among them, electroencephalogram (EEG) signals refer to the electrophysiological signals generated by the synchronous activity of postsynaptic potentials of brain neuron groups recorded from the scalp surface, which are then amplified by electrodes to form a time-series waveform. For example, when assessing the neurological function of a stroke patient, the waveform of electrical activity of the entire cerebral cortex collected from the scalp surface by a 64-channel wireless EEG acquisition unit at a sampling rate of 1000Hz is the EEG signal, in which the event-related potential component recorded in the central channel shows a significant amplitude attenuation.

[0014] Functional near-infrared spectroscopy hemodynamic signal refers to the optical signal obtained by detecting changes in the concentration of oxyhemoglobin and deoxyhemoglobin in the cerebral cortex after near-infrared light penetrates biological tissue, reflecting the blood flow metabolic response caused by neural activity in local brain regions. For example, when a stroke patient performs a hand motor imagery task, the 24-channel fNIRS acquisition unit detects that the increase in oxyhemoglobin concentration in the primary motor cortex region of the damaged hemisphere is significantly lower than that in the corresponding region of the healthy side through dual wavelength light sources of 780nm and 850nm. The concentration change curve constitutes the functional near-infrared spectroscopy hemodynamic signal.

[0015] Transcranial magnetic stimulation (TMS) evoked response signals refer to the muscle electrical activity response signals recorded in the target muscles after a pulsed magnetic field is applied to the scalp surface by a TMS coil, inducing nerve depolarization in the target brain region and transmitting the signal via a descending conduction pathway. For example, when assessing the neural conduction pathways of a stroke patient, the TMS evoked response unit stimulates the primary motor cortex on the affected side at 80% of the motor threshold intensity. The potential waveform recorded from the surface electrodes of the contralateral abductor pollicis brevis muscle is the TMS evoked response signal, in which the motor evoked potential amplitude is 1.2 mV and the latency is 24.5 ms.

[0016] The multi-source neural feature extraction and preprocessing module 120 is used to perform baseline correction, power frequency interference filtering, and motion artifact removal on each signal after spatiotemporal alignment. It extracts event-related potential features, power spectral density features, blood oxygen saturation features, and motion evoked potential amplitude and latency features to construct a full-dimensional neural function feature sequence. Based on the feature gradient changes of the full-dimensional neural function feature sequence, it identifies feature break segments and performs node splicing on the feature break segments that meet the isomorphic conditions to generate neural function continuity recovery markers.

[0017] Baseline correction refers to a signal preprocessing operation that estimates and subtracts slowly drifting DC offsets or low-frequency trend terms from the original neural signal. For example, in the case of a stroke patient's EEG signal, the overall signal drifts upward by 0.3mV due to the slow change in electrode-skin contact impedance during continuous acquisition. Baseline correction restores the signal to zero mean level by calculating the mean of the entire signal segment and subtracting the mean from all sampling points. Power frequency interference filtering refers to the process of suppressing or eliminating electromagnetic interference components generated by the AC power supply system mixed into the neural signal. For example, in the case of a stroke patient's EEG signal superimposed with 50Hz sine wave interference, power frequency interference filtering attenuates the interference component by 40dB using a 50Hz notch filter, essentially eliminating the 50Hz oscillation in the output waveform. Motion artifact removal refers to the process of identifying and removing non-neurogenic interference components introduced into neural signals by head movements, blinking, muscle tension, or changes in body posture. For example, a stroke patient exhibited two obvious head rotations during the data acquisition process. Motion artifact removal uses an independent component analysis algorithm to identify the time component that is highly correlated with the head motion accelerometer signal as an artifact component and remove it from the mixed signal.

[0018] Event-related potential (ERP) characteristics refer to the amplitude and latency quantification parameters of the positive or negative peak of evoked potentials extracted after time-locked superposition and averaging of EEG signals according to specific events. For example, the peak voltage of the P300 component extracted from the EEG waveform 300 ms after the presentation of the target stimulus in a stroke patient is 5.8 μV, and the peak latency is 312 ms, which is an EEPROM characteristic. Power spectral density characteristics refer to the quantification parameters of the energy distribution density of each frequency band on the frequency axis after frequency domain transformation of the EEG signal. For example, the α-band power spectral density characteristic extracted from the resting-state EEG signal of a stroke patient is 3.2 μV² / Hz, with the affected hemisphere showing a decrease of approximately 40% compared to the healthy hemisphere. Oxygen saturation characteristics refer to the quantitative parameters of changes or rates of change in oxyhemoglobin concentration, deoxyhemoglobin concentration, and total hemoglobin concentration extracted from fNIRS hemodynamic signals. For example, the oxygen saturation characteristics extracted from the task-state fNIRS hemodynamic signals of a stroke patient are: a rate of change of oxyhemoglobin concentration in the primary motor cortex on the affected side of 0.08 μM·mm / s, and a rate of change of 0.15 μM·mm / s in the corresponding region on the healthy side. Motor evoked potential amplitude and latency characteristics refer to the peak-to-peak voltage amplitude of the muscle compound action potential and the time delay from stimulus onset to potential appearance extracted from TMS neural evoked response signals. For example, the amplitude characteristic of the motor evoked potential recorded from the abductor pollicis brevis muscle on the affected side of a stroke patient is 1.2 mV, and the latency characteristic is 24.5 ms, while the corresponding parameters on the healthy side are 2.8 mV and 21.0 ms, respectively.

