A dynamic monitoring system for spinal cord injury reconstruction procedures

Abstract

The present application relates to the field of medical dynamic monitoring, in particular to a kind of dynamic monitoring system for spinal cord injury reconstruction process.The present application relates to the field of medical dynamic monitoring, in particular to a kind of dynamic monitoring system for spinal cord injury reconstruction process, including acquisition and fusion module, dynamic modeling module, function evaluation module and adaptive control module.The system acquires nerve and muscle electrophysiological signals through wearable surface electrode and flexible microelectrode array, combines calibration sample to correct timing error, realizes multimodal signal fusion.Construct topological order and timing dependence structure, use timing convolution network to extract latency, activation gradient and response dispersion.Recognize functional transition and abnormal compensatory behavior through function evaluation module, output nerve pathway response timing chart.Based on bayesian optimization strategy, dynamically adjust sampling frequency, feature weight and evaluation period, realize adaptive optimization of monitoring strategy.Thereby realize the accurate modeling and efficient tracking of nerve pathway function reconstruction process.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of medical dynamic monitoring, in particular to a dynamic monitoring system for spinal cord injury reconstruction process. BACKGROUND

[0002] The functional reconstruction after spinal cord injury is the core problem in the field of neural rehabilitation, and monitoring the recovery process of the patient's neural-muscular function is of great significance for rehabilitation evaluation and intervention decision-making.

[0003] The existing neuroelectrophysiological monitoring method mainly relies on intraoperative short-time SEP / MEP measurement, which is limited by the anesthetic state, operation experience and stimulation parameters, and is difficult to be used for long-term dynamic tracking after operation. Although electromyography acquisition has the advantage of non-invasiveness, the independence between channels is strong, and the response structure of the neural-muscular pathway is lacking, which leads to the inability to reconstruct the excitation-response path, and the evaluation result is mostly limited to the signal amplitude level. At the same time, the existing system generally lacks a pathway structure modeling and adaptive evaluation mechanism, which makes it difficult to support multi-cycle, cross-channel functional reconstruction analysis, and also cannot identify functional transition and compensation behavior, which limits its practicality in fine management of rehabilitation.

[0004] Therefore, the present application provides a dynamic monitoring system for spinal cord injury reconstruction process. SUMMARY

[0005] The present application provides a dynamic monitoring system for spinal cord injury reconstruction process, which aims to realize accurate evaluation and closed-loop feedback optimization of neural pathway reconstruction process.

[0006] To achieve the above purpose, the present application provides the following technical scheme:

[0007] The present application provides a dynamic monitoring system for spinal cord injury reconstruction process, which comprises:

[0008] The acquisition and fusion module adopts a wearable surface electrode and a flexible microelectrode array to acquire neural and muscle electrophysiological signals, synchronously samples based on a unified high-precision clock, constructs a compensation matrix combined with a calibration sample and adjusts the signal time sequence structure to correct the time sequence error, realizes registration and fusion of neural-muscular multi-modal signals and outputs a structure-unified time sequence matrix;

[0009] The dynamic modeling module is used for rearranging the pathway structure of the structure-unified time sequence matrix, constructing a cross-node time sequence dependent sequence according to the neural excitation sequence, extracting features of each node time sequence by using a time sequence convolution network, and outputting the latency, activation gradient and response dispersion of each node;

[0010] The function evaluation module constructs a time trend according to the node response characteristics in multiple monitoring periods, generates a neural pathway function reconstruction score, identifies function transition and abnormal compensatory behavior, and outputs a neural pathway response time sequence diagram.

[0011] The adaptive control module dynamically adjusts the sampling frequency, feature weight and evaluation period based on historical monitoring data and injury characteristics through a Bayesian decision update algorithm to realize adaptive optimization of the monitoring strategy.

[0012] As a preferred technical solution of the application, the calibration sample collection step comprises:

[0013] A functional electrical stimulation device is used to apply a control current to the target muscle group to induce the limb to perform a passive standard action;

[0014] Electrophysiological signals and muscle tension change data are collected during the induced action;

[0015] The collected data are used as calibration samples to construct a compensation matrix to correct the time sequence error.

