Method for myelin repair assistance based on neural signal modeling

Abstract

The present application relates to a kind of myelin repair auxiliary method based on neural signal modeling.The method obtains the multi-modal neural signal data such as electroencephalogram signal of individual to be rehabilitated and is preprocessed, forms preprocessed data set;Subsequently, individualized time series modeling process is executed, and the time series characteristic representation including nerve conduction velocity recovery rate, phase consistency and neural pathway integrity is generated, and individualized calibration is carried out by dynamic baseline updating mechanism.The calibrated feature representation is input to the pre-trained deep learning inference engine.The inference engine outputs myelin repair index and imbalance risk score, which is used to represent the degree of repair and predict future abnormal risk.Based on the above index, a set of intervention parameters is generated, including electrical stimulation intervention parameters, virtual rehabilitation training task difficulty and rehabilitation training rhythm, and further forming control instructions to configure virtual rehabilitation training environment and electrical stimulation execution interface, so as to realize the individualized assistance and dynamic optimization of myelin regeneration repair.

Description

Technical Field

[0001] This invention relates to the field of healthcare informatics technology, and in particular to an auxiliary method for myelin repair based on neural signal modeling. Background Technology

[0002] In existing technologies, research on myelin repair after central nervous system injury mainly relies on single-modality detection and assessment methods. These include monitoring nerve conduction function via electroencephalography (EEG), observing white matter structural changes using functional magnetic resonance imaging (fMRI), or detecting conduction velocity and excitability characteristics through peripheral nerve electrophysiology. These methods have been widely used in clinical practice and research to assess the degree of damage and partial recovery of the nervous system, and to provide data support for rehabilitation training or electrical stimulation interventions.

[0003] However, existing technologies have significant shortcomings. First, data from different modalities are often used in isolation, lacking cross-modal fusion and unified modeling, which prevents a comprehensive reflection of the dynamic process of myelin damage and repair. Second, individual differences are not adequately considered, often relying on group averages as a reference and lacking dynamic correction of individual baseline states. Third, in terms of rehabilitation intervention, most methods adjust stimulation parameters based on a single indicator, making it difficult to achieve multi-dimensional comprehensive optimization of training task difficulty, rhythm, and electrical stimulation protocols. These problems limit the accuracy and personalization of rehabilitation assistive methods.

[0004] Therefore, a novel method for assisting myelin repair based on neural signal modeling is needed to overcome the shortcomings of existing technologies. Summary of the Invention

[0005] This application provides a method for assisting myelin repair based on neural signal modeling, so as to achieve personalized assistance and dynamic optimization of myelin regeneration and repair.

[0006] This application provides an auxiliary method for myelin repair based on neural signal modeling, including:

[0007] Multimodal neural signal data of the individual to be rehabilitated is acquired and preprocessed to obtain a preprocessed dataset. The multimodal neural signal data includes electroencephalogram (EEG) signals, functional magnetic resonance imaging (fMRI) signals, and peripheral nerve electrophysiological signals.

[0008] Based on the preprocessed dataset, an individualized temporal modeling process is performed to generate a temporal feature representation characterizing the dynamic process of myelin sheath injury-repair. The temporal feature representation includes at least the neural conduction velocity recovery rate index, the phase consistency index, and the neural pathway integrity index.

[0009] Individualized baseline parameters are calculated using a dynamic baseline update mechanism, and the time series feature representation is calibrated using the individualized baseline parameters to obtain a calibrated time series feature representation.

[0010] The calibration temporal feature representation is used as input and fed into a pre-trained deep learning inference engine containing recurrent units, graph convolutional units, and transfer learning modules. The deep learning inference engine outputs a myelin repair index representing the degree of myelin repair and an imbalance risk score representing the probability of future repair abnormalities. During the training phase of the deep learning inference engine, the prior constraint parameters generated by the individualized temporal modeling process are embedded into the loss construction and regularization of the deep learning inference engine.

[0011] The system receives the myelin repair index and imbalance risk score, and generates a set of intervention parameters, which includes electrical stimulation intervention parameters, virtual rehabilitation training task difficulty, and rehabilitation training rhythm.

[0012] Control commands are generated based on the set of intervention parameters, and the control commands are executed to configure the virtual rehabilitation training environment and the electrical stimulation execution interface.

[0013] The beneficial effects of the technical solution provided in this application include:

[0014] (1) By jointly modeling EEG signals, functional magnetic resonance signals and peripheral nerve electrophysiological signals and generating temporal feature representations, the dynamic process of myelin damage and repair can be comprehensively depicted, which improves the completeness and accuracy of the assessment compared with single signal analysis. (2) By calculating individualized baseline parameters through a dynamic baseline update mechanism and calibrating the feature representations, individual differences between different patients can be effectively eliminated, and the personal credibility of the repair index and risk score can be improved, thereby avoiding the insufficient applicability caused by the reliance on the population mean in traditional methods. (3) By combining a deep learning inference engine with recurrent units, graph convolutional units and transfer learning modules, and introducing prior constraint parameters in the training phase, the myelin repair index and imbalance risk score can be output, which can not only quantify the degree of repair, but also predict future abnormal risks, providing a prospective reference for clinical intervention. (4) By receiving the repair index and risk score, an intervention parameter set containing electrical stimulation parameters, training task difficulty and training rhythm can be generated, and control instructions can be executed to configure the virtual rehabilitation training environment and electrical stimulation interface, which can realize the comprehensive optimization of rehabilitation training and electrical stimulation, and improve the efficiency and effect of myelin regeneration and repair. Attached Figure Description

[0015] Figure 1 This is a flowchart of an auxiliary method for myelin repair based on neural signal modeling provided in the first embodiment of this application. Detailed Implementation

[0016] Many specific details are set forth in the following description to provide a full understanding of this application. However, this application can be implemented in many other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of this application; therefore, this application is not limited to the specific embodiments disclosed below.

[0017] The first embodiment of this application provides an auxiliary method for myelin repair based on neural signal modeling. Please refer to... Figure 1 This figure is a schematic diagram of the first embodiment of this application. The following is in conjunction with... Figure 1 The first embodiment of this application provides a detailed description of an auxiliary method for myelin repair based on neural signal modeling.

[0018] Step S101: Obtain multimodal neural signal data of the individual to be rehabilitated and preprocess it to obtain a preprocessed dataset. The multimodal neural signal data includes electroencephalogram (EEG) signals, functional magnetic resonance imaging (fMRI) signals, and peripheral nerve electrophysiological signals.

[0019] In step S101, it is necessary to acquire and preprocess the multimodal neural signals of the individual to be rehabilitated. This process is not merely a simple data acquisition, but rather ensures strict temporal alignment and spatial correspondence between different modalities to guarantee the reliability of subsequent modeling. Firstly, regarding the acquisition of EEG signals, a multi-channel EEG acquisition device conforming to medical standards should be used. Electrical activity signals from different regions of the cerebral cortex should be recorded in real time using scalp electrodes. The sampling rate, filtering parameters, and electrode impedance should be precisely controlled to ensure the signal covers the main frequency band from 0.5Hz to 70Hz, while avoiding power line interference and high-frequency noise. For the acquisition of functional magnetic resonance imaging (fMRI) signals, functional sequence scanning needs to be performed on standardized MRI equipment. Typically, BOLD (blood oxygen level dependent) signals are used to acquire changes in brain region activity, ensuring synchronization with EEG signal acquisition within the same scanning time window. If necessary, temporal correction can be performed using triggers or timestamps to ensure that data from different modalities can be processed uniformly in the subsequent stages. The acquisition of peripheral nerve electrophysiological signals is mainly accomplished through electromyography electrodes or nerve conduction velocity testing equipment. By placing surface electrodes or needle electrodes in the target nerve innervation area, action potentials and compound electromyographic responses are induced and recorded, thereby reflecting the conduction velocity and excitability characteristics of peripheral nerves.

[0020] After completing multimodal data acquisition, preprocessing is a crucial step to ensure data quality and analytical accuracy. For EEG signals, bandpass filtering is required to eliminate low-frequency drift and high-frequency noise, while independent component analysis is used to remove artifacts related to blinking, ECG, and EMG. For functional magnetic resonance imaging (fMRI) signals, slice time correction, motion correction, spatial normalization, and smoothing are necessary to ensure comparability of data from different time points and individuals. For peripheral neurophysiological signals, baseline drift correction, signal segmentation, and artifact removal are required to ensure accurate identification of the onset and peak values ​​of action potentials. After preprocessing of all modalities, a unified timestamp alignment mechanism should be used to map EEG, fMRI, and electrophysiological signals onto the same timeline, forming time-consistent multimodal data.

[0021] In this process, a standardized data structure for multimodal signals can be established, such as representing them as three-dimensional tensors, with different dimensions corresponding to modality type, time series, and signal intensity, respectively. This allows subsequent individualized time series modeling to directly call upon structured data input without repeated format conversion. Simultaneously, a quality control module should be added during the preprocessing stage to mark or remove data with excessively low signal intensity, high artifact ratios, or too many missing segments to avoid affecting the accuracy of the modeling results. Ultimately, the preprocessed dataset obtained through the above steps should possess characteristics of minimized noise, temporal alignment, and structural uniformity, comprehensively and accurately reflecting the neural state of the individual undergoing rehabilitation at the three levels of cerebral cortex, deep neural activity, and peripheral conduction function, thus laying a reliable data foundation for subsequent time series modeling and myelin repair analysis.

[0022] Furthermore, the acquisition and preprocessing of multimodal neural signal data of the individual to be rehabilitated to obtain a preprocessed dataset includes:

[0023] The artifact removal process takes the original EEG signal, functional magnetic resonance imaging (fMRI) signal, and peripheral nerve electrophysiological signal as inputs. It uses a pre-set artifact template library to perform time-window matching and removal of eye movement artifacts, electrocardiogram artifacts, electromyography artifacts, and power frequency interference on the EEG signal. It performs slice time correction, head movement trajectory correction, and spatial normalization on the fMRI signal. It performs baseline drift correction and stimulus label error correction on the peripheral nerve electrophysiological signal. It outputs a clean multimodal signal set and generates data quality labels for each time window.

[0024] Based on the clean multimodal signal set, a cross-modal time synchronization process is performed, aligning EEG events, MRI volume sampling and electrophysiological stimulation response events according to hardware trigger signals or system timestamps, outputting a time synchronization signal set and assigning an acquisition time marker to each time point;

[0025] Using the time synchronization signal set as input, a feature standardization and mutual information enhancement process is performed. The amplitude of each modality is standardized and the noise energy ratio is corrected according to the value range recorded during the training phase. Within the sliding time window, an alignment confidence score is generated by maximizing the deterministic mutual information metric between each modality. The standardized and aligned result is output as a feature-enhanced aligned signal set, and the alignment confidence score is incorporated into the data quality label of the corresponding time window.

[0026] Based on the feature-enhanced aligned signal set, a quality-weighted combination process is performed. Dynamic modal weights are assigned to EEG signals, functional magnetic resonance signals, and peripheral neurophysiological signals according to data quality labels and alignment confidence scores. Low-quality time windows are masked and bounded interpolation is performed on adjacent time windows to output a quality-weighted signal set.

[0027] The quality-weighted signal set is converted into a unified data structure, the order and dimension of the multimodal feature components are fixed, and the acquisition time marker and data quality marker are encapsulated together to form a time-ordered feature vector sequence as the preprocessed dataset.

