Intelligent Assessment Method for Post-Stroke Cognitive Impairment Based on Dynamic Reconstruction of Brain Networks

Abstract

The present invention relates to the field of healthcare informatics technology and discloses an intelligent assessment method for post-stroke cognitive impairment based on dynamic brain network reconstruction. The method comprises: acquiring multimodal brain imaging data from patients, dynamically compensating for hemodynamic delays and white matter conduction velocity differences through a neurovascular coupling model, and achieving precise spatiotemporal alignment of functional and structural signals; constructing a dynamic entropy change model of a spiking neural network based on the aligned data, encoding changes in functional connectivity strength into pulse emission sequences, and extracting compensatory features such as gamma-band synchronization enhancement; employing an energy optimal transmission model to constrain the rationality of brain network reconstruction and generate a compensatory heat map; and achieving real-time edge computing through pulse timing compression coding. The present invention avoids the limitations of traditional static analysis and, through pulse coding technology, enables the algorithm to achieve microsecond response on embedded devices, providing a reliable stroke rehabilitation assessment tool for primary hospitals.

Description

Technical Field

[0001] The present invention relates to an intelligent assessment method for post-stroke cognitive impairment based on dynamic reconstruction of brain networks, and belongs to the technical field of medical care informatics. Background Art

[0002] In the rehabilitation assessment of stroke patients, dynamically capturing the spatiotemporal evolution of brain network compensation pathways based on multimodal medical data is a key step in improving the accuracy of intelligent assessment systems. Current clinical assessment methods primarily rely on static functional connectivity analysis techniques, such as fMRI data processing based on a fixed time window, and independent assessments combined with structural imaging (such as DTI white matter tracing). While this assessment method can pinpoint the damaged area, it suffers from the following technical bottlenecks:

[0003] Inadequate capture of dynamic compensation features: Static analysis cannot resolve the millisecond-level oscillation characteristics triggered by synaptic plasticity (such as synchronous enhancement of the gamma band), resulting in missed detection of activation signals for the compensation pathway. The industry generally compensates for sensitivity by extending the sampling time window, but this blurs timing resolution and can lead to misjudgment of the compensation window.

[0004] 2. Multimodal spatiotemporal mismatch: The hemodynamic delay of fMRI and the difference in white matter conduction velocity of DTI are not corrected. The independent processing of functional and structural data leads to cross-modal signal misalignment and the generation of false connections. Existing solutions usually use simple time interpolation alignment, but it is difficult to adapt to the asynchronous characteristics under non-uniform neural activity.

[0005] 3. Computing resource dependence limits grassroots applications: Dynamic network modeling relies on the high computing power of GPU clusters. Traditional algorithms cannot perform real-time analysis on embedded devices such as mobile MRI workstations in grassroots hospitals. The industry has attempted to alleviate computing pressure by reducing model complexity, but this has sacrificed the physiological rationality of compensatory heat maps.

[0006] Based on this, existing assessment methods suffer from high rates of misjudgment of compensatory pathways, poor real-time performance, and difficulties in grassroots deployment. In recent years, while some studies have attempted to introduce dynamic graph networks or lightweight models, it remains difficult to balance timing accuracy with physiological plausibility. Furthermore, the lack of modeling of metabolic cost constraints can easily lead to results that violate the laws of neurovascular coupling. Therefore, achieving dynamic and precise alignment of multimodal data, millisecond-level analysis of compensatory oscillation characteristics, and real-time, reliable assessment in low-computing environments have become technical challenges awaiting breakthrough in this field. Summary of the Invention

[0007] The present invention provides an intelligent assessment method for post-stroke cognitive impairment based on dynamic reconstruction of brain networks. Its main purpose is to solve the problems of insufficient capture of dynamic compensation features, multimodal spatiotemporal mismatch and computing resource dependence that limit grassroots applications.