[0016] As a preferred technical solution of the application, the step of correcting the time sequence error comprises:

[0017] The time difference between the neural activation time point and the corresponding electromyographic signal peak value of each collection channel in the calibration sample is extracted, and a channel response delay value is calculated;

[0018] The channel delay values are repeatedly sampled and averaged to construct a channel delay compensation matrix;

[0019] According to the channel delay, a linear translation method is used to perform time axis shift processing on the original sampling signals of each channel to align them to a unified response reference point;

[0020] The phase difference between the neural activation time point and the electromyographic response peak value is further extracted, and a sliding average method is used to adjust the sampling window position of the electromyographic signal;

[0021] Delay compensation and phase correction are combined to be applied to the original signal to realize time sequence alignment and individual standardization of neural-muscular multi-modal electrophysiological data.

[0022] As a preferred technical solution of the application, the step of rearranging the pathway structure of the structure-unified time sequence matrix comprises:

[0023] According to the neural-muscular anatomical topological order corresponding to the electrode channel, each node channel in the time sequence matrix is sorted according to the neural excitation path;

[0024] The activation start time of each node in the response process is identified, a multi-channel time sequence input with causal order is constructed, and the constructed multi-channel time sequence input is used as the modeling input of the time sequence convolution network.

[0025] As a preferred technical solution of the application, the step of extracting features from the time series by using the time series convolution network comprises:

[0026] normalizing and sliding window segmenting the node time series signals;

[0027] extracting short-time dynamic features by using a one-dimensional time series convolution network, and combining a residual structure to retain the signal change trend;

[0028] performing aggregation operation on the output feature vectors to obtain conduction latency, activation gradient and response dispersion index of each monitoring node.

[0029] As a preferred technical solution of the application, the step of generating the neural pathway function reconstruction score comprises:

[0030] extracting response amplitude, latency and pathway activation sequence features of each monitoring node under different monitoring periods;

[0031] calculating the change trend of each pathway index in the time dimension, and comparing with historical samples;

[0032] generating a reconstruction score value based on a statistical model, for quantifying the neural pathway function recovery efficiency.

[0033] As a preferred technical solution of the application, the step of identifying the functional transition comprises:

[0034] tracking the evolution of the response mode of each node in the neural pathway in the continuous monitoring period;

[0035] detecting the first occurrence of the activation path transfer, the significant improvement of response synchrony or the significant shortening of latency event;

[0036] and marking the above event as a trigger event of the functional transition, for determining the key functional inflection point in the rehabilitation process.

[0037] As a preferred technical solution of the application, the step of identifying the abnormal compensatory behavior comprises:

[0038] analyzing the occurrence of unexpected activation path in the neural pathway, the abnormal increase of activation intensity of the monitoring node response or the asymmetric diffusion;

[0039] identifying the potential compensatory pathway based on the time series change map, and judging whether it deviates from the normal activation logic combined with the prior structure model;

[0040] labeling the abnormal response pathway as a compensatory path, for assisting the adjustment of the rehabilitation strategy and the risk prompt.

[0041] As a preferred technical solution of the application, the adaptive optimization step comprises:

[0042] Collecting historical monitoring data and patient injury characteristics to construct a target function for optimizing parameter configuration;

[0043] Based on the Bayesian optimization framework, the sampling frequency, feature weight and evaluation cycle control parameters are modeled as variables;

[0044] The optimal solution is searched in the parameter space using the expected improvement criterion, and the current monitoring strategy is updated to realize closed-loop adjustment and adaptive optimization.

[0045] The beneficial effects of the present application are:

[0046] 1. The present application realizes the synchronous acquisition of neural and muscle multi-modal signals by fusing wearable surface electrodes and flexible microelectrode arrays, and extracts the inter-channel response delay and phase difference combined with the calibration sample, constructs the channel compensation matrix and sliding correction strategy, and unifies the time axis alignment and phase adjustment of each channel signal, completes the cross-channel error correction and standardization, provides a high consistency and personalized signal basis for subsequent modeling, and is significantly superior to the traditional independent channel acquisition and static compensation mode.

[0047] 2. The present application realizes continuous modeling, dynamic evaluation and real-time optimization of the functional state of the neural pathway through the collaborative design of the modeling module and the evaluation and control module, can identify the functional transition and abnormal compensation trend in time during the monitoring process, and adjust the sampling and evaluation strategy accordingly. Compared with the traditional scheme relying on single-cycle observation and static model, the present application can accurately capture the change inflection point of the pathway during the multi-cycle recovery process, and improves the response sensitivity and intervention adaptability of the neural reconstruction process. BRIEF DESCRIPTION OF DRAWINGS

[0048] The accompanying drawings are used to provide a further understanding of the present application, and constitute a part of the specification, together with the embodiments of the present application, to explain the present application, and do not constitute a limitation of the present application. In the drawings:

[0049] Figure 1 is a structural schematic diagram of a dynamic monitoring system for spinal cord injury reconstruction process of the present application;

[0050] Figure 2 is a flowchart of a dynamic monitoring system for spinal cord injury reconstruction process of the present application. DETAILED DESCRIPTION

[0051] The preferred embodiments of the present application are described below in conjunction with the accompanying drawings, and it should be understood that the preferred embodiments described herein are only used to illustrate and explain the present application, and do not limit the present application.