[0028] In practice, the first step is to remove artifacts from the acquired raw EEG, functional magnetic resonance imaging (fMRI), and peripheral neurophysiological signals. This artifact removal process requires targeted interventions for each modality. For EEG signals, a pre-defined artifact template library is used, and the recorded signals are compared with characteristic patterns of eye movement artifacts, electrocardiogram (ECG) artifacts, electromyography (EMG) artifacts, and power line interference within each time window. When a segment highly overlaps with the template features, the system automatically removes it or replaces it with interpolation, thus preserving the cleanest possible neural activity components. For fMRI signals, slice time correction must be performed after 3D image acquisition to ensure that time delays between multiple slices are compensated; then, head movement trajectory correction is used to eliminate signal deviations caused by slight head movements; finally, spatial normalization is performed, resampling all images to a unified coordinate system for subsequent alignment with other modalities. For peripheral neurophysiological signals, slowly changing baseline drift needs to be removed, and trigger markers recorded in the electrical stimulation experiment need to be corrected to ensure temporal consistency between stimulation and response. After this stage, the system will output a clean multimodal signal set and attach a data quality tag to each time window, which records the validity score of the data in that window after artifact removal.

[0029] After obtaining a clean multimodal signal set, the system enters the cross-modal time synchronization phase. The goal of this phase is to align the time references of different modalities, ensuring that EEG events, functional magnetic resonance imaging (fMRI) volumetric sampling points, and peripheral electrophysiological stimulation and response events are all located within the same time coordinate system. Synchronization relies on hardware trigger signals or system timestamps generated during acquisition. The system uses these triggers or timestamps as anchor points to align events across the three modalities, ultimately outputting a time-synchronized signal set. Each time point in this signal set not only contains multimodal data but also clearly marks a unified acquisition time stamp, ensuring that cross-modal features can be accessed under the same time reference during subsequent processing.

[0030] After time alignment, the system takes the time-synchronized signal set as input and further performs feature standardization and mutual information enhancement. In this stage, the amplitude of data from each modality is standardized according to the feature value range recorded during the training phase. For example, the amplitude of EEG is standardized to a comparable range in the millivolt range, the signal intensity of functional magnetic resonance imaging (fMRI) is standardized to a normalized percentage, and electrophysiological signals are standardized to the milliampere or millisecond range. Simultaneously, the noise energy ratio is corrected based on the noise level of each modality to ensure a relatively balanced signal-to-noise ratio across different modalities. Based on this, the system calculates the mutual information value between different modalities within a sliding time window and generates an alignment confidence score while maximizing mutual information. This means that if the EEG signal and fMRI signal show a strong correspondence in neural events within a certain time window, the mutual information value is higher, and the corresponding time window will be assigned a higher alignment confidence score. Finally, the output of this stage is a feature-enhanced aligned signal set, and this confidence score is appended to the data quality label, ensuring that the data is not only standardized but also receives a measure of cross-modal consistency.

[0031] The next step is the quality-weighted combination process, which takes the feature-enhanced aligned signal set as input. The core of this process is to assign dynamic weights to the three modalities based on the data quality labels and alignment confidence scores generated in the previous step. Data with higher weights contribute more to the result during the combination process, while low-quality time windows are either masked or smoothly compensated for by adjacent high-quality time windows through bounded interpolation. In this way, even if the quality of EEG signals degrades due to transient interference during certain time periods, the system can still maintain the integrity and stability of the data by relying on functional magnetic resonance imaging (fMRI) and peripheral electrophysiological signals. This step ultimately outputs a quality-weighted signal set in which the influence of each modality is dynamically adjusted to an optimal combination, ensuring that the result retains the complementarity of multimodal information while avoiding interference from low-quality data.

[0032] Finally, the system transforms the quality-weighted signal set into a unified data structure. During this process, the system fixes the order and dimensions of the multimodal feature components; for example, it arranges indicators such as neural conduction velocity, phase consistency, and pathway integrity in fixed positions, ensuring that subsequent algorithms can directly read them according to the agreed-upon indexes. Acquisition time stamps and data quality markers are also encapsulated along with the data, forming a time-ordered sequence of feature vectors. This sequence is the final form of the preprocessed dataset. It not only contains clean versions of the multimodal neural signals but has also undergone alignment, standardization, quality weighting, and unified encapsulation, enabling it to be directly used in subsequent individualized temporal modeling processes without additional corrections.

[0033] Through the aforementioned continuous artifact removal, time synchronization, feature enhancement and normalization, quality weighting, and data structuring processes, the preprocessed dataset achieves cross-modal unification while fully preserving the temporal and structural characteristics of neural signals.

[0034] Step S102: Perform an individualized temporal modeling process based on the preprocessed dataset to generate a temporal feature representation that characterizes the dynamic process of myelin sheath injury-repair. The temporal feature representation includes at least the neural conduction velocity recovery rate index, the phase consistency index, and the neural pathway integrity index.

[0035] In step S102, an individualized temporal modeling process needs to be performed based on the preprocessed dataset obtained in the previous step to accurately depict the dynamic changes in myelin sheath injury and repair. This process first requires sequentially arranging EEG signals, functional magnetic resonance imaging (fMRI) signals, and peripheral nerve electrophysiological signals according to the time dimension, ensuring that different modalities can form corresponding relationships on the same time axis, thus preventing temporal misalignment between modalities during subsequent modeling. Subsequently, feature extraction algorithms are used to decompose and analyze the data for each time period. For EEG signals, power spectral density estimation and wavelet transform can be used to extract the energy distribution and phase relationship of neural oscillations at different frequency bands. For fMRI signals, temporal signals from regions of interest (ROIs) can be used to extract local activity intensity changes, and diffusion tensor imaging parameters can be combined to infer the integrity of white matter fiber tracts and the functional connectivity between different regions. For peripheral nerve electrophysiological signals, the onset latency, peak time, and amplitude changes of nerve conduction need to be calculated to reflect the nerve impulse transmission speed and pathway conduction efficiency.

[0036] After initial feature extraction, the individualized temporal modeling process requires organizing these features according to a unified data structure, typically using a multidimensional tensor representation. One dimension corresponds to the modality category, another to a time point, and the third to the specific feature value. This approach ensures that the modeling process can simultaneously access feature information from different modalities while maintaining continuity over time. During model construction, suitable modeling tools for temporal data processing must be selected, such as state transition frameworks based on Hidden Markov Models or hybrid structures based on temporal convolutions and recurrent neural networks, to capture the dynamic dependencies between different time segments. To enhance individual variability, the model also needs to incorporate reference baselines relevant to that individual, such as neural signal characteristics in the pre-rehabilitation phase or the contralateral undamaged area. This reference information can be embedded as conditional variables into the model structure, ensuring that the generated feature representation not only reflects the general laws of injury repair but also highlights the individual's specific responses.

[0037] Through the above modeling, the final step is to generate a temporal feature representation that can characterize the myelin sheath injury and repair process. This feature representation is not a single indicator, but a dynamic vector set containing multiple dimensions. This set should at least include a neural conduction velocity recovery rate indicator, which quantifies repair progress by comparing the increase in neural conduction velocity at different time periods; a phase consistency indicator, which reflects the synchronization recovery of the neural network by analyzing the phase locking degree of EEG signals at different frequency bands; and a neural pathway integrity indicator, which infers the changes in the integrity of neural fiber pathways during the injury repair process by jointly using magnetic resonance diffusion parameters and electrophysiological connectivity. In practice, these indicators need to be arranged in chronological order to form a continuous temporal curve or matrix so that they can be input into the deep learning inference engine for further calculation and prediction.

[0038] The core of step S102 lies in leveraging the temporal consistency and feature complementarity of multimodal signals, combined with an individualized reference baseline, to establish a temporal feature representation that can dynamically describe the myelin repair process. This process not only requires multi-level extraction and fusion of data, but also the construction of a temporal modeling framework that conforms to individual characteristics, thereby ensuring that the output feature representation has both cross-modal comprehensiveness and individual-specific accuracy.

[0039] The following details the specific calculation process of the three indicators. In the specific implementation of step S102, the nerve conduction velocity recovery rate indicator first needs to be obtained from the preprocessed dataset. In the peripheral nerve electrophysiological signal part, the preprocessing step has completed noise removal and signal baseline correction, so that the starting point of the stimulation signal and the response waveform of the muscle compound potential can be clearly identified. In the pre-rehabilitation measurement, a long latency time can usually be observed between the application of electrical stimulation and the appearance of muscle response. For example, the muscle activity peak appears only after a long delay after the stimulation signal is emitted. Under the same measurement conditions after rehabilitation, the response appears earlier and the latency time is shortened. In order to quantify the recovery of nerve conduction velocity, data needs to be collected multiple times at different rehabilitation stages, and the average latency time is measured for each stage. Then, the latency time before rehabilitation is compared with the latency time after rehabilitation, and the shortening ratio relative to the initial state is calculated. This ratio reflects the degree of improvement in the conduction velocity of nerve impulses along nerve fibers, thus forming the nerve conduction velocity recovery rate indicator, which is recorded in chronological order.

[0040] Secondly, the phase consistency index is calculated. In the EEG signal section, preprocessed data has removed artifacts such as eye movement and ECG, resulting in clear signals suitable for time-series and frequency band analysis. For specific motion-related frequency bands, such as low-frequency rhythmic signals, the phases of multiple electrode sites at the same time can be compared. If the phase values ​​of different electrodes differ significantly, it indicates a lack of synchronicity among neuronal populations; conversely, if the phase values ​​of different electrodes gradually converge in post-rehabilitation data, it indicates enhanced coordinated activity between brain regions. To obtain a quantitative index, the concentration of phase differences between different electrodes can be statistically analyzed within a fixed time window. If the phase differences are dispersed, the consistency is low; if the phase differences are concentrated within a certain range, the consistency is high. By comparing the consistency distribution at different rehabilitation stages, the trend of the phase consistency index over time can be obtained. This index directly reflects the recovery process of neural network functional synchronicity.

[0041] Finally, the neural pathway integrity index is calculated. In the functional magnetic resonance imaging (fMRI) and diffusion imaging sections, preprocessed image data has eliminated motion artifacts and unified the spatial coordinate system, ensuring direct comparison of data from different time points. In the early stages of rehabilitation, reconstructed white matter fiber bundles often exhibit breaks, discontinuous fiber orientation, and a low fiber count. Image processing software can identify the number, average length, and connectivity of fiber bundles. For example, it can be determined whether fibers in a certain region can completely penetrate the expected origin and end points; if a large number of fibers are broken or fail to reach the target area, the integrity is low. In the preprocessed data from the later stages of rehabilitation, fiber bundles appear more continuous, their number significantly increases, and fiber connectivity is markedly improved. By comparing the improvement in fiber number, length, and connectivity at different time points, a neural pathway integrity index can be generated and presented as a time-varying sequence reflecting the progress of fiber structure repair.

[0042] Through the three specific processes described above, it can be seen that the calculation of all indicators is based on the preprocessed dataset: the neural conduction velocity recovery rate reflects the recovery of conduction velocity through the proportion of latency shortening; the phase consistency indicator reflects the enhanced synchronicity through the concentration of the phase distribution of EEG signals between electrodes; and the neural pathway integrity indicator reflects the degree of neural structural recovery through the improvement of fiber number, length, and connectivity. The values ​​of these indicators gradually improve over time and form a continuous curve in the time dimension, thus constituting a complete temporal feature representation and providing reliable input for subsequent deep learning inference.

[0043] Furthermore, the individualized temporal modeling process performed based on the preprocessed dataset generates a temporal feature representation characterizing the dynamic process of myelin sheath injury-repair. This temporal feature representation includes at least neural conduction velocity recovery rate indices, phase consistency indices, and neural pathway integrity indices, including:

[0044] The preprocessed dataset is divided into multi-scale time segments. The EEG signals, functional magnetic resonance signals and peripheral nerve electrophysiological signals are segmented into short-time segments and long-time segments to obtain a multi-scale time segment sequence with time order.