[0008] To achieve the above objectives, the present invention provides an intelligent assessment method for post-stroke cognitive impairment based on dynamic brain network reconstruction, comprising the following steps:

[0009] Step 1: Acquire multimodal brain imaging data of a post-stroke patient, wherein the multimodal brain imaging data at least includes blood oxygen level-dependent BOLD signal data from functional magnetic resonance imaging (fMRI) and white matter fiber tracking data from diffusion tensor imaging (DTI);

[0010] Step 2: For multimodal brain imaging data, a delay adaptive alignment algorithm is used to dynamically match BOLD signal data and white matter fiber tracking data on a non-uniform time axis. Specifically, the algorithm includes: building a neurovascular coupling model, dynamically estimating and correcting the hemodynamic delay based on the neurovascular coupling model. and the cross-modal time axis misalignment caused by the difference in white matter fiber conduction velocity, generating a spatiotemporally aligned multimodal brain imaging data stream; Estimated based on the following formula:

[0011] ,

[0012] in, is the neurovascular coupling coefficient, is the length of white matter fibers, is the white matter fiber conduction velocity;

[0013] Step 3: Based on the spatiotemporally aligned multimodal brain imaging data stream, a spiking neural network (SNN) dynamic entropy change model was constructed to encode the temporal oscillation characteristics of post-stroke brain network reconstruction. Specifically, the model encodes the changes in dynamic functional connectivity strength between brain network nodes into SNN spike sequences. By analyzing the membrane potential oscillation characteristics of the spike sequences, the temporal entropy change features that characterize the stability of the compensatory pathway and the efficiency of resource consumption are extracted. The temporal entropy change features include a synchronous enhancement pattern in the gamma band (30-80 Hz).

[0014] Step 4: Based on the spatiotemporally aligned multimodal brain imaging data stream, an energy-optimal transmission model is constructed to constrain the rationality of dynamic brain network reconstruction after stroke. Specifically, the dynamic brain network reconstruction process is modeled as an energy-optimal transmission problem, the neurovascular coupling equation is introduced as a constraint, the reconstruction efficiency of key hub nodes and the rationality of resource allocation are calculated, and a compensation heat map is generated to reflect the resource allocation of the compensation pathway.

[0015] In step 5, pulse timing compression coding (STCE) technology is used to perform sparse processing on the pulse emission sequence of the SNN to generate a sparse event stream, and the temporal entropy change characteristics and compensation heat map are reconstructed in real time based on the sparse event stream on the edge embedded chip to achieve rapid and intelligent assessment of post-stroke cognitive impairment.

[0016] In a preferred embodiment, in step 2, the delay adaptive alignment algorithm further comprises: iteratively optimizing parameters of the neurovascular coupling model to minimize the residual error between the BOLD signal data and the white matter fiber tracking data on the time axis.

[0017] In a preferred embodiment, in step 3, the temporal entropy change model quantifies the transient response of synaptic weight adjustment during the compensation process by analyzing the frequency changes and phase synchronization of membrane potential oscillations, and identifies specific compensatory oscillation patterns associated with cognitive function recovery.

[0018] In a preferred embodiment, in step 4, the energy optimal transfer model further introduces a metabolic cost function, which is constructed based on the energy consumption relationship between neuronal activity and hemodynamics, and is used to inhibit the generation of compensatory pathways that are inconsistent with physiological laws and have excessive energy consumption.

[0019] In a preferred embodiment, in step 5, the pulse timing compression coding technology adopts an adaptive sparsification strategy based on dynamic reconstruction of pulse density, specifically including: real-time monitoring of the pulse computing density of each brain area of the brain network, when the pulse density of the key compensatory brain area exceeds the preset threshold, triggering the local pulse cluster merging mechanism, integrating the high-frequency micropulses into low-frequency macropulses and retaining their oscillation phase information; implementing event-driven truncation coding on the pulse sequence of non-critical brain areas, filtering low information entropy pulses through synaptic weight thresholds, and retaining only key conduction events of cross-hemispheric interactions; deploying a pulse buffer pool dynamic scheduling algorithm at the edge end, and dynamically adjusting the parallel processing queue depth of the pulse cluster according to the real-time remaining computing resources to ensure the real-time performance of the evaluation.