[0052] Example 1: as Figure 1As shown, the dynamic monitoring system for spinal cord injury reconstruction process comprises:

[0053] The acquisition and fusion module adopts wearable surface electrodes and flexible microelectrode array to acquire neural and muscle electrophysiological signals, synchronously samples based on a unified high-precision clock, constructs a compensation matrix combined with calibration samples and adjusts the signal time sequence structure to correct the time sequence error, realizes registration and fusion of neural-muscle multi-modal signals and outputs a time sequence matrix with unified structure.

[0054] Further, the calibration sample acquisition step comprises:

[0055] A functional electrical stimulation device is used to apply a control current to the target muscle group to induce the limb to perform a passive standard action.

[0056] The electrophysiological signals and muscle tension change data are acquired during the induced action.

[0057] The acquired data are used as calibration samples to correct the time sequence error by an individualized compensation method.

[0058] Further, the time sequence error correction step comprises:

[0059] The time difference between the neural activation time point and the corresponding electromyographic signal peak value of each acquisition channel in the calibration sample is extracted, and a channel response delay value is calculated.

[0060] The channel delay values are repeatedly sampled and averaged to construct a channel delay compensation matrix.

[0061] According to the channel delay, a time axis shift processing is performed on the original sampling signals of each channel in a linear translation manner to align them to a unified response reference point.

[0062] The phase difference between the neural activation time point and the electromyographic response peak value is further extracted, and a sliding average method is used to adjust the sampling window position of the electromyographic signal.

[0063] The delay compensation and phase correction are jointly applied to the original signal to realize time sequence alignment and individualized standardization of neural-muscle multi-modal electrophysiological data.

[0064] Specifically, the wearable surface electromyographic electrodes are arranged in the target muscle group area to acquire the surface electromyographic signals generated by the muscles; at the same time, flexible microelectrode arrays are arranged at key positions of the neural pathway to acquire the electrical signals generated by the neural activity. The above electrodes are connected to the synchronous acquisition device, and the system aligns the time sequences of all acquisition channels based on a unified high-precision clock to form a cross-modal, multi-channel original sampling data stream.

[0065] After the first use or repositioning of the electrode, the target muscle group is subjected to a preset intensity of control current by the functional electrical stimulation system, and a passive standardized motion response is induced. The electrophysiological signal and muscle tension change data are collected during the induction of the action, and the collected data are used as calibration samples. According to the time difference between the activation time of the neural signal and the peak response of the electromyographic signal after each electrical pulse in the calibration sample, the response delay value m of the neural-muscular pathway is extracted. The average value of 5 repeated samples of each channel is taken to form a channel delay compensation matrix. The phase difference Δφ between the start of electrical stimulation and the peak response is recorded at the same time, which is used for subsequent compensation.

[0066] According to the delay difference of each channel in the delay matrix M, a linear translation compensation function is set:

[0067]

[0068] Wherein is the original sampling signal, is the delay of the i-th channel, and the unit is ms.

[0069] After all the channel signals are processed by the function, they are aligned to the unified response reference point on the time axis, realizing the preliminary time registration of the cross-modal signals.

[0070] In order to correct the intention-response deviation problem caused by nerve injury, the collected neural activation time point is recorded as , and the peak point of the electromyographic signal is , and the phase difference Δφ is used to correct the time window of the electromyographic signal, and the sampling segment is moved forward or backward by the sliding average method to adjust the time sequence structure of the electromyographic response.

[0071] At the same time, the weight coefficient is calculated according to the average value of the response amplitude of each channel, and the amplitude weighted standardization is performed on the signal amplitude, and the output vector is:

[0072]

[0073] All channel signals processed by time sequence correction and amplitude weighting are classified according to the signal type. Let the number of neural signal channels be , the number of electromyographic signal channels be , and the number of sampling points of each channel be , two matrices are constructed respectively:

[0074] Neural signal matrix

[0075] Electromyographic signal matrix

[0076] ​According to the electrode layout and anatomical association, the two modal signals are channel-level registered. Each set of paired channels constitutes a dual-channel time sequence block:

[0077]

[0078] All the paired small matrices Spliced in the row direction in the anatomical order, a unified fusion matrix is constructed:

[0079]

[0080] By wearing a surface electrode and a flexible microelectrode array to cooperatively collect neural and muscle electrophysiological signals, and combining a high-precision clock and a personalized compensation strategy, accurate alignment of different modal signals on the time axis is realized, providing a structural unified and time consistent data basis for subsequent modeling and functional evaluation.