[0019] The full-dimensional neural function feature sequence refers to a time series of multidimensional feature vectors formed by arranging and combining all neurophysiological features, hemodynamic features, and neural evoked conduction features extracted from EEG signals, fNIRS hemodynamic signals, and TMS neural evoked response signals in chronological order within the same time period. For example, a full-dimensional neural function feature sequence is formed by arranging 12 types of parameters—event-related potential features, power spectral density features, blood oxygen saturation features, and motor evoked potential amplitude and latency features—within a 10-minute time window of a stroke patient in segments of 1 second each, forming a 12×600-dimensional feature matrix. Feature gradient change refers to the rate and direction of change of multimodal neural feature vectors between adjacent time windows in the full-dimensional neural function feature sequence. For example, in a stroke patient's full-dimensional neural function feature sequence, between the adjacent time windows of 180s and 185s, the motor evoked potential amplitude suddenly drops from 1.2mV to 0.4mV, and the calculated Euclidean distance of the multimodal feature vectors is 0.73; this distance value is the feature gradient change.

[0020] The "characteristic break segment" refers to a signal segment in the full-dimensional neural function feature sequence where the characteristic gradient change exceeds a preset gradient threshold, causing a discontinuity or abrupt jump in the neural function feature over time. For example, in a stroke patient's feature sequence, the characteristic gradient change at 180s reaches 0.73, exceeding the preset gradient threshold of 0.5. The signal intervals 2s before and after 180s are marked as characteristic break segments. The isomorphism condition refers to the constraint condition that the event-related potential components, power spectral density distribution, and blood oxygen saturation change rate of the EEG signal before and after the break segment are highly similar in statistical characteristics, thus determining that the break is caused by noise or artifacts rather than a real interruption of neural function. For example, in a stroke patient, the correlation coefficient of the event-related potential components before and after the break segment is 0.91, the KL divergence of the power spectral density distribution is 0.12, and the cosine similarity of the blood oxygen saturation change rate is 0.89. All three indicators fall within the preset similarity threshold range, satisfying the isomorphism condition.

[0021] Among them, the neural function continuity recovery marker point refers to the marker node generated after performing node splicing on the feature break segment that meets the isomorphic condition, which is used to mark the original break position in the neural function continuity spectrum that has been repaired; for example, at the 180s position of the full-dimensional neural function feature sequence of a stroke patient, the new data point generated by connecting the endpoints before and after the break through cubic spline interpolation is the neural function continuity recovery marker point, and the feature recovery is smooth after splicing.

[0022] The neurological function impairment grading calculation module 130 is used to input the full-dimensional neurological function feature sequence and the neurological function continuity recovery markers into a deep belief network for deep association analysis, output the neurological function impairment grading result, and perform grading label pruning and splitting operations on the neurological impairment offset region where the feature fluctuation amplitude exceeds a preset threshold in continuous time sequence. Based on the grading label pruning and splitting operations, the feature weights are dynamically adjusted so that the grading labels match the actual damage state of the brain region.

[0023] Deep belief networks (DBNs) are deep probabilistic generative models composed of multiple stacked Restricted Boltzmann Machines (RBMs) that undergo layer-by-layer unsupervised pre-training and supervised fine-tuning. For example, a DBN consisting of three stacked RBMs might take a stroke patient's full-dimensional neurological function feature sequence as input to the first layer. The output of the first layer, fused with neurological function continuity recovery markers, is input to the second layer. The output of the second layer is input to the third layer, and the output of the third layer is mapped by a softmax classification layer to generate probability distributions for each injury level. The neurological function injury grading result refers to the severity category of the injury output by the DBN based on the input full-dimensional neurological function feature sequence and neurological function continuity recovery markers after deep association analysis. For example, if a stroke patient's stroke injury probabilities are 0.87 for each level in the softmax classification layer output after analysis by the DBN, higher than other levels, the output neurological function injury grading result would be level 3 severe injury.