[0045] Based on the multi-scale time segment sequence, instantaneous phase events in the EEG signal are identified, and local activation changes in the functional magnetic resonance signal are aligned on a unified time axis. At the same time, the stimulus and response latencies in the peripheral nerve electrophysiological signal are mapped to the alignment result to obtain cross-modal time correspondence.

[0046] Based on the cross-modal time correspondence, a time-series change trajectory is generated, which reflects the trend of latency shortening or lengthening, the trend of phase concentration or dispersion, and the trend of neural pathway connectivity enhancement or weakening, thus obtaining a trajectory set composed of multiple time-series change trajectories.

[0047] Based on the trajectory set, individualized dynamic features are extracted. The temporal change trajectory of the latency period is converted into a neural conduction velocity recovery rate index, the temporal change trajectory of the phase is converted into a phase consistency index, and the temporal change trajectory of the pathway connectivity is converted into a neural pathway integrity index. A feature vector sequence composed of the three indices is output. The feature vector sequence is arranged in chronological order to form a temporal feature representation.

[0048] In practical implementation, the preprocessed dataset first needs to be divided into multi-scale time segments. Multi-scale time segmentation refers to the system dividing the data into short and long segments according to a set duration, maintaining a nested or overlapping relationship between the two types of segments. For example, short segments can cover a time range of tens to hundreds of milliseconds to capture rapid fluctuations in neural events, while long segments cover a range of several seconds or even tens of seconds to reflect macroscopic trends and chronic changes in functional areas. During the segmentation process, the system must attach a uniform timestamp to each segment to ensure that subsequent processing steps are aligned correctly. This process yields a multi-scale time segment sequence with a clear temporal order, containing complete information from EEG signals, functional magnetic resonance imaging (fMRI) signals, and peripheral neurophysiological signals.

[0049] After obtaining multi-scale time-segment sequences, the system enters the cross-modal alignment phase. At this stage, the system identifies instantaneous phase events in the EEG signals. These events typically manifest as synchronous or inverse changes between different electrodes at a given moment, and their location can be precisely determined through phase difference calculations. Subsequently, the system aligns local activation changes in the functional magnetic resonance imaging (fMRI) signals on a unified time axis. For example, a significant increase or decrease in voxel signal intensity in a brain region during a time segment is identified and correlated with the aforementioned instantaneous phase events. Simultaneously, the stimulus and response latencies recorded in peripheral neurophysiological signals are also corrected and mapped to the same time point. The result of this series of operations is the cross-modal temporal correspondence, which integrates three previously independent signal modalities onto the same time coordinate, providing a foundation for subsequent trajectory generation.

[0050] After establishing cross-modal temporal correspondences, the system generates temporal change trajectories. These trajectories are not abstract concepts, but rather continuous curves explicitly constructed based on the temporal trends of each modality. For latency, the system records the relative changes in latency within each short time segment and connects these points in chronological order to form a trajectory reflecting the trend of latency shortening or lengthening. For phase, the system calculates the phase concentration among multiple electrodes within each time segment and generates a trajectory reflecting the trend of phase concentration or dispersion. For pathway integrity, the system utilizes connectivity information provided by both functional magnetic resonance imaging (fMRI) and electrophysiological signals to calculate the effectiveness and stability of neural pathways and generates a trajectory reflecting the trend of increased or decreased pathway connectivity. These three types of trajectories together constitute a trajectory set, each trajectory being a result with explicit numerical values ​​and time stamps.

[0051] After generating the trajectory set, the system enters the feature extraction stage. This stage requires converting the trajectory set into indicators that can be directly used by subsequent steps. The temporal variation trajectory of latency is parsed into a neural conduction velocity recovery rate indicator, which can be obtained by calculating the magnitude and frequency of latency shortening and can quantitatively reflect the recovery level of neural conduction function. The temporal variation trajectory of phase is parsed into a phase consistency indicator, which indicates whether the phase concentration between electrodes continues to increase within a certain time range and can reflect the synchronization trend of the neural network. The temporal variation trajectory of pathway connectivity is parsed into a neural pathway integrity indicator, which reflects whether the pathway shows stable enhancement or maintenance in the time dimension, avoiding the limitations of single-point-of-time observation. These three types of indicators are uniformly organized into a feature vector sequence, arranged in chronological order, to form a complete temporal feature representation.

[0052] Through the above steps, the original complex multimodal signal is transformed layer by layer. First, the temporal resolution is obtained by dividing the time segments into multiple scales. Then, the unification between different modes is completed by establishing cross-modal time correspondence. Next, the visualization of the continuous dynamic process is obtained by generating the temporal change trajectory. Finally, these dynamic processes are transformed into quantitative indicators through individualized feature extraction.

[0053] Step S103: Calculate individualized baseline parameters through a dynamic baseline update mechanism, and calibrate the time series feature representation using the individualized baseline parameters to obtain a calibrated time series feature representation.

[0054] In step S103, the core task is to calculate individualized baseline parameters through a dynamic baseline update mechanism and use these parameters to calibrate the temporal feature representation, thereby obtaining a calibrated temporal feature representation that better reflects the individual's actual situation. Here, "dynamic baseline update" does not statically use an initial value as a long-term reference, but rather continuously corrects and updates it as the rehabilitation process progresses and data accumulates, enabling the entire system to make judgments based on the latest state at any given time.

[0055] In practice, the first step is to extract reference values ​​for multimodal signals in the early stages of rehabilitation from the preprocessed dataset. For example, in the peripheral neurophysiology section, the average latency at the initial rehabilitation test can be used as the initial baseline for nerve conduction velocity; in the EEG section, the phase difference distribution between different electrodes in the early rehabilitation period can be used as the initial reference range for phase consistency; and in the functional magnetic resonance diffusion imaging section, the initial number and connectivity of fiber bundles can be used as the reference level for pathway integrity. These initial values ​​constitute the earliest individual baseline parameter set.

[0056] As rehabilitation training progresses, new data is continuously collected and enters the preprocessing stage. Under the dynamic baseline update mechanism, each time new data arrives, the system compares the newly obtained neural conduction velocity, phase coherence, and pathway integrity values ​​with historical records. If the new data consistently shows a level better than the initial baseline, and this trend remains consistent within a certain time window, then the original baseline parameters will be updated to the new values. For example, if a patient's average latency was 20 milliseconds at the initial test, and after four weeks of rehabilitation, the average latency from multiple collections stabilizes at around 15 milliseconds with very little fluctuation, the system will update the baseline parameter from 20 milliseconds to 15 milliseconds. The purpose of this is to ensure that subsequent repair rate calculations are not always compared to the initial severe injury state, but rather to the patient's current typical level, making the results more realistically reflect the subtle progress during the rehabilitation process.

[0057] In the EEG phase consistency section, if the phase difference between different electrodes is widely dispersed in the initial state, the system records this dispersion as the initial baseline. During the mid-rehabilitation phase, when newly acquired data shows a significantly higher phase concentration, the system gradually corrects the baseline using a dynamic update mechanism, raising the baseline level to a new concentration range. In this way, even if the patient's progress is very minor in the later stages, these subtle changes can be detected and quantified at the new baseline, preventing them from being masked by large differences in the early stages.

[0058] In functional magnetic resonance imaging (fMRI) and diffusion imaging, the initial baseline may show a low number of fiber bundles and numerous breaks. After rehabilitation training, if newly acquired data consistently shows an increase in fiber number and improved continuity across multiple measurements, the system will adjust the baseline parameters to the new fiber number and connectivity levels. This dynamic baseline update ensures that the model, when processing subsequent data, does not overly rely on early severe injury states but always references the patient's latest rehabilitation progress.

[0059] After baseline updates are completed, these updated individualized baseline parameters need to be applied to the calibration of temporal feature representations. Calibration involves comparing each newly obtained feature value with the current individual baseline, rather than simply with the population mean. For example, the calculation of neural conduction velocity recovery rate no longer relies solely on differences from healthy individuals, but directly uses the patient's baseline latency as a reference, thus reflecting the extent of improvement relative to their own baseline during rehabilitation. Similarly, phase consistency and pathway integrity indices are also corrected to better reflect the dynamic changes of individual patients, rather than being fixed to an immutable standard.

[0060] To achieve this calibration, after generating new feature values, the system directly compares these values ​​with the current baseline and generates standardized calibration features based on the magnitude and direction of the difference. For example, when the latest neural conduction latency is shorter than the current baseline, the system converts this improvement into a positive value and stores it in the corresponding neural conduction velocity recovery rate vector component; if the latency is prolonged, it is converted into a negative value to reflect functional degradation. In the phase consistency section, the system compares the concentration of the current phase difference between electrodes with the baseline concentration and converts the difference into a continuous value. A higher value indicates better consistency, and a lower value indicates decreased consistency. This result is stored as a vector component of the phase consistency index. In the pathway integrity section, the system compares the current fiber number, length, and connectivity with the baseline values ​​one by one, integrating the results into a standardized value that reflects both the degree of pathway enhancement and the magnitude of degradation, and is stored as a vector component of the neural pathway integrity index. All these calibrated feature values ​​are packaged together with the acquisition timestamp and data quality marker to form a structured feature vector. Over time, each newly collected data undergoes the same calibration and packaging process, ultimately forming a sequence of feature vectors arranged in chronological order. This sequence is the calibrated temporal feature representation, which can completely and continuously record the dynamic changes of patients relative to their individual baselines during the rehabilitation process, providing a unified and standardized input for subsequent deep learning inference engines.

[0061] The temporal feature representation obtained after this dynamic baseline update and calibration process not only retains the damage and repair trends revealed by the original features, but also enhances the sensitivity to subtle progress through continuous correction of the individualized baseline. The calibrated temporal feature representation generated in this way is more stable and comparable than the uncalibrated data, and can reflect the gradually changing reality during the rehabilitation process, thus providing a more reliable data foundation for the input of the subsequent deep learning inference engine.

[0062] Furthermore, the step of calculating individualized baseline parameters through a dynamic baseline update mechanism, and calibrating the time-series feature representation using the individualized baseline parameters to obtain a calibrated time-series feature representation, includes:

[0063] Baseline segments are extracted from the temporal feature representation. Stable segments of neural conduction velocity recovery rate index, phase consistency index and neural pathway integrity index are identified within a preset time window, and a set of baseline segments is output.

[0064] Using the baseline segment set as input, abnormal segment removal is performed. Based on statistical thresholds and data quality markers, segments with instantaneous spikes, significant deviations, or low-quality markers are removed, and a valid baseline segment set is output.

[0065] Using the set of effective baseline segments as input, weighted time series fusion is performed, assigning higher weights to baseline segments in recent time windows and lower weights to baseline segments in distant time windows, and outputting dynamically updated individualized baseline parameters.

[0066] Using the individualized baseline parameters as input, the temporal feature representation is calibrated point by point. The neural conduction velocity recovery rate index is adjusted to the velocity recovery difference relative to the individualized baseline parameters, the phase consistency index is adjusted to the consistency difference relative to the individualized baseline parameters, and the neural pathway integrity index is adjusted to the integrity difference relative to the individualized baseline parameters. The calibrated temporal feature representation is then output.

[0067] In practical implementation, the first step is to identify baseline segments from the temporal feature representation. Baseline segment extraction is performed in units of preset time windows. Within each time window, the system performs stability analysis on neural conduction velocity recovery rate, phase consistency, and neural pathway integrity indicators. If, within that window, the fluctuation range of these indicators is below a preset threshold and there is no obvious upward or downward trend, they are considered stable segments. These stable segments are collected to form a set of baseline segments with clear time stamps, serving as the basis for subsequent calculations.