[0081] The dynamic modeling module is configured to rearrange the path structure of the structural unified time sequence matrix, construct a cross-node time sequence dependency sequence according to the neural excitation order, extract features of the time sequence of each node by using a time convolution network, and output a latency, an activation gradient and a response dispersion of each node.

[0082] Further, the step of constructing the node time sequence dependency sequence comprises:

[0083] Rearranging the path structure of the structural unified time sequence matrix, and sorting the node channels according to the neural excitation path order;

[0084] Extracting the activation start time in the response of each node and the delay relationship between nodes;

[0085] Constructing an input sequence with topological constraints and time sequence dependencies for the time convolution network to model.

[0086] Specifically, the fusion matrix output by the acquisition and fusion module , wherein each row corresponds to a time sequence electrophysiological signal of an electrode channel. For each monitoring node, the complete time sequence data thereof is directly extracted, i.e., the th node signal is represented as without any dimension reduction or truncation processing, ensuring the integrity and continuity of the time sequence signal.

[0087] According to the anatomical structure of the nervous system and the neural excitation path, the channels corresponding to different modal signals are rearranged according to their physiological positions and conduction orders. For example, the neural channel signals located in the upper segment of the spinal cord are arranged first, and then the corresponding distal electromyographic channel signals are sequentially arranged. The rearranged signal sequence constitutes a two-dimensional matrix with the number of rows corresponding to the number of nodes and the number of columns corresponding to the number of time points:

[0088]

[0089] wherein denotes the th rearranged node index, .

[0090] For each rearranged channel , its activation starting time point is extracted. The activation time point is determined by the following method:

[0091] Firstly, the sliding window segmentation is performed on each node time series, and the average energy of each segment is calculated. The energy calculation formula is:

[0092]

[0093] wherein is the window length, is the sliding starting time point.

[0094] A global energy threshold is defined, and the activation time point is the time point at which is first satisfied.

[0095] Subsequently, the timing delay between adjacent channels is calculated:

[0096]

[0097] Finally, the rearranged timing matrix of the channel , the corresponding node activation time point and the delay sequence are taken as inputs to construct a structured input sequence for modeling the timing convolution network.

[0098] The input structure retains the original multi-channel time series information while embedding the topological constraints of neural excitation pathways and the timing dependence between nodes, which helps the subsequent model to accurately extract features such as pathway latency, activation gradient and response dispersion.

[0099] Further, the step of extracting features from the time series using the timing convolution network comprises:

[0100] normalizing and sliding window segmenting the node time series signal;

[0101] extracting short-time dynamic features using a one-dimensional timing convolution network, and combining a residual structure to retain the signal change trend;

[0102] performing aggregation operation on the output feature vector to obtain the conduction latency, activation gradient and response dispersion index of each monitoring node.

[0103] Specifically, according to the rearranged time sequence matrix , combined with the activation starting time point of each node and the activation delay sequence between adjacent nodes , a three-channel input tensor is constructed , and the data of each channel is as follows:

[0104] Channel 1: The original time sequence signal of each node is normalized to the interval [0, 1] and used as the main input;

[0105] Channel 2: A certain time window after the activation starting time of each node is marked as "1", and other time points are marked as "0" to form a binary mask sequence, representing the significant response segment;

[0106] Channel 3: For each node, fill in the activation delay value relative to the upstream node , which is used as a reference quantity for topological-time sequence dependence, and is copied in the time dimension to form a constant sequence.

[0107] The above input tensor is sliced in the time dimension according to a fixed window length to generate multiple overlapping time sequence segments. Each segment retains the original multi-channel structure, and all segments are integrated as batch input to meet the input requirements of the time sequence convolution network. In order to ensure the stability of the input data, the time sequence signal is normalized before sliding window, and the abnormal value is truncated by threshold detection. When the tail of the segment is insufficient, zero padding is used to fill it up.

[0108] The processed time sequence is used as the input of a one-dimensional time sequence convolution network to extract the time-frequency dynamic features of the signal within each time window while retaining the overall trend of the signal. The TCN output corresponds to the feature vector of each sliding window, reflecting the local activation pattern and dynamic changes.