[0024] The characteristic fluctuation amplitude refers to the range of numerical changes or deviations from the baseline of a certain type of neural feature in a full-dimensional neural function feature sequence over a continuous time interval. For example, in a stroke patient, the amplitude of a motor evoked potential gradually decreases from 1.2mV to 0.4mV within a 10-minute time window and then partially recovers; the difference between the maximum and minimum amplitudes, 0.8mV, is the characteristic fluctuation amplitude. The preset threshold refers to a pre-set critical value used to determine whether the deviation of a neural feature reaches the condition for triggering the grading label operation. For example, the preset threshold for the characteristic fluctuation amplitude of motor evoked potential amplitude is 0.5mV. If a stroke patient's amplitude fluctuation of 0.8mV exceeds the preset threshold, subsequent grading label pruning and splitting operations are triggered.

[0025] The neurological injury offset region refers to a time-series segment where the amplitude of characteristic fluctuations exceeds a preset threshold, leading to a risk of mismatch between the grading label and the actual brain region injury state. For example, in a stroke patient, after the characteristic fluctuation amplitude exceeds the preset threshold, the time-series region where the amplitude of motor evoked potentials fluctuates abnormally continuously from 120s to 300s is marked as the neurological injury offset region. The grading label trimming and splitting operation refers to the dynamic adjustment operation by which the neurological functional injury grading calculation module 130 re-divides the boundaries of the original integrated grading label within the neurological injury offset region and generates new sub-level label boundaries within the label. For example, in a stroke patient, the original uniform level 3 injury label covers the region from 120s to 300s. The grading label trimming and splitting operation splits a level 2 sub-segment at 210s, and the original label is trimmed into three segments: "level 3-level 2-level 3," making the grading result more precisely match the temporal fluctuations of the brain region injury state.

[0026] The true state of brain injury refers to the comprehensive pathophysiological state of neuronal dysfunction, abnormal blood flow metabolism, or conduction pathway blockage that actually exists in a specific brain region of the central nervous system under the characterization of multimodal neural signals. For example, in a stroke patient, during the fluctuation of the amplitude of motor evoked potentials, the primary motor cortex on the affected side simultaneously shows a decrease in the rate of change of oxyhemoglobin concentration and a decrease in alpha band power. The dynamic evolution of the ischemic penumbra in the cortex revealed by the multimodal neural characteristics is the true state of brain injury.

[0027] The graded and adaptive dynamic neurofeedback training module 140 is used to match rehabilitation training programs according to the graded results of the neurological function damage, and to collect changes in neural signals in real time during the rehabilitation training process to determine the optimal window for neural plasticity.

[0028] The rehabilitation training program refers to a personalized neurorehabilitation intervention plan selected based on the neurological functional impairment grading results, including training type, stimulation parameters, training duration, and training frequency. For example, after a stroke patient is assessed as having a grade 3 severe injury, the matched rehabilitation training program is neuromodulation enhancement training, which includes 40 minutes of transcranial direct current stimulation with the anode placed on the primary motor cortex of the affected side and a stimulation current of 1.5 mA, combined with passive joint movement training. The optimal window for neuroplasticity refers to the time interval during rehabilitation training where the patient's neurological functional signals show a positive trend of plasticity and the training intervention effect is most significant. For example, in a stroke patient, after 18 minutes of neuromodulation enhancement training, the amplitude of motor evoked potentials gradually increased from 0.4 mV to 0.9 mV, and the rate of change in blood oxygen saturation improved simultaneously. The time period around the 18th minute was determined as the optimal window for neuroplasticity, and the current stimulation parameters were maintained to consolidate the plasticity effect.

[0029] It should be noted that the specific rules for determining the optimal window for neural plasticity are as follows: The moving average slopes of the amplitude of motor evoked potentials and the moving average slopes of the rate of change of blood oxygen saturation are calculated in real time over five consecutive time windows (each window being one minute long). When the slope of the amplitude of motor evoked potentials is positive for three consecutive windows and the cumulative increase exceeds 15% of the baseline value, and the slope of the rate of change of blood oxygen saturation is positive for three consecutive windows and the cumulative increase exceeds 10% of the baseline value, the subsequent 10-minute time interval from that moment is marked as the optimal window for neural plasticity, and the current stimulation parameters are maintained. If the above conditions are not met, the determination is repeated every 30 seconds. This rule is pre-labeled based on the statistical characteristics of plasticity bursts in clinical rehabilitation data.