[0068] After forming the baseline fragment set, further outlier removal is required. Outliers typically manifest as sudden spikes, significant deviations from long-term trends, or fragments already marked as low-quality during preprocessing. The system performs statistical analysis on each baseline fragment, calculating its mean, variance, and consistency with adjacent fragments. If a fragment's metrics exceed a set statistical threshold, or if a quality flag indicates severe noise interference, the fragment is removed. This process outputs a cleaned, valid baseline fragment set, eliminating outliers and low-quality interference, ensuring the reliability of subsequent calculations.

[0069] After obtaining a valid set of baseline segments, the system performs weighted temporal fusion. The purpose of this process is to generate dynamically updated individualized baseline parameters. Specifically, the system assigns different weights to different baseline segments based on their temporal distance. More recent baseline segments are given higher weights because they better reflect the individual's latest recovery status, while older baseline segments are given lower weights to maintain the continuity of the overall trend. This weighting is not a one-time event but is continuously updated over time, ensuring that new observations are gradually incorporated into the baseline calculation. The individualized baseline parameters generated in this way are not fixed values ​​but rather a dynamically changing reference point that evolves with the recovery process.

[0070] After obtaining dynamically updated individualized baseline parameters, the system applies them to point-by-point calibration, directly impacting the entire temporal feature representation. For the neural conduction velocity recovery rate index, the system calculates the difference between the velocity recovery rate and the individualized baseline parameters at each time point, thus obtaining the velocity recovery difference, reflecting the extent of improvement or degradation of neural conduction compared to its baseline. For the phase consistency index, the system compares the consistency level at each time point with the individualized baseline parameters, obtaining the consistency difference, which measures the relative change in phase concentration. For the neural pathway integrity index, the system compares the connectivity value at each time point with the individualized baseline parameters, obtaining the integrity difference, revealing whether the pathway is enhanced or weakened compared to the baseline. After this point-by-point calibration, all feature data are no longer absolute values, but rather, with the individual's own dynamic baseline as a reference, form a calibrated temporal feature representation.

[0071] Through this complete dynamic baseline update and calibration process, the preprocessed temporal feature representation is further personalized and stabilized, enabling it to more accurately reflect the progress of myelin repair in subsequent deep learning inference engines.

[0072] Step S104: The calibrated temporal feature representation is fed into a pre-trained deep learning inference engine containing recurrent units, graph convolutional units, and transfer learning modules; the deep learning inference engine outputs a myelin repair index representing the degree of myelin repair and an imbalance risk score representing the probability of future repair abnormalities; wherein, during the training phase of the deep learning inference engine, the prior constraint parameters generated by the individualized temporal modeling process are embedded into the loss construction and regularization of the deep learning inference engine.

[0073] In step S104, the input to the deep learning inference engine is the calibrated temporal feature representation generated in step S103. This representation uses a sequence of feature vectors arranged in chronological order as its basic unit. Each vector at each time point contains at least three dimensions: neural conduction velocity recovery rate, phase consistency, and neural pathway integrity, along with acquisition time stamps and data quality markers. To adapt the data to the inference engine, the input adaptation layer first segments the original sequence into fixed-length time slices, such as every five seconds or every hundred sampling points. For segments marked as low quality or missing, the system removes them based on the data quality markers or repairs them using interpolation methods, and normalizes all features to a predefined numerical range during the training phase. This processing results in a consistent and stable input batch.

[0074] The recurrent unit (ROU) receives each time slice from the batch, with the input being the feature vector of each time step and the output being a sequence of temporal latent representations corresponding to each time step. The RRU can employ a Long Short-Term Memory (LSTM) network or a gated recurrent unit (GRU), and its function is to capture short-term fluctuations and long-term trends of features over time. The internal state of this unit is continuously passed within a time slice, but reset between time slices to avoid cross-segment interference. To stabilize the differences between individuals at the temporal scale, before entering the first step of computation, the RRU adjusts the time constant and the forget gate threshold based on the individualized baseline parameters generated in step S103. This adjustment does not change the unit's training parameters, but rather alters the update speed of the input flowing through the unit through additional scaling rules, thereby ensuring that the output temporal latent representation remains sensitive and stable to individual rhythm differences.

[0075] The input to the graph convolutional unit consists of two parts: a temporal implicit representation generated by the recurrent unit at each time step, and graph structure information describing the relationships between various indicators. The nodes in this graph include at least three categories: neural conduction velocity recovery rate, phase consistency, and neural pathway integrity. If necessary, they can be further subdivided into multiple nodes to represent features of different brain regions or segments of neural fiber bundles. Edge weights are derived from template relationships established during training, such as the coupling degree between neural conduction velocity and pathway integrity. During inference, the system makes small, bounded adjustments to these edge weights based on the correlation coefficients of adjacent indicators and data quality labels. This adjustment is deterministic fine-tuning and does not involve parameter learning. Through neighborhood aggregation operations, the graph convolutional unit encodes the interaction relationships between different indicators into a new representation, outputting a structure-enhanced representation sequence aligned with the time step.

[0076] The transfer learning module follows the graph convolutional unit. Its input is a sequence of structure-enhanced representations, and its output is a sequence of representations aligned to the individual domain. During the training phase, the module establishes multiple alignment mappings from the source domain to the target individual domain and stores them in a domain adaptation table. Each mapping is recorded as a scaling factor and bias value for the feature channels, forming a parameter table. During the inference phase, the module reads the individualized baseline parameters generated in step S103 and, combined with metadata such as the acquisition device model and data source, retrieves the closest mapping entry from the domain adaptation table. If the individual baseline parameters fall within a certain numerical range, the system selects the mapping table corresponding to that range. Subsequently, the module performs linear correction on each channel of the input representation, that is, first multiplies the channel value by the corresponding scaling factor, and then adds the bias value. The resulting representation distribution is closer to the target individual domain, reducing systematic biases caused by different populations, different acquisition devices, and different scenarios. The entire process does not involve new learning; it only performs deterministic transformations of the stored mappings, ensuring repeatability and field feasibility.

[0077] After domain alignment, the temporal convergence and stability evaluation layer summarizes the representation sequences at the time-slice scale. The input to this layer is the individual domain representation sequence processed by the transfer learning module, and the output includes two types of summaries: one reflecting centrality and asymptotic trends, obtained by weighted averaging of representations at each time step within the time slice, with weights assigned based on data quality labels and the latest baseline; the other reflecting volatility and instantaneous anomaly density, obtained by statistically analyzing differences between adjacent time steps and the number of times the data exceeds the baseline range. All calculations are deterministic statistical operations, requiring no parameter learning.

[0078] The output layer consists of two parallel prediction heads, both receiving the aforementioned set of summary values ​​as input. The first prediction head generates a myelin repair index, calculated primarily based on centrality and progressive trends, while also considering the proportion of components related to neural pathway integrity in the structural enhancement representation. It outputs a quantitative result with dimensions consistent with the baseline parameters in step S103, facilitating comparisons across time slices. The second prediction head generates an imbalance risk score, primarily considering volatility and anomaly density, while also incorporating short-term signs of decline and inconsistencies between indicators. It outputs a quantified risk value for future time windows. During the inference phase, both prediction heads perform boundary pruning and reliability remapping on the results. For example, values ​​exceeding the reasonable neurophysiological range are pruned to the upper or lower bound, and data with low confidence are marked as low-confidence. These rules are defined during the training phase and fixed through prior constraint parameters; they are only executed during the inference phase and are not updated.

[0079] The execution sequence of step S104 is as follows: First, the calibrated temporal feature representation from step S103 is received and divided into normalized time slices through the input adaptation layer. Then, the recurrent unit processes these time slices stepwise to generate temporal latent representations. Next, the graph convolutional unit performs graph structure enhancement on the temporal latent representations at each time step to generate structure-enhanced representations. Afterward, the transfer learning module performs channel-level correction according to the individual domain selection rules to obtain individual domain representations. Subsequently, the temporal convergence and stability evaluation layer performs statistics on the data within the time slices to obtain the summary quantity. Finally, the two prediction heads of the output layer generate the myelin repair index and the imbalance risk score, respectively, and pass them to the subsequent intervention parameter generation step.

[0080] During the training phase, the deep learning inference engine does not rely solely on labeled data for supervision. Instead, it embeds prior constraint parameters from the individualized temporal modeling process as additional guiding signals into the loss construction and regularization. These prior constraint parameters refer to the physiological patterns and individual characteristics obtained during the individualized temporal modeling process in step S102, such as the reasonable range of variation in neural conduction velocity, the normal concentration range of EEG phase consistency, and the structural continuity threshold of neural pathway integrity. These parameters are used as soft or hard boundaries during training to ensure that the model's learning results not only conform to the training data but also follow neurophysiological rationality.

[0081] Specifically, the loss construct during the training phase is no longer a single prediction error, but consists of two parts. One part is the deviation between the model's predicted value and the true label, used to ensure the model has basic fitting ability. The other part is a regularization term generated based on prior constraint parameters, used to penalize predictions or internal representations that deviate from reasonable physiological ranges. For example, if the predicted curve of the neural conduction velocity recovery rate exceeds the upper and lower limits extracted during the modeling process, the system will impose additional penalties on this deviation; if the phase consistency index exhibits unreasonable and violent oscillations in the time dimension, it will also trigger stability constraints, thereby forcing the model parameters to adjust in a more reasonable direction. In this way, the prior parameters can both constrain the output results and constrain the distribution of the intermediate layer representations, ensuring that the entire inference engine remains on a track that conforms to the logic of the individual rehabilitation process.

[0082] To make this process easier to understand, let's take a detailed example. During the training phase, suppose a patient's neural conduction latency during early rehabilitation is 20 milliseconds. The reasonable recovery range obtained from individualized temporal modeling is a latency reduction of no more than 30% and no less than the typical value of 15 milliseconds for healthy individuals. In this case, the system will convert the constraint "latency recovery of no more than 30% and no less than 15 milliseconds" into a priori parameters. When training the deep learning engine, if the model predicts a latency reduction of more than 30% for the neural conduction velocity recovery rate, or a reduction to less than 15 milliseconds, this prediction will trigger a constraint penalty, increasing the loss value. This forces the model to adjust weights during parameter updates to avoid generating results that do not conform to prior rules. Similarly, if the phase consistency of EEG signals is defined by prior modeling as "gradually increasing but not suddenly decreasing within a certain training period," but the model's prediction shows a significant decrease in a short time slice, the loss function will also increase the corresponding penalty value, enabling the model to gradually learn to maintain a reasonable trend. In this way, prior constraint parameters are explicitly incorporated into the loss construction and regularization, enabling the deep learning inference engine not only to learn the surface patterns of the data but also to maintain conformity with neurophysiological mechanisms.

[0083] In one embodiment, the deep learning inference engine includes an input adaptation layer, recurrent units, graph convolutional units, a transfer learning module, a temporal convergence and stability evaluation layer, and an output layer. The engine first receives calibrated temporal feature representations through the input adaptation layer. This input is not a single data point, but rather a sequence of feature vectors arranged in chronological order. Each time point includes neural conduction velocity recovery rate, phase consistency, and neural pathway integrity indicators, as well as acquisition time stamps and data quality markers. The primary task of the input adaptation layer is to segment these feature vector sequences into standardized time slices of fixed length, suitable for batch processing. During this process, if some time points are incomplete due to device noise or signal loss, the input adaptation layer will remove them based on the data quality markers or interpolate within a reasonable range to ensure the continuity and consistency of the input. Furthermore, the input adaptation layer will standardize all features according to the order and value range recorded during the training phase, keeping their distribution stable and preventing offsets between different batches of input. After processing, the output is a structured, standardized time slice sequence.