[0020] In a preferred embodiment, the local pulse cluster merging mechanism embeds a timestamp vector carrying the original oscillation phase when generating macropulses to ensure the integrity of the oscillation synchronization analysis in subsequent SNN modeling.

[0021] In a preferred embodiment, in the event-driven truncation coding, the synaptic weight threshold is dynamically adjusted based on the brain network reconstruction data and cognitive function recovery of historical stroke patients.

[0022] In a preferred embodiment, the method further includes the steps of: using a machine learning model to predict the cognitive function recovery trajectory of post-stroke patients based on the temporal entropy change characteristics and the compensatory heat map, and generating personalized rehabilitation suggestions.

[0023] In a preferred embodiment, the machine learning model includes but is not limited to a support vector machine, a neural network or a regression model.

[0024] In a preferred embodiment, after step 5, the method further includes the step of visually displaying the temporal entropy change characteristics and the compensation heat map.

[0025] Compared with the background technology problems, the beneficial effects of the present invention are:

[0026] 1. Dynamically compensate for time axis offsets in multimodal data through a neurovascular coupling model, eliminating cross-interference between hemodynamic delays and white matter conduction velocity differences, ensuring precise spatiotemporal matching of functional and structural signals. Based on this, the system utilizes the unique pulse sequence encoding mechanism of spiking neural networks to convert transient fluctuations in dynamic functional connectivity strength into resolvable membrane potential oscillation patterns, effectively capturing compensatory pathway initiation signals that are difficult to identify using traditional static analysis methods, such as the initial activation characteristics of synaptic plasticity. This deep coupling of spatiotemporal calibration and dynamic encoding enables the system to extract key compensatory patterns, such as Gamma band synchronization enhancement, from millisecond-level neural oscillations, providing clinically relevant evidence for earlier intervention timing.

[0027] 2. The brain network reconstruction process is modeled as an optimal energy transmission problem. The physiological rationality of the compensatory path is constrained by the neurovascular coupling equation, which suppresses the risk of false functional connections caused by ignoring metabolic costs in traditional algorithms. On this basis, the dynamic entropy change characteristics are sparsely processed using pulse timing compression coding technology. Through the pulse density adaptive resource scheduling strategy, the computational priority of key compensatory brain areas is achieved on the embedded chip. This collaborative design of energy constraint and hardware adaptation ensures that the compensatory thermal map conforms to the laws of cerebral hemodynamics, and enables the evaluation system to maintain microsecond response in a low computing power environment, breaking the real-time bottleneck of primary medical scenarios, realizing the deep integration of brain imaging physiological parameters and computational models, and improving the intelligence level of health data processing in edge computing.

[0028] 3. Imitating the regulatory principles of glial cells on synaptic activity, a local pulse cluster merging and phase retention mechanism is designed to compress the computational load while fully preserving the oscillation synchronization information. Further combined with the metabolic cost dynamic decision-making model, computing resources are allocated preferentially to clinical high-value compensatory feature analysis such as cross-frequency coupling, achieving accurate capture of key biomarkers and intelligent filtering of redundant data. This pulse encoding strategy is linked with metabolic value assessment, allowing the system to maintain low-power operation while avoiding the feature confusion and phase distortion problems caused by full-band calculations in traditional methods, thereby enhancing the automatic decoding of medical data and clinical feature extraction capabilities through intelligent pulse processing mechanisms.