[0109] For multiple sliding window feature vectors of the same monitoring node, they are arranged in time sequence to form a feature matrix , where is the number of sliding windows, is the dimension of a single sliding window feature vector. Based on the aggregated features, the following quantitative indicators are calculated:

[0110] Conduction latency: for the feature dimension reflecting the activation intensity in , set a threshold ; find the first sliding window that satisfies the feature value greater than in time sequence, and the index is the time point corresponding to the conduction latency ;

[0111] Activation gradient: in The characteristic dimension reflecting the activation intensity is selected, and the first-order difference on the time sequence is calculated The activation gradient is defined as the average value of the difference sequence, reflecting the overall activation rate change of the node in the monitoring period.

[0112] Response dispersion: the value sequence of the same characteristic dimension on the time axis The standard deviation is calculated as the response dispersion index, indicating the stability of the activation state of the node in the entire period.

[0113] The adaptability of signal preprocessing is improved through normalization and sliding window processing, short-term dynamic features are extracted using one-dimensional time series convolution network, and residual structure is used to retain overall trends, finally realizing accurate quantification of key response features of each monitoring node, providing reliable data foundation for subsequent function evaluation.

[0114] The function evaluation module constructs the time trend according to the node response features in multiple monitoring periods, generates a neural pathway function reconstruction score, identifies functional transition and abnormal compensatory behavior, and outputs a neural pathway response time series graph.

[0115] Further, the step of generating the neural pathway function reconstruction score comprises:

[0116] Extract the response amplitude, latency and pathway activation sequence features of each monitoring node in different monitoring periods.

[0117] Calculate the change trend of each pathway index in the time dimension and compare it with the historical sample.

[0118] Based on the statistical model, a reconstruction score value is generated to quantify the recovery efficiency of the neural pathway function.

[0119] Specifically, in each preset monitoring period, the results output by the aforementioned dynamic modeling module are called to extract the following three types of feature indicators of all monitoring nodes: conduction latency, activation gradient and response dispersion.

[0120] The three types of indicators of each monitoring node in the current period are respectively arranged into vectors , , , representing the conduction latency, activation gradient and response dispersion in the first period, respectively.

[0121] In a plurality of continuous rehabilitation monitoring periods, the evolution trend of each type of index on the time axis is constructed respectively:

[0122] The latency value of each monitoring node in is linearly regressed and fitted to obtain the slope of the change of the latency of the node.

[0123] For each element of , evaluate its promotion speed in periods;

[0124] For each element of , calculate the change range and mean fluctuation amplitude to determine whether the response stability has improved.

[0125] To obtain the quantitative results of the overall neural pathway function recovery degree, set the target change direction of each type of index: latency reduction, gradient enhancement, and dispersion reduction, and define the expected improvement interval.

[0126] Standardize the trend values of each node on the three types of indexes to the interval, and calculate the overall neural pathway reconstruction score value by weighted summation according to the preset weights.

[0127] The change trend of the score value is used to evaluate the patient's rehabilitation progress in each period, and a higher value indicates that the neural pathway activation is closer to the healthy pathway state, serving as the basis for adjusting the rehabilitation plan and evaluating the stage effect.

[0128] Further, the step of identifying the functional transition comprises:

[0129] In the continuous monitoring period, track the evolution of the response pattern of each node in the neural pathway;

[0130] Detect the first occurrence of activation path transfer, significant improvement in response synchrony, or significant reduction in latency event;

[0131] And mark the above events as trigger events of functional transition, which are used to determine the key functional inflection point in the rehabilitation process.

[0132] In the set continuous monitoring period, extract the conduction latency, activation gradient, and response dispersion of each node, forming a period sequence ;

[0133] For the conduction latency , construct a time series for each monitoring node, calculate its change rate using the first-order difference method, and if a sudden drop occurs in a certain period , the drop amplitude exceeds the set threshold , it is determined as a latency transition;

[0134] Compare the activation path sequence: in periods and , construct the corresponding pathway paths and If the directed connection relationship changes, i.e. path jump or secondary path early activation, it is marked as an activation path transition event;

[0135] In addition, the Pearson correlation coefficient matrix between each cycle node response is calculated, and if the average synchrony significantly improves, in this embodiment, if the overall correlation coefficient improves by more than 0.3, it is determined that the response synchrony jumps;

[0136] The cycle that meets any condition, i.e. the function transition trigger cycle, is identified as a key turning point, and the corresponding node and time point are marked in the response timing diagram.