[0084] After the standardized time-slice sequence enters the recurrent unit (RCU), each time slice is input to the unit step-by-step. The RCU uses a state transfer mechanism to capture rapid short-term fluctuations and slow long-term trend evolutions of features. Between different time slices, the RCU automatically performs a reset to prevent interference between different segments. To address inherent rhythmic differences between individuals, the RCU adjusts the time constant at a fixed scale during the inference phase, referencing individualized baseline parameters. This adjustment does not change the core parameters of the unit but is achieved by controlling the rate of state updates, making the output more closely match the rhythmic characteristics of the current individual. Finally, the RCU outputs a sequence of temporal implicit representations corresponding one-to-one with each time step; these representations are abstractions of the original features after dynamic temporal modeling.

[0085] The graph convolutional unit takes a temporal implicit representation sequence and a pre-defined graph structure as input. This graph structure explicitly describes the connections and edge weights between neural conduction velocity recovery rate indicators, phase consistency indicators, and neural pathway integrity indicators. In each time slice, the graph convolutional unit performs neighborhood aggregation on the interactions between indicators, thereby achieving cross-indicator and cross-region information exchange. To ensure stability during the inference phase, the unit does not learn new parameters during operation; instead, it only performs bounded fine-tuning of the edge weights in the graph structure based on the coordinated changes of indicators within the current time slice and data quality markers. Through this mechanism, the output structure-enhanced representation sequence not only contains information about the changes in individual indicators but also explicitly encodes the correlations and dependencies between them.

[0086] After obtaining the structure-enhanced representation sequence, the transfer learning module is responsible for eliminating systematic biases introduced by different individuals, devices, or acquisition environments. This module takes the structure-enhanced representation sequence and individual domain selection information as input. The individual domain selection information is determined based on individualized baseline parameters and acquisition environment information. The transfer learning module retrieves alignment maps matching the individual domain selection information from the domain adaptation table. Each alignment map is stored as a scaling factor and bias value for each feature channel. During the inference phase, the transfer learning module sequentially applies scaling and offset to each input feature channel, performing linear recalibration and offset correction. This process is completely deterministic and does not involve new learning, thus ensuring alignment of different data sources within a unified representation space. The output is the individual domain-aligned representation sequence, which reflects the true feature distribution of the current individual better than the structure-enhanced representation.

[0087] The representation sequences aligned to individual domains enter the temporal convergence and stability evaluation layer. In this layer, the system integrates representations from multiple time steps, using time slices as units. Weighted summarization yields the central level and asymptotic trend; the weighting is based on data quality markers and the weight distribution of the most recent baseline, ensuring a larger proportion of high-quality data in the final result. Simultaneously, this layer also calculates the differences between adjacent time steps and records the number of times indicators cross boundaries, thereby deriving the fluctuation amplitude and instantaneous anomaly density. These operations ensure that the model can not only identify overall trends but also accurately capture local instabilities. The final output set of time-slice-level summaries provides the foundation for subsequent decision-making layers.

[0088] In the output layer, a time-slice-level summary set is fed into two parallel output pathways. The first pathway primarily relies on the central level, progressive trend, and the proportion of components related to neural pathway integrity to generate a myelin repair index, which quantitatively reflects the individual's current degree of myelin repair. The second pathway relies on fluctuation amplitude, instantaneous abnormality density, and cross-index inconsistency to generate an imbalance risk score, which predicts whether the repair process may experience abnormalities or regression in the future. During the inference phase, neither pathway updates its parameters. Instead, fixed boundary pruning and reliability remapping ensure that the output values ​​fluctuate within a reasonable neurophysiological range, enhancing the interpretability and stability of the results.

[0089] The training and inference phases of the entire deep learning inference engine are strictly separated. During training, the domain adaptation tables of recurrent units, graph convolutional units, and transfer learning modules, as well as the boundaries and remapping rules of the output layer, are all learned offline by embedding prior constraint parameters generated during individualized temporal modeling into the loss construction and regularization, and then solidified in the system. During inference, the engine operates strictly according to the aforementioned data flow and deterministic rules, without involving any new learning, thus ensuring real-time performance and stability.

[0090] In one specific embodiment, the training and inference phases of the deep learning inference engine are strictly separated. During the training phase, researchers collected a large amount of multimodal neural signal data from rehabilitation patients. After preprocessing, temporal modeling, and dynamic baseline calibration, calibrated temporal feature representations for training were formed. These feature representations not only include neural conduction velocity recovery rate indicators, phase consistency indicators, and neural pathway integrity indicators, but also individual baseline parameter information and quality markers.

[0091] During training, the learning objectives of the recurrent unit, graph convolutional unit, and transfer learning module are not only to minimize prediction error, but also to adhere to the prior constraint parameters output by the individualized temporal modeling process. For example, when the individualized temporal modeling process indicates that the rate of change of the neural conduction velocity recovery rate index during normal recovery should not exceed a certain threshold, this constraint is written into the loss function, preventing the recurrent unit from learning unreasonable high-speed change patterns when updating parameters. Similarly, the graph convolutional unit is also constrained by prior constraint parameters when aggregating the relationships between different indices. For instance, under normal circumstances, the correlation between phase consistency and neural pathway integrity should remain within a certain range. If it deviates from this range during training, the loss function will increase the penalty, thereby forcing the model to learn a weight distribution that conforms to physiological laws. The domain adaptation table of the transfer learning module is established during training. It learns a fixed set of scaling factors and bias values ​​based on the differences between different individuals and different acquisition devices. These parameters are stored in the form of a mapping table to ensure that subsequent data from different sources can be aligned in the same feature space.

[0092] In the output layer, prior constraints are embedded through regularization during the training phase to prevent the myelin repair index and imbalance risk score from exceeding physiological ranges. For example, if a patient's indicators fluctuate greatly in the short term, but according to prior constraints, this fluctuation is still within a recoverable range, the output layer will limit the output to a reasonable range through boundary pruning and remapping rules to avoid incorrectly giving extremely high risk scores. All these rules are solidified after training, forming a fixed domain adaptation table, boundary range, and remapping function.

[0093] When entering the inference phase, the deep learning inference engine no longer updates any parameters, but instead operates strictly according to the parameters and rules fixed during the training phase. Taking a new patient as an example, when its calibrated temporal feature representation is input into the system, the recurrent unit directly uses the already trained internal parameters for state updates, but no longer adjusts these parameters; the graph convolutional unit only fine-tunes the edge weights according to the time slice co-variation within the allowed boundaries, without learning new connection methods; the transfer learning module selects the most suitable scaling factor and bias value from the domain adaptation table saved during the training phase and performs fixed alignment transformations; the output layer maps the intermediate results to a repair index and risk score that conform to neurophysiological rationality, according to the boundaries and remapping rules fixed during training.

[0094] To illustrate with a detailed example: During the training phase, the system learns that the neural conduction velocity recovery rate of a certain type of patient typically increases gradually from the baseline to 120% of the baseline within a 30-day recovery period, while the phase consistency index typically increases by no more than 30%. These two ranges are then fixed as prior constraint parameters and incorporated into the loss function. After training, during the inference phase, if new patient input data shows a neural conduction velocity recovery rate that surges to 150% of the baseline within three days, the system automatically determines that this output is irregular based on the rules fixed during training. The output layer adjusts the result to a reasonable range through boundary pruning and remapping, and simultaneously increases the weight in the imbalance risk score, indicating potential anomalies or data bias.

[0095] In this way, the prior constraint parameters established during the training phase not only guide the model's learning process, but also ensure the reliability and interpretability of the results during the inference phase through domain adaptation tables, boundary and remapping rules.

[0096] Step S105: Receive the myelin repair index and imbalance risk score, and generate an intervention parameter set, which includes electrical stimulation intervention parameters, virtual rehabilitation training task difficulty, and rehabilitation training rhythm.

[0097] In step S105, the system first receives the output of the deep learning inference engine from step S104. Each time slice includes a myelin repair index, an imbalance risk score, and a corresponding time stamp and confidence level. To ensure consistency between upstream and downstream interfaces, the system can also read the amount of summary generated by the inference engine within that time slice to reflect the centrality level, gradual trend, and fluctuation amplitude. Here, the input and output relationships of each unit of the deep learning inference engine in S105 are clarified again to ensure the traceability of this step: the recurrent unit receives the calibrated temporal feature representation and outputs the temporal latent representation in S104; the graph convolution unit receives the temporal latent representation and graph structure information and outputs the structure enhancement representation; the transfer learning module receives the structure enhancement representation and individual domain selection information and outputs the representation sequence aligned with the individual domains; the temporal convergence and stability evaluation layer receives the individual domain representation and outputs the statistical summary amount for that time slice; and the output layer converts the summary amount into the myelin repair index and the imbalance risk score during the inference stage. Step S105 strictly uses these explicit outputs as inputs to generate a set of intervention parameters and submits them to the next step of execution.

[0098] During implementation, before entering the intervention parameter generation process, the system synchronously loads individual profiles and constraint information, specifically including the individualized baseline parameters formed in step S103, previous adverse reaction markers, clinical contraindications, and device capability boundaries. Device capability boundaries specify the safe range and discrete levels available for the current electrical stimulation execution interface and the virtual rehabilitation training environment, such as the permissible stimulation current intensity and step granularity, upper and lower limits of pulse width and frequency, maximum duty cycle and ramp / slope time, and the available target size, range of motion, speed of motion, and joint freedom restrictions for the virtual training task. This information, along with the output of the inference engine, constitutes the complete input set for this step. The output is a set of intervention parameters valid within the time slice, including electrical stimulation intervention parameters, virtual rehabilitation training task difficulty, and rehabilitation training rhythm, along with the effective time period and feedback markers to facilitate closed-loop processing.

[0099] The parameter generation process follows a deterministic path. First, the system uses input convergence logic to standardize the myelin repair index and the imbalance risk score. If the confidence level is low or the data quality marker indicates a gap in the current time slice, the stable value from the previous time slice is used, and the range of change is limited to prevent sudden fluctuations. Then, a safety constraint layer loads device capability boundaries and medical contraindications, completely masking unusable intervals to ensure that subsequent algorithms search only within the permissible space. Next, the system enters the policy mapping and smoothing control stage. Policy mapping interprets the myelin repair index as a quantitative scale of "current carrying capacity" and the imbalance risk score as a scale of "short-term instability risk": the higher the carrying capacity, the higher the permissible training intensity and electrical stimulation dose range; the higher the instability risk, the lower the intensity and dose, and the longer the recovery interval. To prevent parameters from oscillating between adjacent time slices, the system introduces hysteresis and a rate of change limit in all channels. Only when the indicator crosses the set threshold and remains stable within the minimum holding time is the parameter allowed to switch to a higher or lower level. At the same time, all upward adjustments adopt a gradual increase in level, while downward adjustments adopt a combination of immediate decrease in level and slow recovery to prioritize safety.

[0100] Electrical stimulation intervention parameters are calculated using a channelized generator. The generator receives the myelin repair index, imbalance risk score, individualized baseline parameters, and device capability boundaries, and outputs intensity, pulse width, frequency, duty cycle, and ramp / ramp times in a fixed order. The initial reference value for intensity is taken from the midpoint between the "comfort threshold" and "effective threshold" in the individualized baseline parameters, and shifts upwards or downwards according to the myelin repair index, directly returning to near the comfort threshold in high-risk situations. Pulse width is constrained by the summation volume related to pathway integrity; a shorter pulse width and lower intensity are used when the structural recovery signal is weak, while a medium pulse width with lower intensity is allowed when the structural recovery signal is strong, to avoid excessive charge per unit time. The choice of frequency is closely related to phase consistency; low to medium frequencies are used when network synchronization is poor to prevent fatigue, while medium frequencies are allowed when network synchronization is good to improve effective recruitment. Duty cycle and ramp / ramp times are dominated by the imbalance risk score; when the risk is high, the working time is shortened, the rest time is prolonged, and the ramp / ramp times are lengthened to reduce abrupt stimulation changes. All parameters are immediately trimmed through a safety constraint layer after generation to ensure they do not exceed the boundaries of the device and the physiology.