[0029] 4. Through the organic integration of the pulse neural network dynamic entropy change model and the machine learning algorithm, a multi-dimensional correlation analysis is performed on the Gamma band synchronization enhancement slope of the temporal entropy change feature and the resource allocation pattern of the compensatory heat map, automatically identifying the compensatory oscillation pattern that is strongly correlated with cognitive function recovery, avoiding the excessive reliance of traditional evaluation methods on experience, and providing explainable decision support for the formulation of personalized rehabilitation plans by quantifying the stability of the compensatory path and the efficiency of resource utilization, significantly improving the accuracy and timeliness of rehabilitation interventions. Under the constraints of the neurovascular coupling equation, the compensatory heat map generated by the optimal energy transfer model can accurately reflect the rationality of resource redistribution between brain regions, avoiding the algorithm from generating connection patterns that violate physiological laws. At the same time, the pulse buffer pool dynamic scheduling algorithm automatically adjusts the parallel processing depth based on real-time computing resources to ensure the stable output of clinical-level evaluation results on commonly used equipment in grassroots hospitals. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 Schematic diagram of the delayed adaptive alignment processing flow of brain image data according to the present invention.

[0031] Figure 2 This is a flowchart of the spatiotemporal alignment of multimodal brain imaging data and spiking neural network modeling in the present invention.

[0032] Figure 3 This is a flow chart of the multimodal brain image alignment data calibration based on dynamic compensation of the present invention.

[0033] Figure 4 This is a flow chart of pulse density sparse compression coding and sparse event stream generation according to the present invention.

[0034] The purpose, features and advantages of the present invention will be further described with reference to the accompanying drawings and in conjunction with the embodiments. DETAILED DESCRIPTION

[0035] It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0036] The present invention provides an intelligent assessment method for post-stroke cognitive impairment based on dynamic brain network reconstruction, comprising the following steps:

[0037] Step 1: Acquire multimodal brain imaging data of a post-stroke patient, wherein the multimodal brain imaging data at least includes blood oxygen level-dependent BOLD signal data from functional magnetic resonance imaging (fMRI) and white matter fiber tracking data from diffusion tensor imaging (DTI);

[0038] Step 2: For multimodal brain imaging data, a delay adaptive alignment algorithm is used to dynamically match BOLD signal data and white matter fiber tracking data on a non-uniform time axis. Specifically, the algorithm includes: building a neurovascular coupling model, dynamically estimating and correcting the hemodynamic delay based on the neurovascular coupling model. and the cross-modal time axis misalignment caused by the difference in white matter fiber conduction velocity, generating a spatiotemporally aligned multimodal brain imaging data stream; Estimated based on the following formula:

[0039] ,

[0040] in, is the neurovascular coupling coefficient, is the length of white matter fibers, is the white matter fiber conduction velocity;

[0041] Step 3: Based on the spatiotemporally aligned multimodal brain imaging data stream, a spiking neural network (SNN) dynamic entropy change model was constructed to encode the temporal oscillation characteristics of post-stroke brain network reconstruction. Specifically, the model encodes the changes in dynamic functional connectivity strength between brain network nodes into SNN spike sequences. By analyzing the membrane potential oscillation characteristics of the spike sequences, the temporal entropy change features that characterize the stability of the compensatory pathway and the efficiency of resource consumption are extracted. The temporal entropy change features include a synchronous enhancement pattern in the gamma band (30-80 Hz).

[0042] Step 4: Based on the spatiotemporally aligned multimodal brain imaging data stream, an energy-optimal transmission model is constructed to constrain the rationality of dynamic brain network reconstruction after stroke. Specifically, the dynamic brain network reconstruction process is modeled as an energy-optimal transmission problem, the neurovascular coupling equation is introduced as a constraint, the reconstruction efficiency of key hub nodes and the rationality of resource allocation are calculated, and a compensation heat map is generated to reflect the resource allocation of the compensation pathway.

[0043] In step 5, pulse timing compression coding (STCE) technology is used to perform sparse processing on the pulse emission sequence of the SNN to generate a sparse event stream, and the temporal entropy change characteristics and compensation heat map are reconstructed in real time based on the sparse event stream on the edge embedded chip to achieve rapid and intelligent assessment of post-stroke cognitive impairment.

[0044] In a preferred embodiment, in step 2, the delay adaptive alignment algorithm further comprises: iteratively optimizing parameters of the neurovascular coupling model to minimize the residual error between the BOLD signal data and the white matter fiber tracking data on the time axis.

