CN-121506504-B - Myelin repair auxiliary method based on neural signal modeling

CN121506504BCN 121506504 BCN121506504 BCN 121506504BCN-121506504-B

Abstract

The invention relates to a myelin sheath repair auxiliary method based on neural signal modeling. The method comprises the steps of obtaining multi-mode nerve signal data such as brain electrical signals of an individual to be rehabilitated, preprocessing the multi-mode nerve signal data to form a preprocessed data set, then executing an individuation time sequence modeling process, generating time sequence characteristic representations containing indexes such as nerve conduction speed recovery rate, phase consistency, nerve path integrity and the like, and carrying out individuation calibration through a dynamic baseline updating mechanism. The calibrated feature representation is input to a pre-trained deep learning inference engine. The inference engine outputs myelin repair index and imbalance risk score for characterizing repair levels and predicting future abnormal risk. And generating an intervention parameter set based on the indexes, wherein the intervention parameter set comprises an electric stimulation intervention parameter, a virtual rehabilitation training task difficulty and a rehabilitation training rhythm, and further forming a control instruction to configure a virtual rehabilitation training environment and an electric stimulation execution interface, so that personalized assistance and dynamic optimization of remyelination repair are realized.

Inventors

CHEN RUIBING
WU YU
ZHANG LEI
FANG WENJIE
SHAO QI
ZHANG HUA
LEI WENZHI
QI TIAN
FANG YUE
LI YAOXIN

Assignees

中国人民解放军海军军医大学

Dates

Publication Date: 20260512
Application Date: 20260112

Claims (6)