[0101] The difficulty of virtual rehabilitation training tasks is determined by the task scheduler. The scheduler reads the trend and central level of the myelin repair index to determine the task family and difficulty level within that time slice, and then sets protective elements based on the imbalance risk score. The task family can revolve around target tracking, rhythmic coordination, or strength endurance. The difficulty level is achieved through a combination of factors such as target size, target speed, allowable error window, range of motion, joint degrees of freedom, and intensity of cognitive interference. When the myelin repair index continues to rise and fluctuates little, the target can be made smaller, the speed can be slightly increased, the allowable error window can be narrowed, and the degrees of freedom can be increased to improve the fine control load. When the imbalance risk increases, the target is enlarged, the speed is reduced, the error window is widened, the degrees of freedom are reduced, and more visual and auditory guidance is enabled. If necessary, it switches to a "guided-follow" mode, where the system provides partial motion trajectory assistance in the virtual environment to prevent mismatch.

[0102] The rehabilitation training rhythm is output by a pacemaker, which determines the duration of a single work segment and the total training duration based on the myelin repair index, and the work-rest ratio and minimum recovery interval based on the imbalance risk score. When the workload is high and the risk is low, longer work segments and standard rests are permitted; when the workload is moderate or the risk is high, work segments are shortened and intervals are lengthened, with complete rest segments inserted between adjacent time slices if necessary. The pacemaker is also responsible for generating consistent rhythmic signals for electrical stimulation and the virtual task, aligning muscle recruitment with the task's beat phase, and reducing fatigue and frustration caused by sensorimotor mismatch.

[0103] To ensure a clear implementation path, the system's workflow for this time slice begins with the convergence logic reading two metrics and the summary quantity from the inference engine and correcting the confidence level. Next, the safety constraint layer loads individualized baseline parameters, device capability boundaries, and a contraindication list. Then, the electrical stimulation generator, task scheduler, and rhythmizer calculate candidate values ​​for their respective channels in parallel. A smoothing control strategy then performs hysteresis, amplitude limiting, and rate limiting. Finally, it outputs three types of parameters: electrical stimulation intervention parameters, virtual rehabilitation training task difficulty, and rehabilitation training rhythm. These, along with the effective time period and feedback markers, are packaged together and passed to step S106 as the intervention parameter set. The output of the intervention parameter set is deterministic and does not rely on online learning. If the input is missing or of insufficient quality, a conservative strategy is triggered, automatically reducing the parameters to low intensity and tolerance difficulty, shortening the working segment, and extending the rest segment until a new, high-quality time slice becomes available.

[0104] Here's an example. Within a certain time slice, the myelin repair index output by the deep learning inference engine is moderately high and shows an upward trend for two consecutive time slices. The imbalance risk score is low, and the summary volume shows small fluctuations while the proportion of the pathway integrity component is increasing. Based on this, the input convergence logic confirms that it can enter the progressive up-adjustment channel. The safety constraint layer reads that the patient's electrical stimulation comfort threshold and effective threshold are located in the lower-middle range of the device intensity range, the maximum allowable frequency and duty cycle are both at the medium configuration of the device, and the patient has a history of adverse reactions to rapid intensity mutations. The electrical stimulation generator adjusts the intensity slightly above the comfort threshold, setting a moderate pulse width and low to medium frequency, lengthening the rise and fall times to avoid rapid abrupt changes, and maintaining the duty cycle within the device's allowable range. The task scheduler smoothly upgrades the virtual task from tracking a larger target at a slower speed to tracking a medium target size at a slightly faster speed, adjusting the allowable error window from loose to moderate, and increasing the joint degrees of freedom from a single plane to two planes. The rhythmic generator slightly extends the working segment while maintaining the same rest length, based on the original work-rest ratio, and broadcasts a consistent rhythm signal to the electrical stimulation generator and the task scheduler, ensuring that the electrical stimulation pulse train and the task beat remain sensorimotorally aligned. The smoothing control strategy imposes rate limits on the variation amplitude of the three channels, ensuring that the parameters of the current time slice change only in small steps compared to the previous time slice, and packages all results, along with the effective time of the time slice and the feedback marker, into an intervention parameter set output. When the next time slice begins, if the inference engine assesses an increased risk of imbalance or greater volatility, the strategy will immediately reduce the intensity, ease the task difficulty, and extend the pause period, avoiding frequent switching back and forth by using a predefined hysteresis threshold.

[0105] In this embodiment, the system can continuously read the myelin repair index and the imbalance risk score in the same processing thread, and establish a joint interpretation using both as the sole input. The joint interpretation is not a simple judgment of a single moment value, but rather uses a recent time window of fixed length as a reference, while retaining the judgment result of the previous time window as a hysteresis benchmark. The system first maps the current recovery level to a predefined grade scale based on the numerical range of the myelin repair index; then, based on the changing trend of the imbalance risk score, it compares the current value with the median level and extreme fluctuation points within the time window to identify whether the risk is rising, stabilizing, or falling. The output of the joint interpretation is a recovery status label, which is limited to three categories: steady recovery, controllable risk, or abnormal warning. It is subject to dual constraints of a minimum holding time and a switching threshold. The label is only allowed to be updated when a new judgment is continuously maintained for more than the minimum holding time and exceeds the switching threshold, to avoid frequent fluctuations near the boundaries. The system outputs the recovery status label to subsequent modules and simultaneously transmits the statistics of the time window used for judgment, ensuring that downstream systems can trace back the judgment basis.

[0106] The generation of the candidate set of electrical stimulation parameters is entirely driven by a parameter mapping rule set. This rule set is stored as a key-value table, with each rule using a recovery status label as the search key. Additional filtering conditions can be added, such as device capability boundaries and past adverse reaction markers. Each rule explicitly specifies the allowable range, recommended operating point, and step granularity for pulse amplitude, pulse frequency, and pulse duration, and records the compatibility constraints between these three parameters. For example, at higher pulse frequencies, the pulse duration must be shortened to control the charge per pulse, or at lower pulse amplitudes, a slightly longer pulse duration is allowed to maintain the recruitment effect. In the generation process, the system uses the recovery status label as input, retrieves matching rules from the parameter mapping rule set, uses the device capability boundary as a hard constraint to trim the allowable range, and then uses past adverse reaction markers as soft constraints to safely bias the recommended operating point, resulting in an individualized set of candidate electrical stimulation parameters. To avoid abrupt changes across time slices, the system uses the previously executed electrical stimulation intervention parameters as a starting point and applies the maximum instantaneous step limit to perform small-step updates within the candidate interval. If the candidate interval shifts significantly downward compared to the previous round, it prioritizes downgrading and delays upgrading to ensure safety first.

[0107] The virtual rehabilitation training task configuration uses an individualized set of electrical stimulation candidate parameters and rehabilitation status labels as inputs, and determines the training task type, difficulty level, and duration within the same time window. The training task type is selected from a preset task family. During steady recovery, fine control and coordination tasks are prioritized; when risks are manageable, basic tracking and rhythmic guidance tasks are assigned; and during abnormal alerts, guided-following and passive participation maintenance tasks are activated. The training task difficulty level is not determined in isolation but dynamically matched under the constraints of the electrical stimulation candidate parameter set: when pulse amplitude and pulse frequency are in the upper half of the candidate range, the training task difficulty level is limited to medium or slightly lower than medium to avoid applying excessive load to both sensory and motor sides simultaneously; when candidate parameters tend to be conservative, visual and attentional loads are allowed to be moderately increased in the virtual environment to maintain participation. The training task duration is adjusted based on the volatility of the imbalance risk score. Volatility is jointly given by the number of out-of-bounds occurrences within the time window and the concentration of differences between adjacent time points. Increased volatility shortens the training task duration and inserts more micro-rest periods, while decreased volatility extends the duration in minimum increments. The process outputs a structured virtual rehabilitation training task configuration, which clearly records the specific values ​​of the training task type, training task difficulty level, and training task duration, as well as consistency markers with the electrical stimulation candidate parameter set, for reference in the next step.

[0108] The generation of rehabilitation training rhythms uses virtual rehabilitation training task configurations as input and combines the stability of rehabilitation status tags to arrange training frequency, training rest intervals, and training progression over time. Within the current time window, the system first sets the basic training frequency based on the difficulty level of the training task, then sets the basic training rest interval based on the duration of the training task and previous adverse reaction markers, and subsequently adds rhythmic corrections based on the stability of the rehabilitation status tags. When recovery is steady and tags remain stable, the training frequency is allowed to gradually increase with the minimum increment, the training rest interval is allowed to decrease with the minimum decrement step size, and the training progression is allowed to slowly increase weekly or within several time windows. When the risk is controllable, the training frequency remains unchanged or decreases slightly, the training rest interval is slightly lengthened, and the training progression remains constant. When an abnormal warning occurs, the training frequency is immediately reduced to the safe lower limit, the training rest interval is lengthened to the safe upper limit, the training progression is reset to zero, and a buffer period begins. To avoid negative superposition between the rhythm and the set of electrical stimulation candidate parameters within the same time window, the system performs a consistency check after the arrangement is completed: if the sum of the training task duration and the training rest interval is insufficient to support the minimum effective working segment recommended by the set of electrical stimulation candidate parameters, the pulse frequency is preferentially reduced without exceeding the device capability limits and patient limitations, and the rhythm is recalculated until the check is passed.

[0109] After generating the rehabilitation training rhythm, the system enters the aggregation phase, converting the set of candidate electrical stimulation parameters into intervention parameters. The virtual rehabilitation training task configuration and the rehabilitation training rhythm are packaged into an intervention parameter set based on the same session time reference. When converting to intervention parameters, the system uses the recommended operating point within the candidate interval as a benchmark and applies maximum instantaneous step limits and hysteresis rules to generate the final value for the current time window. If the current time window is an abnormal warning, the system automatically selects the lower quartile operating point within the candidate interval and locks the upward adjustment channel until the rehabilitation status label returns to controllable risk or steady recovery. Phase alignment markers are added to the virtual rehabilitation training task configuration and rehabilitation training rhythm during packaging. These markers indicate the phase relationship between the training frequency beat and the electrical stimulation trigger time, ensuring a consistent sensory-motor experience at the execution end. The final form of the intervention parameter set is a structured record, which includes three parts: electrical stimulation intervention parameters, virtual rehabilitation training task configuration, and rehabilitation training rhythm. It also retains rehabilitation status labels, time window statistics, rule matching records, execution results of maximum instantaneous step limit, and consistency verification results, so that the decision path can be traced back when a downgrade or circuit breaker occurs at the execution end.

[0110] To ensure the feasibility of this solution, the system also sets up clear fallback paths for the three stages mentioned above. When the joint interpretation fails to generate a rehabilitation status label due to missing input, a conservative label is triggered, and the set of candidate electrical stimulation parameters is limited to the safe lower-middle range of the device's capability boundary. Simultaneously, the virtual rehabilitation training task configuration is switched to a low-difficulty training task type, the lowest training task difficulty level, and the shortest training task duration. The rehabilitation training rhythm reverts to a combination of low training frequency, long training rest intervals, and zero training progression. When the parameter mapping rule set lacks matching entries, the system performs small-step extrapolation based on the most recently valid electrical stimulation intervention parameters and the virtual rehabilitation training task configuration, and writes a default rule flag into the intervention parameter set, prohibiting upward adjustment without rule coverage. When the consistency check fails and no feasible combination can be found within the allowed number of iterations, the system forcibly reduces the training task difficulty level and pulse amplitude, and marks this time window as a conservative execution window to ensure safety.