[0137] Further, the step of identifying abnormal compensatory behavior comprises:

[0138] Analyzing the occurrence of unexpected activation path in neural pathway, monitoring the abnormal increase of activation intensity or asymmetric diffusion of node response;

[0139] Based on the timing change map, potential compensatory pathways are identified, and whether they deviate from the normal activation logic is judged based on the prior structure model;

[0140] The abnormal response path is marked as a compensatory path, which is used to assist in adjusting the rehabilitation strategy and risk prompt.

[0141] For each cycle, analyze the new activation node set outside the neural activation path , and compare it with the predefined normal path model, if There is no structural connection between the node and the main path, or the activation delay is abnormally shortened (such as less than 5ms), which is marked as an unexpected activation path;

[0142] The activation gradient of the node value is standardized by Z-score, and if the activation gradient value of a single node in a path exceeds 2 times the standard deviation of the mean value, it is considered as abnormal increase of activation intensity;

[0143] If multiple nodes are activated in a spatially aggregated manner, resulting in a significant expansion of the response dispersion on one side, while the other side is not responding or responding weakly, it is identified as asymmetric diffusion of response;

[0144] The above abnormal activation path is constructed as an independent path, and whether it deviates from the neural-muscle physiological pathway logic is judged based on the structure model, if it deviates, the path is marked as an abnormal compensatory path, and is used to prompt potential risks.

[0145] Finally, the main activation path in each cycle is plotted with time period as horizontal axis and node anatomical position as vertical axis.

[0146] For the identified functional transition events, mark "transition points" on the corresponding cycle nodes;

[0147] For the identified abnormal compensatory pathways, use dashed lines or special colors to indicate their first appearance cycle and pathway structure;

[0148] Form an interactive neural pathway response timing diagram, provide clinicians with rehabilitation inflection point judgment basis, and serve as a visual reference tool for rehabilitation intervention strategy adjustment.

[0149] Adapt the control module to dynamically adjust the sampling frequency, feature weight and evaluation period based on historical monitoring data and injury characteristics, and realize adaptive optimization of the monitoring strategy through Bayesian decision update algorithm.

[0150] Further, the adaptive optimization step includes:

[0151] Collect historical monitoring data and patient injury characteristics to construct a target function for optimizing parameter configuration;

[0152] Based on the Bayesian optimization framework, variable modeling is performed on the sampling frequency, feature weight and evaluation period control parameters;

[0153] Use the expected improvement criterion to search for the optimal solution in the parameter space, and update the current monitoring strategy to realize adaptive optimization of the monitoring strategy.

[0154] The system first constructs a target function to evaluate the advantages and disadvantages of the current monitoring strategy based on historical monitoring data and patient injury characteristics . Among them, the input parameters correspond to the sampling frequency , feature weight and evaluation period respectively; the output result is a comprehensive score value representing the overall monitoring effect under the current parameter combination. The score function form is as follows:

[0155]

[0156] : The reconstruction score of the current cycle output by the neural pathway function evaluation module;

[0157] : The confidence interval width of the continuous evaluation result, reflecting the stability of the index;

[0158] : The resource consumption estimate (acquisition time, number of electrodes used, etc.) of the system under the current parameter combination;

[0159] : The weight coefficient calculated based on the normalization of the score function partial derivative, which is as follows:

[0160]

[0161]

[0162] This setting ensures that the scoring function can reflect the impact of different factors on the effectiveness of rehabilitation monitoring strategies in real time.

[0163] The system models the objective function based on the existing historical data sample set using Gaussian process regression in the parameter space to obtain the mean and variance of each set of parameter combinations to estimate their potential monitoring effectiveness. The expected improvement criterion is used as the sampling strategy to search for the next set of optimal parameters in the parameter space that can maximize the improvement of monitoring effectiveness

[0164] . After the optimal solution is determined, the system automatically updates the sampling frequency, feature weighting coefficients, and evaluation execution period for the next monitoring cycle.

[0165] After each monitoring cycle is completed, the current parameter combination and evaluation results are added to the sample set and used for the next round of Bayesian model updating, forming a closed-loop mechanism for continuous iterative optimization. To prevent frequent changes in the strategy, a parameter change threshold is also set, which means that the strategy will only be updated when the improvement exceeds the set threshold.