[0030] The injury score dynamic update and visualization output module 150 is used to re-trigger neural signal acquisition, feature extraction and grading calculation according to a preset cycle, update the neural function injury grading results, and correct the rehabilitation training plan when the grading change exceeds the preset level change threshold, and generate a visualized diagnostic report that includes injury grading, functional area localization and rehabilitation potential prediction.

[0031] The preset cycle refers to the fixed time interval at which a new round of neural signal acquisition, feature extraction, and grading calculation updates are automatically triggered. For example, after the initial neurological function impairment grading assessment is completed, the preset cycle is 10 minutes, and at the 10-minute mark, the entire assessment process for a stroke patient is automatically retried. The preset grade change threshold refers to the pre-set critical value for determining whether the degree of change in the neurological function impairment grading result is sufficient to trigger the correction of the rehabilitation training program. For example, if the preset grade change threshold is 1 grade change, and a stroke patient is initially assessed as having a grade 3 injury but is updated to a grade 2 injury after 10 minutes, the grade change reaches 1 grade, meeting the preset grade change threshold and triggering the correction of the rehabilitation training program.

[0032] Among these, injury grading refers to the category label indicating the severity of injury in the neurological function injury grading results; for example, a stroke patient's injury grade in the initial assessment is grade 3 severe injury, and after an updated assessment at 10 minutes, the injury grade changes to grade 2 moderate injury. Functional area localization refers to the process or result of determining the specific cerebral cortical functional areas involved in neurological function injury based on the characteristic abnormal distribution of each brain region in multimodal neural signals; for example, in a stroke patient, the location of the abnormal source of event-related potentials and the area of ​​decreased blood oxygen saturation both point to the hand representative area of ​​the left hemisphere primary motor cortex, and the functional area localization is the left primary motor cortex. Rehabilitation potential prediction refers to the quantitative prediction of the degree and speed of future neurological function recovery of the patient based on multimodal neural function characteristics and markers of continuous recovery of neural function; for example, in a stroke patient, the density of markers of continuous recovery of neural function in the full-dimensional neural function characteristic sequence is high and the characteristic gradient shows a gradual upward trend, the rehabilitation potential prediction is medium to high level, and the Fugl-Meyer scale upper limb motor function score is expected to improve by 15 points within 4 weeks.

[0033] Among them, the visual diagnostic report refers to a diagnostic output file that integrates and presents information such as injury grading, functional area localization, rehabilitation potential prediction and clinical scale scoring in a graphical and chart-like form; for example, the visual diagnostic report generated for a stroke patient includes the injury grading area marked on the three-dimensional topographic map of the cerebral cortex, the coordinates of the functional area localization, the rehabilitation potential trend curve, and radar charts of the Fugl-Meyer scale and Barthel index.

[0034] The technical solution of this embodiment simultaneously acquires and spatiotemporally aligns electroencephalogram (EEG) signals, functional near-infrared spectroscopy hemodynamic signals, and transcranial magnetic stimulation (TMS) evoked neural response signals. It extracts event-related potentials (ERPs), power spectral density, blood oxygen saturation, and the amplitude and latency characteristics of motor evoked potentials to construct a full-dimensional neural function feature sequence. It identifies characteristic fracture segments and splices them to generate continuous recovery markers. The feature sequence and markers are input into a deep belief network for deep association analysis, and graded label pruning and splitting are performed on the injury offset area to dynamically adjust feature weights. Based on the grading results, a rehabilitation training program is matched, and neural signals are acquired in real time to determine the optimal window for neural plasticity. The grading results are periodically updated, and the rehabilitation training program is corrected when changes exceed a threshold. This solves the problems of insufficient grading granularity, blurred diagnostic boundaries, inaccurate matching of rehabilitation programs, and low diagnostic consistency caused by traditional diagnostic reliance on single-modal static parameters and subjective manual assignment. It improves the accuracy of neural function injury grading, diagnostic consistency, and the accuracy of rehabilitation program matching.

[0035] In one alternative embodiment, the multimodal neural signal synchronous acquisition module 110 includes a wireless EEG acquisition unit, a functional near-infrared spectroscopy acquisition unit, and a transcranial magnetic stimulation (TMS) neural evoked response unit. The wireless EEG acquisition unit, the functional near-infrared spectroscopy acquisition unit, and the TMS neural evoked response unit synchronously acquire whole-brain cortical neural electrical signals, dual-wavelength cerebral hemodynamic signals, and neural pathway evoked response signals covering the intensity and frequency ranges of motor thresholds at the same sampling rate.