[0111] Step S106: Generate control instructions based on the set of intervention parameters, and execute the control instructions to configure the virtual rehabilitation training environment and the electrical stimulation execution interface.

[0112] First, a session clock is established as the time reference for the entire system. This can be sourced from the operating system's high-precision clock or an external hardware time synchronization module, ensuring that the virtual rehabilitation training environment and the electrical stimulation execution interface share the same timing reference. Control commands are bound to an effective time period and a start timestamp upon generation, and the intervention parameter set is split into two types of loads: a task configuration load for the virtual rehabilitation training environment and a stimulation configuration load for the electrical stimulation execution interface. The task configuration load includes the training scene identifier to be enabled, target size and color scheme, target movement speed and trajectory type, allowable error window settings, joint degree-of-freedom restrictions, cognitive interference intensity, guidance and follow-up mode switching, prompting and feedback strategies, and rhythm-related beat configurations. The stimulation configuration load includes stimulation channel selection, initial intensity level, pulse width level, frequency level, duty cycle level, ramp and descent times, work-rest rhythm, intensity smoothing strategy, maximum instantaneous step limit, and safety circuit breaker threshold. Both types of workloads are encapsulated into a single control command. The control command also includes a session identifier, a patient anonymity identifier, a device capability boundary summary, a source hash of the input parameters, and a return tag, which facilitates post-execution verification and auditing.

[0113] Before issuing control commands, the execution agent performs an on-site consistency check between the intervention parameter set and the device's capability boundaries. If any boundary violations or conflicts are found (e.g., the range of motion required by the virtual task exceeds the patient's current limits, or the combination of stimulation parameters exceeds the device's allowed pulse train configuration), it automatically reverts to the nearest feasible level according to a predefined degradation strategy, and records the reason and specific fields for this adjustment in the remarks field of the control command. After the consistency check passes, the control channels are opened simultaneously: the channel facing the virtual rehabilitation training environment submits the task configuration load through a local API or network RPC. Upon receiving the load, the environment immediately enters the preloading process, loads the corresponding scene, resources, and interaction scripts, and returns a "pre-ready" confirmation; the channel facing the electrical stimulation execution interface submits the stimulation configuration load through physical links such as serial, USB, Ethernet, or Bluetooth, according to the data frame format published by the device manufacturer. After the device completes parameter loading and waveform synthesis preparation, it returns a "ready" confirmation. After receiving two "ready" signals and aligning them with the session clock, the execution agent broadcasts the same start timestamp and beat signal to both, enabling the key interaction points of the virtual task to be phase-aligned with the stimulus pulse train with millisecond-level precision, thus avoiding sensory-motor mismatch.

[0114] During execution, the virtual rehabilitation training environment interacts according to the task configuration load. The system renders the target, updates the trajectory, calculates hits or deviations, and provides visual and auditory feedback based on the allowable error window. The electrical stimulation execution interface generates pulse trains according to the stimulation configuration load in a "work-rest" rhythm, smoothly transitioning in the order of ramp-steady-ramp. Any adjustments to intensity and pulse width follow the maximum instantaneous step limit to ensure no abrupt transitions. To ensure operational safety, the execution agent continuously monitors two feedback channels throughout the effective period: the virtual environment feeds back the task status, current frame error, participation indicators, and device rendering load; the electrical stimulation device feeds back real-time output current, voltage, impedance, self-test alarms, and any triggered protection events. Once either feedback triggers the fuse threshold, such as detecting an abnormal impedance or a sharp drop in participation, the execution agent immediately broadcasts a pause signal, simultaneously issuing an emergency downshift or disconnect command to the electrical stimulation device and a pause and prompt overlay display command to the virtual environment. When the alarm is cleared and the minimum recovery interval is met, the system recovers to the safe level in small steps according to a conservative curve.

[0115] To achieve a closed loop, this step can also create a data acquisition thread alongside the execution thread. Following the sampling frequency and timestamp specifications of step S101, new multimodal neural signal data is re-acquired and cached. The acquisition thread writes the session clock timestamp into each packet header and pushes the data back to the preprocessing flow as a "post-intervention data block" after completing its time slice. Simultaneously, the execution agent archives the control commands, device confirmation frames, key transmission points, and any degradation or circuit breaker events as a serialized log. The log contains the complete context used to reproduce the experiment, facilitating subsequent auditing and reuse of training data.

[0116] For example, before the start of a certain time slice, the intervention parameter set shows that the electrical stimulation intervention parameters are at a moderate intensity and low to medium frequency, the virtual rehabilitation training task difficulty is moderate, the allowable error window is moderate, the joint degrees of freedom are two planes, and the rehabilitation training rhythm requires a slightly longer working segment and a standard rest segment. During consistency verification, the control command finds that the patient's device capability boundary records "rapid intensity mutations have caused discomfort," so it automatically extends the ramp time and tightens the maximum instantaneous step limit. Subsequently, the virtual environment returns to pre-ready confirmation, the electrical stimulation device returns to ready confirmation, and the execution agent broadcasts the same start timestamp and beat signal to both ends based on the session clock. After execution, the virtual environment presents a smaller, slightly faster, and more complex target according to the beat, and issues a progressive prompt when the patient's deviation approaches the edge of the allowable error window; the electrical stimulation device enters steady-state output according to the set ramp within each working segment, smoothly terminates with a ramp at the end of the segment, and is completely silent during the rest segment. Midway through the time slice, the device feedback showed a brief increase in electrode impedance, which had not yet exceeded the fuse threshold. However, the combination of the imbalance risk score and the fluctuation summary indicated a possible short-term fatigue. The execution agent immediately adopted a conservative strategy, lowering the duty cycle by one level and extending the rest segment at the end of the time slice. Simultaneously, it notified the virtual environment to slightly reduce the target speed and slightly widen the error window. All adjustments were completed within the maximum instantaneous step limit and took effect at the next beat boundary. After the time slice ended, the acquisition thread stored the post-intervention multimodal neural signal data, along with the control commands, device confirmation frames, and feedback summaries, for reuse in subsequent steps. Through this process, control commands were precisely packaged, synchronized, executed, and transmitted back. The virtual rehabilitation training environment and the electrical stimulation execution interface were configured consistently, and reliable implementation and re-evaluation of the intervention parameter set were achieved under the protection of full safety and closed-loop acquisition.

[0117] In one embodiment, the system first performs structured parsing of the intervention parameter set. This set, generated in the previous step, comprises three parts: electrical stimulation intervention parameters, virtual rehabilitation training task configuration, and rehabilitation training rhythm. Upon receiving this set, the system unpacks each of the three parameter types and assigns them a unique control channel identifier to ensure no confusion occurs during subsequent data transmission and command generation. Each control channel corresponds to a specific parameter category, such as the electrical stimulation intervention parameter channel, the virtual rehabilitation training task configuration channel, and the rehabilitation training rhythm channel. All channels are organized on a unified control bus and output as parameter sequences with channel identifiers. This structured output ensures that different types of parameters can be accurately identified and invoked in subsequent modules.

[0118] In the electrical stimulation intervention section, the system uses the electrical stimulation intervention parameters from the parameter sequence as input to generate an electrical stimulation control instruction set. This instruction set consists of three parts: pulse amplitude setting instructions, pulse frequency setting instructions, and pulse duration setting instructions. To ensure that these instructions can be safely executed in a real device, the system introduces a device capability boundary verification mechanism. This mechanism verifies each item in the generated electrical stimulation control instruction set, comparing it with the maximum output voltage, current, pulse width, and frequency allowed by the device hardware. If a parameter exceeds the allowable range, the system automatically trims or lowers it to a safe range and records the reason for the adjustment in the control log, thus outputting a verified electrical stimulation control instruction set. The advantage of this approach is that it avoids theoretical parameters directly affecting the patient, ensuring the safety and stability of the execution process.

[0119] In the virtual rehabilitation training section, the system takes the virtual rehabilitation training task configuration and rehabilitation training rhythm from the parameter sequence as input to generate a virtual rehabilitation training control instruction set. This instruction set consists of instructions for loading training task types, setting training task difficulty levels, and setting training task duration, and embeds the training frequency, rest intervals, and training progression defined in the rehabilitation training rhythm. During this process, the system coordinates the execution of the virtual training task with the output rhythm of the electrical stimulation based on the constraints of the intervention parameter set. For example, when the electrical stimulation parameters are set in a high-intensity range, the system automatically reduces the difficulty level of the training task or increases the rest interval in the virtual rehabilitation training control instruction set, thereby avoiding excessive dual load on the patient within the same time window. The final output virtual rehabilitation training control instruction set not only contains the parameters of the training task itself but also carries rhythmic information, ensuring that the virtual environment and the electrical stimulation execution interface operate at a unified rhythm.

[0120] During the synchronous execution phase, the system simultaneously loads the verified electrical stimulation control instruction set and the integrated virtual rehabilitation training control instruction set into the scheduler for synchronous scheduling based on a unified time reference. The synchronous scheduling process aligns the timestamps of both types of instructions to ensure that pulse triggering and training task switching strictly adhere to the set rehabilitation training rhythm. When generating the synchronous execution instruction sequence, the system further checks for rhythmic conflicts between electrical stimulation and virtual training. For example, if both high-frequency pulse stimulation and high-difficulty task operations are required at the same time, the scheduler will automatically delay or reduce the intensity of one based on priority rules to avoid overloading the patient. The final generated synchronous execution instruction sequence is sent to the electrical stimulation execution interface and the virtual rehabilitation training environment through a standardized interface protocol, enabling real-time configuration and execution of electrical stimulation intervention parameters, virtual rehabilitation training task configuration, and rehabilitation training rhythm.

[0121] Through the above process, it is clear how to start from the set of intervention parameters, go through parsing, instruction generation, verification, integration and synchronization scheduling, and finally execute the intervention plan in the actual equipment and virtual environment.

[0122] A second embodiment of this application provides an electronic device, the electronic device comprising:

[0123] processor;

[0124] The memory is used to store a program, which, when read and executed by the processor, performs a myelin repair assist method based on neural signal modeling provided in the first embodiment of this application.

[0125] The third embodiment of this application provides a computer-readable storage medium storing a computer program thereon. When the program is executed by a processor, it performs a myelin repair assist method based on neural signal modeling provided in the first embodiment of this application.

[0126] Although this application discloses preferred embodiments as described above, it is not intended to limit this application. Any person skilled in the art can make possible changes and modifications without departing from the spirit and scope of this application. Therefore, the scope of protection of this application should be determined by the scope defined in the claims of this application.