1. An electronic device, the electronic device comprising: A processor; a memory for storing a program which, when read for execution by the processor, performs a myelin repair assistance method based on neural signal modeling, comprising: Acquiring multi-modal nerve signal data of an individual to be rehabilitated, and preprocessing the multi-modal nerve signal data to acquire a preprocessing data set, wherein the multi-modal nerve signal data comprises an electroencephalogram signal, a functional magnetic resonance signal and a peripheral nerve electrophysiological signal; Performing an individualized time sequence modeling process based on the preprocessed data set, generating a time sequence feature representation representing a myelin sheath damage-repair dynamic process, the time sequence feature representation including at least a nerve conduction velocity recovery rate index, a phase consistency index, and a nerve pathway integrity index; Calculating an individualized baseline parameter through a dynamic baseline updating mechanism, and calibrating the time sequence characteristic representation by the individualized baseline parameter to obtain a calibrated time sequence characteristic representation; The method comprises the steps of taking a calibration time sequence characteristic representation as input, sending the calibration time sequence characteristic representation into a pre-trained deep learning reasoning engine comprising a circulation unit, a graph convolution unit and a migration learning module, and outputting a myelin repair index representing myelin repair degree and an unbalance risk score representing future repair abnormality probability through the deep learning reasoning engine, wherein the deep learning reasoning engine embeds prior constraint parameters generated in the individuation time sequence modeling process into a loss structure and regularization of the deep learning reasoning engine in a training stage; Receiving a myelin repair index and an unbalance risk score, and generating an intervention parameter set, wherein the intervention parameter set comprises an electrical stimulation intervention parameter, a virtual rehabilitation training task difficulty and a rehabilitation training rhythm; generating a control instruction according to the intervention parameter set, and executing the control instruction to configure a virtual rehabilitation training environment and an electric stimulation execution interface; calculating an individualized baseline parameter through a dynamic baseline updating mechanism, calibrating the time sequence characteristic representation by the individualized baseline parameter to obtain a calibrated time sequence characteristic representation, wherein the method comprises the following steps of: Extracting the baseline segment from the time sequence characteristic representation, identifying stable segments of nerve conduction velocity recovery rate index, phase consistency index and nerve path integrity index in a preset time window, and outputting a baseline segment set; Taking the baseline fragment set as input, executing abnormal fragment elimination, removing fragments with instantaneous peaks, significant deviations or low-quality marks according to a statistical threshold and a data quality mark, and outputting an effective baseline fragment set; taking the effective baseline segment set as input, executing weighted time sequence fusion, giving higher weight to the baseline segment of the near-term time window, giving lower weight to the baseline segment of the far-term time window, and outputting dynamically updated individualized baseline parameters; And carrying out point-by-point calibration on the time sequence characteristic representation by taking the individuation baseline parameter as input, adjusting a nerve conduction velocity recovery rate index to a velocity recovery difference value relative to the individuation baseline parameter, adjusting a phase consistency index to a consistency difference value relative to the individuation baseline parameter, adjusting a nerve path integrity index to an integrity difference value relative to the individuation baseline parameter, and outputting a calibration time sequence characteristic representation.
2. The electronic device of claim 1, wherein the acquiring and preprocessing the multimodal neural signal data of the individual to be rehabilitated to obtain the preprocessed data set comprises: Performing an artifact eliminating process, namely performing eye movement artifact, electrocardio artifact, myoelectricity artifact and time window matching elimination of power frequency interference on the electroencephalogram by using a preset artifact template library by taking an original electroencephalogram, a functional magnetic resonance signal and a peripheral nerve electrophysiological signal as inputs, performing slicing time correction, head movement track correction and space normalization on the functional magnetic resonance signal, performing baseline drift correction and stimulation mark correction on the peripheral nerve electrophysiological signal, outputting a clean multi-mode signal set and generating a data quality mark for each time window; Based on the cleaning multi-mode signal set, executing a cross-mode time synchronization process, aligning an electroencephalogram event, a magnetic resonance volume sampling event and an electrophysiological stimulation response event according to a hardware trigger signal or a system time stamp, outputting a time synchronization signal set and giving an acquisition time mark for each time point; Performing characteristic standardization and mutual information enhancement processes by taking the time synchronization signal set as input, carrying out amplitude standardization and noise energy ratio correction on the amplitude values of all modes according to the value range recorded in the training stage, generating alignment confidence scores by maximizing deterministic mutual information measurement among all modes in a sliding time window, outputting standardized and alignment enhanced results as characteristic enhancement alignment signal sets, and integrating the alignment confidence scores into data quality marks of corresponding time windows; Based on the characteristic enhancement alignment signal set, executing a quality weighted combination process, distributing dynamic modal weights to the electroencephalogram signals, the functional magnetic resonance signals and the peripheral nerve electrophysiological signals according to the data quality marks and the alignment confidence scores, shielding a low-quality time window, carrying out bounded interpolation on an adjacent time window, and outputting a quality weighted signal set; And converting the quality weighted signal set into a unified data structure, fixing the sequence and the dimension of the multi-mode characteristic components, and encapsulating the acquisition time mark and the data quality mark together to form a characteristic vector sequence which is arranged in time sequence as the preprocessing data set.
3. The electronic device of claim 1, wherein the performing an individualized time series modeling process based on the preprocessed data set generates a time series characterization representation that characterizes myelin damage-repair dynamics, the time series characterization representation including at least a nerve conduction velocity recovery index, a phase consistency index, and a nerve pathway integrity index, comprising: Dividing the preprocessing data set into multi-scale time segments, and dividing the electroencephalogram signals, the functional magnetic resonance signals and the peripheral nerve electrophysiological signals into short-time segments and long-time segments to obtain a multi-scale time segment sequence with a time sequence; According to the multi-scale time segment sequence, identifying an instantaneous phase event in an electroencephalogram signal, aligning local activation change in a functional magnetic resonance signal on a unified time axis, and simultaneously, corresponding the stimulus and response latency in a peripheral nerve electrophysiological signal to an alignment result to obtain a cross-modal time corresponding relation; Generating a time sequence change track according to the cross-modal time corresponding relation, forming a time sequence change track reflecting a trend of shortening or extending the latency period on a time axis, forming a time sequence change track reflecting a trend of concentrating or dispersing phases, forming a time sequence change track reflecting a trend of enhancing or weakening the connectivity of the neural pathway, and obtaining a track set consisting of a plurality of time sequence change tracks; According to the track set, extracting individual dynamic characteristics, converting a time sequence change track of a latency period into a nerve conduction speed recovery rate index, converting a time sequence change track of a phase into a phase consistency index, converting a time sequence change track of channel connectivity into a nerve channel integrity index, and outputting a feature vector sequence consisting of three indexes, wherein the feature vector sequence is arranged according to a time sequence to form a time sequence feature representation.