[0045] In a preferred embodiment, in step 3, the temporal entropy change model quantifies the transient response of synaptic weight adjustment during the compensation process by analyzing the frequency changes and phase synchronization of membrane potential oscillations, and identifies specific compensatory oscillation patterns associated with cognitive function recovery.

[0046] In a preferred embodiment, in step 4, the energy optimal transfer model further introduces a metabolic cost function, which is constructed based on the energy consumption relationship between neuronal activity and hemodynamics, and is used to inhibit the generation of compensatory pathways that are inconsistent with physiological laws and have excessive energy consumption.

[0047] In a preferred embodiment, in step 5, the pulse timing compression coding technology adopts an adaptive sparsification strategy based on dynamic reconstruction of pulse density, specifically including: real-time monitoring of the pulse computing density of each brain area of the brain network, when the pulse density of the key compensatory brain area exceeds the preset threshold, triggering the local pulse cluster merging mechanism, integrating the high-frequency micropulses into low-frequency macropulses and retaining their oscillation phase information; implementing event-driven truncation coding on the pulse sequence of non-critical brain areas, filtering low information entropy pulses through synaptic weight thresholds, and retaining only key conduction events of cross-hemispheric interactions; deploying a pulse buffer pool dynamic scheduling algorithm at the edge end, and dynamically adjusting the parallel processing queue depth of the pulse cluster according to the real-time remaining computing resources to ensure the real-time performance of the evaluation.

[0048] In a preferred embodiment, the local pulse cluster merging mechanism embeds a timestamp vector carrying the original oscillation phase when generating macropulses to ensure the integrity of the oscillation synchronization analysis in subsequent SNN modeling.

[0049] In a preferred embodiment, in the event-driven truncation coding, the synaptic weight threshold is dynamically adjusted based on the brain network reconstruction data and cognitive function recovery of historical stroke patients.

[0050] In a preferred embodiment, the method further includes the steps of: using a machine learning model to predict the cognitive function recovery trajectory of post-stroke patients based on the temporal entropy change characteristics and the compensatory heat map, and generating personalized rehabilitation suggestions.

[0051] In a preferred embodiment, the machine learning model includes but is not limited to a support vector machine, a neural network or a regression model.

[0052] In a preferred embodiment, after step 5, the method further includes the step of visually displaying the temporal entropy change characteristics and the compensation heat map.

[0053] Example 1: In this example, It represents the time offset between the BOLD signal and the DTI data caused by hemodynamic lag and white matter conduction velocity differences. To ensure the time alignment of multimodal data in the context of dynamic neural activity, this offset needs to be modeled and calculated based on the neurovascular coupling characteristics of individual brain regions. The length of the white matter tract path within a specific brain region obtained by DTI white matter fiber tracking, in millimeters (mm), can be extracted from the three-dimensional structural reconstruction of diffusion tensor imaging after standard atlas registration. The mean conduction velocity of white matter fibers, expressed in millimeters per millisecond (mm / ms), is usually set at 0.2 to 0.6 mm / ms based on the range reported in the clinical literature. An initial estimate can be derived from the P300 latency in individual functional provocation tests. It is a coupling coefficient used to correct the structural differences between nerves and blood vessels that are not explicitly modeled in the model. It is usually set as a dimensionless empirical constant. In practical applications, it can be obtained through small sample iterative optimization. Its initial value can be set between 0.8 and 1.2, and the specific value is automatically updated according to the minimum time axis residual criterion.

[0054] During the processing, the BOLD signal and DTI data are mapped to a unified spatial coordinate system, and the time axis is non-uniformly sampled and reconstructed so that each data point carries its spatial position and estimated timestamp. For each BOLD sampling point, the corresponding spatial position and DTI data association path are calculated. , correcting its timestamp. The resulting spatiotemporal aligned data stream has the synchronous expression characteristics of cross-modal neural signals under the same reference time axis.