Claims

1. An electronic device, characterized in that, The electronic device includes: processor; The memory stores a program that, when read and executed by the processor, performs a myelin repair-assisted method based on neural signal modeling, including: Multimodal neural signal data of the individual to be rehabilitated is acquired and preprocessed to obtain a preprocessed dataset. The multimodal neural signal data includes electroencephalogram (EEG) signals, functional magnetic resonance imaging (fMRI) signals, and peripheral nerve electrophysiological signals. Based on the preprocessed dataset, an individualized temporal modeling process is performed to generate a temporal feature representation characterizing the dynamic process of myelin sheath injury-repair. The temporal feature representation includes at least the neural conduction velocity recovery rate index, the phase consistency index, and the neural pathway integrity index. Individualized baseline parameters are calculated using a dynamic baseline update mechanism, and the time series feature representation is calibrated using the individualized baseline parameters to obtain a calibrated time series feature representation. The calibration temporal feature representation is used as input and fed into a pre-trained deep learning inference engine that includes recurrent units, graph convolutional units, and transfer learning modules. The deep learning inference engine outputs a myelin repair index that represents the degree of myelin repair and an imbalance risk score that represents the probability of future repair abnormalities. During the training phase, the deep learning inference engine embeds the prior constraint parameters generated by the individualized temporal modeling process into the loss construction and regularization of the deep learning inference engine. The system receives the myelin repair index and imbalance risk score, and generates a set of intervention parameters, which includes electrical stimulation intervention parameters, virtual rehabilitation training task difficulty, and rehabilitation training rhythm. Control instructions are generated based on the set of intervention parameters, and the control instructions are executed to configure the virtual rehabilitation training environment and the electrical stimulation execution interface; The step of calculating individualized baseline parameters through a dynamic baseline update mechanism, and calibrating the time-series feature representation using the individualized baseline parameters to obtain a calibrated time-series feature representation, includes: Baseline segments are extracted from the temporal feature representation. Stable segments of neural conduction velocity recovery rate index, phase consistency index and neural pathway integrity index are identified within a preset time window, and a set of baseline segments is output. Using the baseline segment set as input, abnormal segment removal is performed. Based on statistical thresholds and data quality markers, segments with instantaneous spikes, significant deviations, or low-quality markers are removed, and a valid baseline segment set is output. Using the set of effective baseline segments as input, weighted time series fusion is performed, assigning higher weights to baseline segments in recent time windows and lower weights to baseline segments in distant time windows, and outputting dynamically updated individualized baseline parameters. Using the individualized baseline parameters as input, the temporal feature representation is calibrated point by point. The neural conduction velocity recovery rate index is adjusted to the velocity recovery difference relative to the individualized baseline parameters, the phase consistency index is adjusted to the consistency difference relative to the individualized baseline parameters, and the neural pathway integrity index is adjusted to the integrity difference relative to the individualized baseline parameters. The calibrated temporal feature representation is then output.

2. The electronic device according to claim 1, characterized in that, The process of acquiring and preprocessing multimodal neural signal data of the individual to be rehabilitated to obtain a preprocessed dataset includes: The artifact removal process takes the original EEG signal, functional magnetic resonance imaging (fMRI) signal, and peripheral nerve electrophysiological signal as inputs. It uses a pre-set artifact template library to perform time-window matching and removal of eye movement artifacts, electrocardiogram artifacts, electromyography artifacts, and power frequency interference on the EEG signal. It performs slice time correction, head movement trajectory correction, and spatial normalization on the fMRI signal. It performs baseline drift correction and stimulus label error correction on the peripheral nerve electrophysiological signal. It outputs a clean multimodal signal set and generates data quality labels for each time window. Based on the clean multimodal signal set, a cross-modal time synchronization process is performed, aligning EEG events, MRI volume sampling and electrophysiological stimulation response events according to hardware trigger signals or system timestamps, outputting a time synchronization signal set and assigning an acquisition time marker to each time point; Using the time synchronization signal set as input, a feature standardization and mutual information enhancement process is performed. The amplitude of each modality is standardized and the noise energy ratio is corrected according to the value range recorded during the training phase. Within the sliding time window, an alignment confidence score is generated by maximizing the deterministic mutual information metric between each modality. The standardized and aligned result is output as a feature-enhanced aligned signal set, and the alignment confidence score is incorporated into the data quality label of the corresponding time window. Based on the feature-enhanced aligned signal set, a quality-weighted combination process is performed. Dynamic modal weights are assigned to EEG signals, functional magnetic resonance signals, and peripheral neurophysiological signals according to data quality labels and alignment confidence scores. Low-quality time windows are masked and bounded interpolation is performed on adjacent time windows to output a quality-weighted signal set. The quality-weighted signal set is converted into a unified data structure, the order and dimension of the multimodal feature components are fixed, and the acquisition time marker and data quality marker are encapsulated together to form a time-ordered feature vector sequence as the preprocessed dataset.

3. The electronic device according to claim 1, characterized in that, The process of performing individualized temporal modeling based on the preprocessed dataset generates temporal feature representations characterizing the dynamic process of myelin sheath injury-repair. These temporal feature representations include at least neural conduction velocity recovery rate indices, phase consistency indices, and neural pathway integrity indices, including: The preprocessed dataset is divided into multi-scale time segments. The EEG signals, functional magnetic resonance signals and peripheral nerve electrophysiological signals are segmented into short-time segments and long-time segments to obtain a multi-scale time segment sequence with time order. Based on the multi-scale time segment sequence, instantaneous phase events in the EEG signal are identified, and local activation changes in the functional magnetic resonance signal are aligned on a unified time axis. At the same time, the stimulus and response latencies in the peripheral nerve electrophysiological signal are mapped to the alignment results to obtain cross-modal time correspondence. Based on the cross-modal time correspondence, a time-series change trajectory is generated, which reflects the trend of latency shortening or lengthening, the trend of phase concentration or dispersion, and the trend of neural pathway connectivity enhancement or weakening, thus obtaining a trajectory set composed of multiple time-series change trajectories. Based on the trajectory set, individualized dynamic features are extracted. The temporal change trajectory of the latency period is converted into a neural conduction velocity recovery rate index, the temporal change trajectory of the phase is converted into a phase consistency index, and the temporal change trajectory of the pathway connectivity is converted into a neural pathway integrity index. A feature vector sequence consisting of the three indices is output. The feature vector sequence is arranged in chronological order to form a temporal feature representation.

4. The electronic device according to claim 1, characterized in that, The deep learning inference engine includes an input adaptation layer, a recurrent unit, a graph convolutional unit, a transfer learning module, a temporal convergence and stability evaluation layer, and an output layer. The input adaptation layer takes the calibration time-series feature representation as input. The calibration time-series feature representation is a sequence of feature vectors arranged in chronological order. Each time point includes at least the neural conduction velocity recovery rate index, phase consistency index, and neural pathway integrity index, and is accompanied by acquisition time stamp and data quality stamp. The input adaptation layer divides the feature vector sequence into standardized time slices by fixed-length segmentation, performs missing point removal or bounded interpolation based on the data quality stamp, and performs consistency processing on the feature order and value range, outputting a standardized time slice sequence. The recurrent unit takes the standardized time slice sequence as input, receives the feature vector step by step in each time slice and maintains the internal state to capture short-term fluctuations and slow trends, and automatically resets between time slices; during the inference process, the recurrent unit performs fixed-scale adjustment based on the individualized baseline parameters to eliminate rhythm differences without changing the model parameters, and outputs the temporal implicit representation sequence. The graph convolutional unit takes the temporal implicit representation sequence and a preset graph structure as input. The graph structure defines the connection relationships and edge weights between the neural conduction velocity recovery rate index, the phase consistency index, and the neural pathway integrity index. During the inference process, the graph convolutional unit only performs bounded fine-tuning of the edge weights based on the degree of co-change and data quality labels within the current time slice, without involving new parameter learning. It achieves cross-index and cross-region information interaction through neighborhood aggregation and outputs a structure-enhanced representation sequence. The transfer learning module takes the structure-enhanced representation sequence and individual domain selection information as input, and the individual domain selection information is determined based on individualized baseline parameters and acquisition environment information. The transfer learning module retrieves the alignment mapping corresponding to the individual domain selection information from the domain adaptation table. The alignment mapping is stored in the form of feature channel scaling factors and bias values. During the inference process, linear recalibration and offset correction are performed on each channel without introducing new learning, and the output is a representation sequence aligned with the individual domain. The temporal convergence and stability evaluation layer takes the individual domain aligned representation sequence as input, and obtains the central level and asymptotic trend through weighted aggregation within the time slice scale. The weighting coefficient is determined by the data quality label and the weight of the nearest baseline. At the same time, it counts the differences between adjacent time steps and the number of out-of-bounds anomalies to obtain the fluctuation amplitude and instantaneous anomaly density, and outputs a time slice-level summary set. The output layer takes the time-slice-level summary set as input and sets up two parallel output paths. The first path generates the myelin repair index based on the central level, progressive trend and neural pathway integrity-related components. The second path generates the imbalance risk score based on the fluctuation amplitude, instantaneous abnormal density and cross-index inconsistency. During the inference process, the two paths only perform boundary clipping and reliability remapping without updating parameters, and output the myelin repair index and imbalance risk score that correspond one-to-one with the input time slice. During the training phase, the domain adaptation tables of the recurrent units, graph convolutional units, and transfer learning modules, as well as the boundary and remapping rules of the output layer, are learned offline and solidified by introducing prior constraint parameters generated during the individualized temporal modeling process into the loss construction and regularization. During the inference phase, the system operates only according to the data flow and deterministic rules.

5. The electronic device according to claim 1, characterized in that, The receiver receives the myelin repair index and imbalance risk score to generate an intervention parameter set, which includes electrical stimulation intervention parameters, virtual rehabilitation training task difficulty, and rehabilitation training rhythm, including: The myelin repair index and the imbalance risk score are jointly interpreted. Based on the numerical range of the myelin repair index and the changing trend of the imbalance risk score, a rehabilitation status label is generated. The rehabilitation status label is divided into three categories: steady recovery, controllable risk, and abnormal warning. The rehabilitation status label is output as the basis for subsequent processing. Using the rehabilitation status label as input, the parameter mapping rule set is called to generate a set of candidate electrical stimulation parameters. The set of candidate electrical stimulation parameters includes pulse amplitude, pulse frequency and pulse duration. The value range of the set of candidate electrical stimulation parameters is determined by the rehabilitation status label, and an individualized set of candidate electrical stimulation parameters is output. Using the individualized set of electrical stimulation candidate parameters and rehabilitation status labels as input, a virtual rehabilitation training task configuration is generated. The virtual rehabilitation training task configuration includes training task type, training task difficulty level and training task duration. The training task type is determined by the rehabilitation status label. The training task difficulty level is dynamically matched within the range of the electrical stimulation candidate parameter set. The training task duration is corrected according to the volatility of the imbalance risk score. The virtual rehabilitation training task configuration is then output. Using the virtual rehabilitation training task configuration as input, a rehabilitation training rhythm is generated, which includes training frequency, training rest interval and training progression amplitude. Finally, electrical stimulation intervention parameters, virtual rehabilitation training task configuration and rehabilitation training rhythm are output, forming an intervention parameter set.

6. The electronic device according to claim 1, characterized in that, The step of generating control instructions based on the intervention parameter set and executing the control instructions to configure the virtual rehabilitation training environment and the electrical stimulation execution interface includes: The set of intervention parameters is analyzed, and the electrical stimulation intervention parameters, virtual rehabilitation training task configuration and rehabilitation training rhythm are separated respectively. A corresponding mapping relationship is established in a unified control channel, and a parameter sequence with channel identifier is output. Using the electrical stimulation intervention parameters in the parameter sequence as input, an electrical stimulation control instruction set is generated. The electrical stimulation control instruction set includes pulse amplitude setting instructions, pulse frequency setting instructions, and pulse duration setting instructions. The legality of the electrical stimulation control instruction set is verified through a device capability boundary verification mechanism, and the verified electrical stimulation control instruction set is output. Using the virtual rehabilitation training task configuration and rehabilitation training rhythm in the parameter sequence as input, a virtual rehabilitation training control instruction set is generated. The virtual rehabilitation training control instruction set includes instructions for loading training task type, setting training task difficulty level, and setting training task duration. The training frequency, training rest interval, and training progression amplitude in the rehabilitation training rhythm are embedded into the virtual rehabilitation training control instruction set, and an integrated virtual rehabilitation training control instruction set is output. The verified electrical stimulation control instruction set and the integrated virtual rehabilitation training control instruction set are synchronously scheduled under a unified time reference to generate a synchronous execution instruction sequence, which is distributed to the electrical stimulation execution interface and the virtual rehabilitation training environment through an interface protocol to complete the real-time configuration and execution of electrical stimulation intervention parameters, virtual rehabilitation training task configuration, and rehabilitation training rhythm.

Images