[0036] In the above-mentioned optional methods, the wireless EEG acquisition unit, functional near-infrared spectroscopy acquisition unit, and transcranial magnetic stimulation nerve evoked unit are used to simultaneously acquire whole-brain cortical neural electrical signals, dual-wavelength cerebral hemodynamic signals, and neural pathway evoked response signals covering the range of motor threshold intensity and frequency at the same sampling rate. This solves the problems of asynchronous acquisition and incomplete signal coverage of single-modal devices and improves the synchronization and coverage of multimodal neural signal acquisition.

[0037] In one alternative approach, the multimodal neural signal synchronous acquisition module 110 is specifically used to: synchronously trigger the event-related potential time stamp of the EEG signal and the optical polar space coordinates of the functional near-infrared spectral hemodynamic signal through a joint timestamp to complete the spatiotemporal alignment.

[0038] In the above-mentioned optional methods, the event-related potential time stamp of the EEG signal and the optical polar space coordinates of the functional near-infrared spectroscopy hemodynamic signal are further synchronized by a joint timestamp to complete the spatiotemporal alignment, which solves the problems of inconsistent time references and spatial coordinate misalignment of different modal neural signals and improves the accuracy of spatiotemporal alignment of multimodal neural signals.

[0039] In an alternative embodiment, the multi-source neural feature extraction and preprocessing module 120 is specifically used for: The Euclidean distance between the multimodal neural feature vectors within adjacent time windows is calculated as the feature gradient change. When the Euclidean distance exceeds a preset gradient threshold, it is determined to be the feature break segment.

[0040] In the above-mentioned optional methods, the Euclidean distance between the multimodal neural feature vectors within adjacent time windows is further calculated as the feature gradient change, and the feature break segment is determined when the Euclidean distance exceeds the preset gradient threshold. This solves the problems of ambiguous localization of neural feature mutation boundaries and inaccurate identification of break segments, and improves the objectivity and accuracy of feature break segment detection.

[0041] In an alternative embodiment, the multi-source neural feature extraction and preprocessing module 120 is specifically used for: When the correlation coefficient of the event-related potential components of the EEG signals before and after the fracture, the KL divergence of the power spectral density distribution, and the cosine similarity of the rate of change of blood oxygen saturation of the functional near-infrared spectral hemodynamic signal are all within the preset similarity threshold range, it is determined that the isomorphism condition is met. Then, cubic spline interpolation is performed on the endpoints of the two characteristic fracture segments that meet the isomorphism condition to generate the neural function continuity recovery marker points.

[0042] In the above-mentioned optional methods, it is further determined that the isomorphic condition is met when the correlation coefficient of the event-related potential components, the KL divergence of the power spectral density distribution, and the cosine similarity of the rate of change of blood oxygen saturation are all within the preset threshold range. Then, cubic spline interpolation is performed on the endpoints of the two characteristic fracture segments that meet the isomorphic condition to generate neural function continuity restoration markers. This solves the problems of fuzzy fracture segment splicing conditions and inaccurate pseudo-continuity markers, and improves the rationality of marker generation and the accuracy of neural function state restoration.

[0043] In an alternative embodiment, the neurological function impairment grading calculation module 130 is specifically used for: A deep belief network consisting of multiple layers of restricted Boltzmann machines is constructed. The full-dimensional neural function feature sequence is input into the first layer of the restricted Boltzmann machine. The output of the previous layer of the restricted Boltzmann machine and the neural function continuity recovery markers are input into the middle layer of the restricted Boltzmann machine. The output of the last layer of the restricted Boltzmann machine is connected to a softmax classification layer, and the neural function damage grading result is output by the softmax classification layer.

[0044] In the above-mentioned optional approach, a deep belief network with multiple layers of stacked restricted Boltzmann machines is further constructed. The full-dimensional neural function feature sequence is input into the first layer of the restricted Boltzmann machine, and the output of the previous layer and the continuity recovery marker are input into the middle layer of the restricted Boltzmann machine. The last layer is connected to the softmax classification layer to output the hierarchical results. This solves the problem that traditional static models cannot deeply analyze the association of multimodal neural features, and improves the analytical depth and output stability of the hierarchical results.