[0166] Example Two:

[0167] In a patient with spinal cord injury admitted to a certain rehabilitation center, a dynamic monitoring system for spinal cord injury reconstruction process is applied, as shown in

[0168] , and the following is the actual deployment and use process: Figure 2 First, a flexible microelectrode array is placed on both sides of the patient's lumbar spine to collect neural electrical signals from the surrounding neural pathways of the spinal cord injury area in real time. At the same time, wearable surface electrodes are attached to key motor muscle groups such as the bilateral rectus femoris, iliopsoas, and tibialis anterior muscles to collect electromyographic signals. All collection channels are connected to a unified high-precision collection interface, and the system automatically completes cross-channel signal synchronization.

[0169]

[0170] ​Considering the patient's inability to actively complete standard actions, a clinical rehabilitation therapist cooperates with the electrical stimulation device to apply controllable current to the areas innervated by the femoral nerve and the sciatic nerve, inducing passive actions such as knee extension and ankle dorsiflexion, and synchronously collecting neural signals, electromyographic signals, and muscle tension changes during the actions. The system records the activation delay time, signal intensity, and response form under multiple rounds of electrical stimulation, constructs an individual channel response delay parameter and signal phase difference model, and automatically generates a corresponding compensation matrix. Subsequently, the system adjusts the original signals for delay, phase correction, and channel weight normalization, and outputs a set of structurally unified and time-aligned neural-muscular fusion signal matrices.

[0171] After data fusion is completed, the system rearranges all channel signals according to the "spinal cord nerve - nerve root - proximal muscle group - distal muscle group" conduction sequence based on patient anatomical information, and constructs a structural topology sequence of neural excitation-muscle response. At the same time, the system scans each channel signal to identify the starting time point of the first continuous activation waveform, and calculates the response delay between channels to generate a time-dependent relationship diagram of the excitation path.

[0172] The system inputs the rearranged signal matrix and activation delay structure into a pre-trained time convolution network model, extracts conduction latency, activation trend change rate, and response stability of each node, and other feature indicators. After the model output, the current cycle result is compared with the patient's previous cycle data. If the latency is significantly shortened, the activation sequence is switched, or a new signal channel activation path appears in the pathway, the system will mark the cycle as a "functional transition" that may exist; if some muscle groups abnormally discharge without being excited by the superior nerve, it is recorded as an "abnormal compensation" risk, and the doctor is prompted to review.

[0173] At the same time, the system automatically retrieves the past 4 monitoring data, and combines the patient's post-injury initial classification, surgical method, and recovery speed, etc. label data, uses Bayesian optimization strategy to adaptively adjust the monitoring cycle, feature selection priority and sampling channel, so that when the system enters the next cycle of monitoring, it can focus on more representative nodes and features, improving the evaluation efficiency and accuracy.

[0174] Finally, the patient experienced two typical activation path transition events at the 3rd and 6th weeks after surgery, respectively, in which the activation latency of the rectus femoris and tibialis anterior muscle was shortened by more than 45%, the neural pathway synchronization was enhanced, and the functional score curve output by the system gradually increased, prompting the doctor to decide to conduct passive walking assistance training of the lower limbs in advance, and after the system prompted the existence of abnormal compensation path at the 8th week, the rehabilitation strategy was adjusted in time, avoiding the risk of muscle overload during training.

[0175] The embodiment two fully embodies the application effect of the system in the real clinical environment to realize the whole process application effect from data acquisition, standardized fusion, dynamic modeling, function identification to monitoring strategy optimization, and proves the practicability and innovativeness in the spinal cord injury reconstruction rehabilitation evaluation.

[0176] Finally, it should be noted that: the above only for the preferred embodiments of the present application, and is not intended to limit the present application, although the foregoing detailed description of the application is made with reference to the foregoing embodiments, for those skilled in the art, it still can be modified, or the equivalent replacement of the technical solutions recorded in the foregoing embodiments. Any modification, equivalent replacement, improvement, etc. within the spirit and principles of the present application shall be included within the scope of the present application.