4. The electronic device of claim 1, wherein the deep learning inference engine comprises an input adaptation layer, a circulation unit, a graph convolution unit, a transfer learning module, a timing aggregation and stability assessment layer, and an output layer; The input adaptation layer takes a calibration time sequence characteristic as input, the calibration time sequence characteristic is a characteristic vector sequence which is arranged in time sequence, each time point at least comprises a nerve conduction speed recovery rate index, a phase consistency index and a nerve path integrity index, and is additionally provided with a collection time mark and a data quality mark; The circulation unit takes the standardized time slice sequence as input, receives the characteristic vector in each time slice step by step, maintains the internal state to capture short-term fluctuation and slow trend, and automatically resets between time slices; The graph convolution unit takes the time sequence implicit expression sequence and a preset graph structure as input, and the graph structure defines the connection relation and side weight among nerve conduction speed recovery rate indexes, phase consistency indexes and nerve path integrity indexes; the graph convolution unit performs bounded fine adjustment on the boundary weights only according to the cooperative variation degree and the data quality mark in the current time slice in the reasoning process, does not involve new parameter learning, realizes cross-index and cross-region information interaction through neighborhood aggregation, and outputs a structure enhancement representation sequence; The migration learning module takes a structure enhanced representation sequence and individual domain selection information as input, the individual domain selection information is determined based on an individual baseline parameter and acquisition environment information, the migration learning module searches an alignment mapping corresponding to the individual domain selection information in a domain adaptation table, the alignment mapping is stored in a characteristic channel scaling factor and offset value form, linear recalibration and offset correction are carried out on each channel in the reasoning process, new learning is not introduced, and a representation sequence aligned by the individual domain is output; The time sequence convergence and stability assessment layer takes the representation sequence aligned by the individual domains as input, obtains a central level and a progressive trend in a time slice scale through weighted summarization, and determines a weighting coefficient by a data quality mark and a latest baseline weight; The output layer takes the time slice level abstract quantity set as input, sets two parallel output channels, wherein the first channel generates a myelin sheath repair index according to central level, progressive trend and related components of neural channel integrity, the second channel generates an unbalance risk score according to fluctuation amplitude, instantaneous abnormal density and cross index inconsistency, and the two channels only execute boundary clipping and reliability remapping without updating parameters in the reasoning process and output the myelin sheath repair index and the unbalance risk score which are in one-to-one correspondence with the input time slices; in the training stage, the circulation unit, the graph convolution unit, the domain adaptation table of the transfer learning module and the boundary and remapping rules of the output layer are used for offline learning and solidification by introducing prior constraint parameter embedding loss construction and regularization generated in the individuation time sequence modeling process, and the reasoning stage only operates according to the data flow and the certainty rules.
5. The electronic device of claim 1, wherein the receiving myelin repair index and imbalance risk score generates an intervention parameter set comprising electrical stimulation intervention parameters, virtual rehabilitation training task difficulty, and rehabilitation training rhythms, comprising: the myelin sheath repair index and the unbalance risk score are jointly interpreted, a recovery state label is generated according to the numerical interval of the myelin sheath repair index and the variation trend of the unbalance risk score, the recovery state label is divided into three types of stable recovery, risk control and abnormal warning, and the recovery state label is output as a judgment basis for subsequent processing; The rehabilitation state label is used as input, a parameter mapping rule set is called, an electrical stimulation candidate parameter set is generated, the electrical stimulation candidate parameter set comprises pulse amplitude, pulse frequency and pulse duration, the value range of the electrical stimulation candidate parameter set is determined through the rehabilitation state label, and the individualized electrical stimulation candidate parameter set is output; Generating virtual rehabilitation training task configuration by taking the individualized electrical stimulation candidate parameter set and the rehabilitation state label as inputs, wherein the virtual rehabilitation training task configuration comprises a training task type, a training task difficulty level and a training task duration, the training task type is determined by the rehabilitation state 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 fluctuation of the unbalance risk score, and the virtual rehabilitation training task configuration is output; And the virtual rehabilitation training task is configured as input to generate a rehabilitation training rhythm, wherein the rehabilitation training rhythm comprises training frequency, training rest interval and training progressive amplitude, and finally, electric stimulation intervention parameters, virtual rehabilitation training task configuration and rehabilitation training rhythm are output to form an intervention parameter set.
6. The electronic device of claim 1, wherein the generating control instructions from the set of intervention parameters and executing the control instructions to configure a virtual rehabilitation training environment with an electrical stimulation execution interface comprises: analyzing the intervention parameter set, separating the electric stimulation intervention parameter, the virtual rehabilitation training task configuration and the rehabilitation training rhythm from each other, establishing a corresponding mapping relation in a unified control channel, and outputting a parameter sequence with a channel identifier; Taking electrical stimulation intervention parameters in the parameter sequence as input, generating an electrical stimulation control instruction set, wherein the electrical stimulation control instruction set comprises a pulse amplitude setting instruction, a pulse frequency setting instruction and a pulse duration setting instruction, verifying the validity of the electrical stimulation control instruction set through a device capability boundary verification mechanism, and outputting a verified electrical stimulation control instruction set; The virtual rehabilitation training control instruction set comprises a training task type loading instruction, a training task difficulty level setting instruction and a training task duration setting instruction, training frequency, a training rest interval and a training progressive amplitude in the rehabilitation training rhythm are embedded into the virtual rehabilitation training control instruction set, an integrated virtual rehabilitation training control instruction set is output, synchronous scheduling is carried out on the checked electric stimulation control instruction set and the integrated virtual rehabilitation training control instruction set under a unified time reference, a synchronous execution instruction sequence is generated, and the synchronous execution instruction sequence is distributed to an electric stimulation execution interface and a virtual rehabilitation training environment through an interface protocol, so that real-time configuration and execution of electric stimulation intervention parameters, virtual rehabilitation training task configuration and rehabilitation training rhythm are completed.