[0045] In an alternative embodiment, the neurological function impairment grading calculation module 130 is specifically used for: When performing hierarchical label clipping and splitting operations, discrimination is based on the multimodal neural injury offset intensity and feature coupling degree. The multimodal neural injury offset intensity is calculated according to the following formula: ; In the formula, This represents the intensity of multimodal neural injury offset. The amplitude of motion-evoked potentials within the current time window. The amplitude of the motor evoked potential is used as a reference for normal brain regions. The latency period of motion-evoked potentials within the current time window. This is the reference latency for motor evoked potentials in normal brain regions. This represents the rate of change in blood oxygen saturation. This is the power spectral density shift. For amplitude offset weight, The weighting is the latency shift. This is the blood oxygen coupling coefficient. This is the power spectral density shift adjustment coefficient. This is the weighted offset between amplitude and latency.

[0046] It should be noted that the multimodal neural injury offset intensity formula combines the relative offset of motor evoked potential amplitude with the weighted index of blood oxygen saturation change rate, combines the relative offset of motor evoked potential latency with the logarithmic adjustment of power spectral density offset, and superimposes the product term of amplitude and latency relative offset. Through amplitude offset weight, latency offset weight, blood oxygen coupling coefficient, power spectral density offset adjustment coefficient, and linkage offset weight, it reflects the nonlinear coupling effect of electrophysiological characteristics, hemodynamic characteristics, and neural conduction time characteristics in the multimodal neural injury process, thus forming a comprehensive quantitative measure of neural injury offset intensity. The above formula provides the basis for the neural function injury grading calculation module 130 to trigger the grading label clipping and splitting operation. When the multimodal neural injury offset intensity exceeds the preset threshold, it indicates that the brain region injury characteristics have shifted significantly, and dynamic adjustment of grading labels needs to be initiated.

[0047] In the above-mentioned optional methods, the multimodal neural injury offset intensity and feature coupling degree are further judged when performing hierarchical label clipping and splitting operations. The damage offset intensity is quantified by the linkage calculation of amplitude offset, latency offset, blood oxygen saturation change rate and power spectral density offset, which solves the problem that a single indicator cannot reflect the comprehensive offset degree of multimodal neural injury and improves the quantitative accuracy of the triggering basis of clipping and splitting operations.

[0048] In an alternative embodiment, the neurological function impairment grading calculation module 130 is specifically used for: When the hierarchical label cropping and splitting operation is triggered, the hierarchical score compression factor is calculated according to the following formula: ; In the formula, This is a rating compression factor. As the dominant damage characteristic value, These are characteristic values ​​of normal brain regions. This represents the intensity of multimodal neural injury offset. The feature gradient modulus of the full-dimensional neural function feature sequence. This is the compression rate adjustment coefficient. This represents the gradient coupling coefficient.

[0049] It should be noted that the grading score compression factor formula is based on the normalized difference between the dominant injury feature value and the normal brain region feature value, combined with the exponential decay term of the multimodal neural injury offset intensity and the hyperbolic tangent function term of the feature gradient modulus of the full-dimensional neural function feature sequence. The decay rate and gradient coupling degree are controlled by the compression rate adjustment coefficient and the gradient coupling coefficient. The compression factor is obtained by subtracting the above product from the unit value of 1, so that the grading score compression factor decreases as the injury offset intensity increases and the feature gradient becomes steeper. The above formula provides the basis for the neural function injury grading calculation module 130 to dynamically adjust the feature weights, and uses the grading score compression factor to correct the contribution of injury features in the grading decision, so that the grading label more accurately matches the actual injury state of the brain region.

[0050] Among them, the magnitude offset weight Latency offset weight Blood oxygen coupling coefficient β, power spectral density shift adjustment coefficient Linkage offset weight and compression rate adjustment coefficient Gradient coupling coefficient All parameters were calibrated using a combination of grid search and cross-validation based on clinical trial data: full-dimensional neurofunctional feature sequences were collected from at least 200 patients with different injury levels. The injury grading results annotated by clinical experts were used as the gold standard. The optimal values ​​were obtained within a pre-defined parameter space by minimizing the cross-entropy loss function between the grading results and the gold standard. Typical value ranges are: =0.4~0.6, =0.3~0.5, =0.2~0.4, =0.5~0.8, =0.1~0.3, =0.6~0.9, =0.3~0.7. This embodiment can be adjusted within the above range or recalibrated using the same method according to the specific application scenario.

[0051] In the above-mentioned optional methods, a grading score compression factor is further calculated when the grading label clipping and splitting operation is triggered. The dynamic compression adjustment is performed by combining the difference between the dominant damage feature value and the normal brain region feature value with the multimodal neural damage offset intensity and feature gradient modulus. This solves the problem of reduced damage level discrimination caused by fixed compression of grading scores and improves the flexibility of score compression and the precision of damage boundary characterization.