Claims

1. A dynamic monitoring system for spinal cord injury reconstruction procedures, characterized by, The method comprises the following steps: a collection and fusion module, which adopts a wearable surface electrode and a flexible microelectrode array to collect neural and muscle electrophysiological signals, synchronously samples based on a unified high-precision clock, constructs a compensation matrix based on a calibration sample, adjusts the time sequence structure to correct the time sequence error, realizes the registration and fusion of neural-muscle multi-modal signals, and outputs a structure-unified time sequence matrix; a dynamic modeling module, which is used for rearranging the path structure of the structure-unified time sequence matrix, constructing a cross-node time sequence dependence sequence according to a neural excitation sequence, extracting features of each node time sequence by using a time sequence convolution network, and outputting a latency, an activation gradient and a response dispersion of each node; a function evaluation module, which constructs a time sequence trend according to the response characteristics of the nodes in multiple monitoring periods, generates a neural path function reconstruction score, identifies a function transition and an abnormal compensatory behavior, and outputs a neural path response time sequence diagram; an adaptive control module, which dynamically adjusts a sampling frequency, a feature weight and an evaluation period by using a Bayesian decision update algorithm based on historical monitoring data and injury characteristics, and realizes adaptive optimization of a monitoring strategy; the identification of the function transition and the abnormal compensatory behavior comprises the following steps: tracking evolution of response modes of each node in the neural path in continuous monitoring periods; detecting a first-occurred activation path transfer, a significant improvement in response synchrony or a significant shortening of latency event; and marking the above event as a trigger event of the function transition, which is used for determining a key function inflection point in a rehabilitation process; analyzing occurrence of an unexpected activation path in the neural path, abnormal increase or asymmetric diffusion of an activation intensity of a monitoring node response; identifying a potential compensatory path based on a time sequence change map, and judging whether the compensatory path deviates from a normal activation logic based on a prior structure model; and marking the neural path with abnormal activation as a compensatory path, which is used for assisting adjustment and risk prompting of a rehabilitation strategy.

2. A dynamic monitoring system for spinal cord injury reconstruction process as claimed in claim 1 wherein, the calibration sample collection step comprises the following steps: applying a control current to a target muscle group by using a functional electrical stimulation device to induce a limb to perform a passive standard action; collecting electrophysiological signals and muscle tension change data during the induced action; using the electrophysiological signals and the muscle tension change data as calibration samples to construct a compensation matrix and correct the time sequence error.

3. A dynamic monitoring system for spinal cord injury reconstruction process as claimed in claim 1 wherein, the correction of the time sequence error comprises the following steps: extracting a time difference between a neural activation time point and a corresponding electromyogram peak value of each collection channel in the calibration sample, and calculating a channel response delay value; repeatedly sampling and averaging the delay values of the channels to construct a channel delay compensation matrix; performing time axis shift processing on original sampling signals of the channels in a linear translation manner according to the channel delay, so that the original sampling signals are aligned to a unified response reference point; further extracting a phase difference between the neural activation time point and the electromyogram response peak value, and adjusting a sampling window position of the electromyogram signal in a sliding average manner; and jointly applying delay compensation and phase correction to the original signals to realize time sequence alignment and individual standardization of neural-muscle multi-modal electrophysiological data.

4. The dynamic monitoring system for spinal cord injury reconstruction process as claimed in claim 1 wherein, the step of rearranging the path structure of the structure-unified time sequence matrix comprises the following steps: sorting each node channel in the time sequence matrix according to a neural-muscle anatomical topology sequence corresponding to an electrode channel; The activation start time of each node in the response process is identified, a multi-channel time sequence input with a causal order is constructed, and the modeling input of the time sequence convolution network is constructed.

5. The dynamic monitoring system for spinal cord injury reconstruction process as claimed in claim 1 wherein, The step of extracting features of the time sequence by using the time sequence convolution network comprises: normalizing and sliding window segmenting the node time sequence signal; extracting short-time dynamic features by using a one-dimensional time sequence convolution network, and combining a residual structure to retain the signal change trend; performing aggregation operation on the output feature vector to obtain the conduction latency, activation gradient and response dispersion index of each monitoring node.

6. The dynamic monitoring system for spinal cord injury reconstruction process as claimed in claim 1 wherein, The step of generating the neural pathway function reconstruction score comprises: extracting the response amplitude, latency and pathway activation order features of each monitoring node in different monitoring periods; calculating the change trend of each pathway index in the time dimension and comparing with the historical sample; generating a reconstruction score value based on a statistical model, which is used to quantify the neural pathway function recovery efficiency.

7. The dynamic monitoring system for spinal cord injury reconstruction process as claimed in claim 1 wherein, The adaptive optimization comprises: collecting historical monitoring data and patient injury characteristics to construct a target function for optimizing parameter configuration; based on a Bayesian optimization framework, variable modeling is performed on the sampling frequency, feature weight and evaluation period control parameters; an optimal solution is searched in the parameter space by using an expected improvement criterion, and the current monitoring strategy is updated to realize closed-loop adjustment and adaptive optimization.

Images