[0055] Regarding neural network modeling, to improve the interpretability of spiking neural networks (SNNs) in identifying poststroke compensatory pathways, the original spatiotemporally aligned signals were first subjected to brain region registration and network node mapping. Each node represented a specific functional cluster, and the connection weights between nodes were constructed based on the changes in functional connectivity strength and encoded as a discretized spike sequence. This sequence was input into a pretrained SNN model, and the changes in its membrane potential were mapped to the frequency component fluctuations in the time series, with a particular focus on increased synchronization in the gamma band (30-80Hz).

[0056] To avoid relying solely on high-frequency oscillations as an evaluation basis, the system further introduces membrane potential temporal entropy as a joint characteristic of compensatory stability and resource utilization efficiency. Entropy calculation is based on a multi-window sliding aggregation approach, sequentially counting changes in the distribution of firing time intervals at key nodes in the spike train to quantify its rhythmicity and regular fluctuations. Increased synchronization in the gamma band is manifested as a decrease in entropy and an increase in frequency concentration within the corresponding window. Combining metabolic cost constraints, a compensatory heat map is constructed to spatially assess whether the resource distribution of compensatory pathways aligns with the neural-vascular distribution logic, thus preventing pseudo-compensatory pathways with abnormally high energy consumption from being mistakenly identified as valid connections. During the STCE compression phase, the system counts changes in spike density per unit time at each neural node. If the spike density in a particular region exceeds the device's capacity within a short time window, a spike cluster merging mechanism is activated. This mechanism uses a phase-preserving method to combine similar spikes into a single macro-pulse, embedding a timestamp vector representing the original oscillation rhythm. This merging strategy ensures that the SNN can accurately interpret the original synchronization characteristics during subsequent processing.

[0057] For non-critical brain areas, when the pulse density is low and the information entropy is insufficient, the system implements event-driven truncation coding. Based on the synaptic weight distribution threshold in historical samples, it dynamically filters pulses with low information content and retains event streams related to cross-hemispheric interactions or critical path coupling, further reducing the computational burden of the embedded system. In this process, the synaptic weight threshold is obtained based on historical stroke patient rehabilitation trajectory data training and is dynamically adjusted based on the current prediction error of the model to avoid information loss or resource waste caused by fixed thresholds. At the edge execution layer, a pulse buffer pool dynamic scheduling module is implemented and deployed. Based on the real-time feedback of the remaining computing resources and processing delays of the chip, the parallel processing depth and data extraction rate of the pulse sequence are dynamically adjusted to ensure stable output of heat maps and entropy change indicators for clinical reference without affecting real-time performance. These are all extended implementation methods known to ordinary technicians in this field.

[0058] Example 2: This example combines Figures 1 to 4 , further illustrate the technical solution of the present invention. Figure 1This is a schematic diagram of the delay-adaptive alignment process for brain imaging data. It primarily describes how the input blood oxygenation-dependent BOLD signal data and white matter fiber tracking data are dynamically estimated and corrected using a neurovascular coupling model to generate a spatiotemporally aligned multimodal brain imaging data stream. The process first imports the blood oxygenation-dependent BOLD signal data and white matter fiber tracking data through the Input Data module. Next, the delay-adaptive alignment module constructs a neurovascular coupling model to establish the physical constraints between the two data types. Based on this model, hemodynamic delays and conduction velocity differences are further estimated based on the model to obtain preliminary parameter estimates. These parameters are then iteratively updated to improve the estimation accuracy through iterative optimization of the model parameters. The system then enters the dynamic estimation and correction step, aligning the data under a non-uniform time axis. The final result is output through the Generate Spatiotemporally Aligned Data Stream module, providing the spatiotemporally aligned multimodal brain imaging data stream as output data.