[0052] In one alternative approach, the graded adaptive dynamic neurofeedback training module 140 stores rehabilitation training programs corresponding to the graded results of neurological functional impairment at each level, wherein grade 1 impairment corresponds to basic motor imagery training, grade 2 impairment corresponds to mirror neuron activation training, grade 3 impairment corresponds to neuromodulation enhancement training, and grade 4 impairment corresponds to passive neural stimulation training.

[0053] In the above-mentioned optional methods, rehabilitation training programs corresponding to the graded neurological function damage results at each level are further stored in the graded adaptation dynamic neurofeedback training module 140, so that grades 1 to 4 damage correspond to basic motor imagery training, mirror neuron activation training, neuromodulation enhancement training and passive nerve stimulation training, respectively. This solves the problems of chaotic matching between rehabilitation training programs and damage levels and homogenization of training content, and improves the pertinence and personalization accuracy of program grading.

[0054] In one alternative approach, the injury score dynamic update and visualization output module 150 incorporates the Fugl-Meyer scale and the Barthel index. The visualized diagnostic report also includes the Fugl-Meyer scale score and the Barthel index score. The injury score dynamic update and visualization output module 150 performs a bidirectional correlation mapping between the neurological function injury grading result and the Fugl-Meyer scale score and the Barthel index score.

[0055] The Fugl-Meyer Assessment is a standardized clinical evaluation tool used to quantitatively assess the degree of recovery of sensory and motor function after stroke. It includes dimensions of motor function, sensation, balance, range of motion, and pain. For example, a stroke patient who underwent the assessment scored 36 points on the upper limb motor function section of the Fugl-Meyer Assessment, out of a possible 66 points. The Barthel Index is a standardized scale used to assess a patient's independence in activities of daily living (ADL). It includes scores for the ability to perform basic ADL activities such as eating, bathing, dressing, and walking. For example, a stroke patient who underwent the assessment scored 55 points on the Barthel Index, out of a possible 100 points, indicating that they require moderate assistance with ADL activities.

[0056] In the above-mentioned optional methods, the Fugl-Meyer scale and Barthel index are further integrated into the dynamic update and visualization output module 150 for injury scoring. This allows for a bidirectional correlation mapping between the neurological functional impairment grading results and the Fugl-Meyer scale score and Barthel index score, solving the problem of the disconnect between the single internal grading results and clinical assessment standards, and improving the consistency between the diagnostic report and the clinical rehabilitation assessment system.

[0057] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A brain-computer interface neurofeedback rehabilitation training system, characterized in that, The system includes: The multimodal neural signal synchronous acquisition module is used to synchronously acquire electroencephalogram (EEG) signals, functional near-infrared spectroscopy hemodynamic signals, and transcranial magnetic stimulation (TMS) evoked neural response signals and perform spatiotemporal alignment. The multi-source neural feature extraction and preprocessing module is used to perform baseline correction, power frequency interference filtering, and motion artifact removal on the spatiotemporally aligned signals, extract event-related potential features, power spectral density features, blood oxygen saturation features, and motion evoked potential amplitude and latency features, construct a full-dimensional neural function feature sequence, identify feature break segments based on the feature gradient changes of the full-dimensional neural function feature sequence, and perform node splicing on the feature break segments that meet the isomorphic conditions to generate neural function continuity recovery markers; The neurological function damage grading module is used to input the full-dimensional neurological function feature sequence and the neurological function continuity recovery markers into a deep belief network for deep association analysis, output the neurological function damage grading results, and perform grading label pruning and splitting operations on the neurological damage offset region where the feature fluctuation amplitude exceeds a preset threshold in continuous time sequence. Based on the grading label pruning and splitting operations, the feature weights are dynamically adjusted so that the grading labels match the actual damage state of the brain region. A graded and adaptive dynamic neurofeedback training module is used to match rehabilitation training programs based on the graded results of neurological function impairment, and to collect changes in neural signals in real time during the rehabilitation training process to determine the optimal window for neural plasticity. The injury score dynamic update and visualization output module is used to re-trigger neural signal acquisition, feature extraction and grading calculation according to a preset cycle, update the neural function injury grading results, and correct the rehabilitation training plan when the grading change exceeds a preset level change threshold, generating a visualized diagnostic report that includes injury grading, functional area localization and rehabilitation potential prediction.

2. The brain-computer interface neurofeedback rehabilitation training system according to claim 1, characterized in that, The multimodal neural signal synchronous acquisition module includes a wireless EEG acquisition unit, a functional near-infrared spectroscopy acquisition unit, and a transcranial magnetic stimulation neural evoked response unit. The wireless EEG acquisition unit, the functional near-infrared spectroscopy acquisition unit, and the transcranial magnetic stimulation neural evoked response unit synchronously acquire whole-brain cortical neural electrical signals, dual-wavelength cerebral hemodynamic signals, and neural pathway evoked response signals covering the intensity and frequency ranges of motor thresholds at the same sampling rate.