Description

Myelin repair auxiliary method based on neural signal modeling Technical Field The invention relates to the technical field of medical health care informatics, in particular to a myelin sheath repair auxiliary method based on neural signal modeling. Background In the prior art, myelin repair research following central nervous system injury mainly relies on single-modality detection and evaluation means, such as monitoring nerve conduction function by electroencephalogram signals, or observing white matter structural changes by using functional magnetic resonance imaging, or detecting conduction velocity and excitation characteristics by peripheral nerve electrophysiology. The methods are widely applied to clinical and scientific research, are used for evaluating the damage degree and partial recovery condition of the nervous system, and provide a certain data support for rehabilitation training or electric stimulation intervention. However, the prior art has obvious defects that firstly different mode data are usually used in isolation, the cross-mode fusion and unified modeling are lacked, so that the dynamic process of myelin damage and repair cannot be comprehensively reflected, secondly, individual differences cannot be fully considered, the average value of groups is often relied on as a reference, the dynamic correction of the baseline state of an individual is lacked, thirdly, in the aspect of rehabilitation intervention, most methods only regulate stimulation parameters based on a single index, and the multi-dimensional comprehensive optimization of training task difficulty, rhythm and electrical stimulation scheme is difficult to realize. The above problems limit the level of accuracy and personalization of rehabilitation assistance methods. In view of this, a new myelin repair assistance method based on neural signal modeling needs to be proposed to overcome the deficiencies of the prior art. Disclosure of Invention The application provides a myelin sheath repair auxiliary method based on neural signal modeling, which is used for realizing personalized auxiliary and dynamic optimization of myelin sheath regeneration repair. The application provides a myelin sheath repair auxiliary method based on nerve signal modeling, which comprises the following steps: Acquiring multi-modal nerve signal data of an individual to be rehabilitated, and preprocessing the multi-modal nerve signal data to acquire a preprocessing data set, wherein the multi-modal nerve signal data comprises an electroencephalogram signal, a functional magnetic resonance signal and a peripheral nerve electrophysiological signal; Performing an individualized time sequence modeling process based on the preprocessed data set, generating a time sequence feature representation representing a myelin sheath damage-repair dynamic process, the time sequence feature representation including at least a nerve conduction velocity recovery rate index, a phase consistency index, and a nerve pathway integrity index; Calculating an individualized baseline parameter through a dynamic baseline updating mechanism, and calibrating the time sequence characteristic representation by the individualized baseline parameter to obtain a calibrated time sequence characteristic representation; The method comprises the steps of taking a calibration time sequence characteristic representation as input, sending the calibration time sequence characteristic representation into a pre-trained deep learning reasoning engine comprising a circulation unit, a graph convolution unit and a migration learning module, and outputting a myelin repair index representing myelin repair degree and an unbalance risk score representing future repair abnormality probability through the deep learning reasoning engine, wherein the prior constraint parameter generated in the individuation time sequence modeling process is embedded into a loss structure and regularization of the deep learning reasoning engine in the training stage of the deep learning reasoning engine; Receiving a myelin repair index and an unbalance risk score, and generating an intervention parameter set, wherein the intervention parameter set comprises an electrical stimulation intervention parameter, a virtual rehabilitation training task difficulty and a rehabilitation training rhythm; and generating a control instruction according to the intervention parameter set, and executing the control instruction to configure a virtual rehabilitation training environment and an electric stimulation execution interface. The technical scheme provided by the application has the beneficial effects that: (1) By means of joint modeling of the electroencephalogram signals, the functional magnetic resonance signals and the peripheral nerve electrophysiological signals and generation of time sequence characteristic representation, dynamic processes of myelin damage and repair can be comprehensively described, and compared with single signa