3. The brain-computer interface neurofeedback rehabilitation training system according to claim 2, characterized in that, The multimodal neural signal synchronous acquisition module is specifically used to: synchronize the event-related potential time stamp of the EEG signal with the optical polar space coordinates of the functional near-infrared spectral hemodynamic signal through a joint timestamp to complete the spatiotemporal alignment.

4. The brain-computer interface neurofeedback rehabilitation training system according to claim 1, characterized in that, The multi-source neural feature extraction and preprocessing module is specifically used for: The Euclidean distance between the multimodal neural feature vectors within adjacent time windows is calculated as the feature gradient change. When the Euclidean distance exceeds a preset gradient threshold, it is determined to be the feature break segment.

5. The brain-computer interface neurofeedback rehabilitation training system according to claim 4, characterized in that, The multi-source neural feature extraction and preprocessing module is specifically used for: When the correlation coefficient of the event-related potential components of the EEG signals before and after the fracture, the KL divergence of the power spectral density distribution, and the cosine similarity of the rate of change of blood oxygen saturation of the functional near-infrared spectral hemodynamic signal are all within the preset similarity threshold range, it is determined that the isomorphism condition is met. Then, cubic spline interpolation is performed on the endpoints of the two characteristic fracture segments that meet the isomorphism condition to generate the neural function continuity recovery marker points.

6. The brain-computer interface neurofeedback rehabilitation training system according to claim 1, characterized in that, The neurological function impairment grading calculation module is specifically used for: A deep belief network consisting of multiple layers of restricted Boltzmann machines is constructed. The full-dimensional neural function feature sequence is input into the first layer of the restricted Boltzmann machine. The output of the previous layer of the restricted Boltzmann machine and the neural function continuity recovery markers are input into the middle layer of the restricted Boltzmann machine. The output of the last layer of the restricted Boltzmann machine is connected to a softmax classification layer, and the neural function damage grading result is output by the softmax classification layer.

7. The brain-computer interface neurofeedback rehabilitation training system according to claim 6, characterized in that, The neurological function impairment grading calculation module is specifically used for: When performing hierarchical label clipping and splitting operations, discrimination is based on the multimodal neural injury offset intensity and feature coupling degree. The multimodal neural injury offset intensity is calculated according to the following formula: ； In the formula, This represents the intensity of multimodal neural injury offset. The amplitude of motion-evoked potentials within the current time window. The amplitude of the motor evoked potential is used as a reference for normal brain regions. The latency period of motion-evoked potentials within the current time window. This is the reference latency for motor evoked potentials in normal brain regions. This represents the rate of change in blood oxygen saturation. This is the power spectral density shift. For amplitude offset weight, The latency shift weight, This is the blood oxygen coupling coefficient. This is the power spectral density shift adjustment coefficient. This is the weighted offset between amplitude and latency.

8. The brain-computer interface neurofeedback rehabilitation training system according to claim 7, characterized in that, The neurological function impairment grading calculation module is specifically used for: When the hierarchical label cropping and splitting operation is triggered, the hierarchical score compression factor is calculated according to the following formula: ； In the formula, This is a rating compression factor. As the dominant damage characteristic value, These are characteristic values ​​of normal brain regions. This represents the intensity of multimodal neural injury offset. The feature gradient modulus of the full-dimensional neural function feature sequence. This is the compression rate adjustment coefficient. This represents the gradient coupling coefficient.

9. The brain-computer interface neurofeedback rehabilitation training system according to claim 1, characterized in that, The graded adaptive dynamic neurofeedback training module stores rehabilitation training programs corresponding to the graded results of neurological functional impairment at each level. Grade 1 impairment corresponds to basic motor imagery training, Grade 2 impairment corresponds to mirror neuron activation training, Grade 3 impairment corresponds to neuromodulation enhancement training, and Grade 4 impairment corresponds to passive neural stimulation training.

10. The brain-computer interface neurofeedback rehabilitation training system according to claim 1, characterized in that, The damage score dynamic update and visualization output module incorporates the Fugl-Meyer scale and the Barthel index. The visualized diagnostic report also includes the Fugl-Meyer scale score and the Barthel index score. The damage score dynamic update and visualization output module performs a bidirectional correlation mapping between the neurological function damage grading results and the Fugl-Meyer scale score and the Barthel index score.

Images