[0059] Figure 2 The main steps and their logical connections from raw brain image acquisition to neural network modeling are demonstrated. First, in a multimodal acquisition system, BOLD signals and white matter fiber tracking data are synchronously acquired to form the raw input for subsequent processing. After a spatiotemporal calibration step, the raw data enters the dynamic alignment engine module, where blood flow delay compensation and conduction velocity correction operations are performed sequentially to correct the time axis mismatch between the BOLD signal and DTI data caused by hemodynamic lag and white matter fiber conduction differences. These two processing measures jointly complete dynamic temporal alignment and ensure the synchronization of multimodal data. After alignment, functional connectivity changes are converted into pulse trains through a membrane potential encoding process and input into the spiking neural network module, which is used to extract the temporal oscillation characteristics and compensatory patterns in brain network remodeling after stroke.

[0060] Figure 3This is a flow chart for multimodal brain imaging data alignment based on dynamic compensation. It demonstrates how to implement the entire data calibration process through the dynamic compensation mechanism, based on the fusion of fMRI BOLD signals and DTI white matter fiber data. The system first receives fMRI BOLD signals and DTI white matter fiber data as input and introduces neurovascular coupling coefficients and conduction velocity correction parameters based on their respective characteristics. Furthermore, the spatiotemporal alignment engine unifies the two data types into a common timeframe and feeds them into the dynamic compensation module for deep correction. Within this module, hemodynamic delays and white matter conduction differences are independently identified and modeled. The former primarily reflects the lag in the BOLD signal response relative to neural activity, while the latter characterizes the heterogeneity of conduction velocity along the DTI pathway. Together, the two constitute a complete dynamic compensation pathway. After correction, the system outputs a calibrated data stream—a multimodal fused data stream that has been corrected for both hemodynamic and conduction velocity deviations. This data stream is then used for subsequent analysis steps such as neural network modeling or compensatory pattern recognition.

[0061] Figure 4 A flowchart for pulse density sparse compression coding and sparse event stream generation is presented, describing the density monitoring and coding compression process that the original pulse sequence undergoes before entering the edge computing terminal. First, the system receives the original pulse sequence input and dynamically monitors the pulse activity frequency through the pulse density monitoring module. If the monitoring result exceeds the threshold, a local pulse cluster merging operation is triggered, adjacent high-frequency pulses are aggregated, and then macro-pulse generation is performed to construct a macro-coding unit that retains the original oscillation phase information, and finally outputs it to the sparse event stream. If the set threshold is not exceeded, the event-driven truncation coding stage is entered, and a resource-saving compression strategy is implemented for low-density pulse sequences. Only events meaningful for cross-brain region interactions are selected through the key conduction event retention mechanism, and a sparse event stream is also formed. This process implements branched processing and compression optimization of the original pulse signal, ensuring that high-value sparse event data can still be output for subsequent time series modeling even with limited embedded platform resources.

[0062] Example 3: In this example, the details of the dynamic alignment of multimodal data and the pulse neural network modeling process are specifically described. First, in the dynamic alignment process, the hemodynamic delay term The calculation has been done by the formula Define it. It represents the white matter fiber path length obtained by diffusion tensor imaging (DTI), in millimeters (mm). In practice, the connection distance between corresponding brain regions can be extracted by standard brain atlas registration; The white matter fiber conduction velocity is expressed in millimeters per millisecond (mm / ms). The reference literature usually sets the range as 0.2–0.6 mm / ms, and the individual initial value can be estimated through experimental methods such as P300 evoked potential. It is a dimensionless coupling correction coefficient that mainly reflects the individual differences between neural activity and blood flow response. It is recommended that the initial value be in the range of 0.8–1.2 and be iteratively updated by minimizing the BOLD signal and DTI time series residual.

[0063] In actual implementation, for each BOLD data point, the time offset is estimated using the aforementioned formula based on the correlation between its spatial mapping and the DTI pathway. Its timestamp is then corrected through non-uniform timeline interpolation to ensure that the functional and structural signals are fused on the same reference timeline, thereby improving cross-modal data consistency and analytical accuracy. During the spiking neural network (SNN) construction phase, the aligned brain network node signals are mapped into a spike sequence, and the membrane potential trend is calculated using a discrete time window. This sequence is input into a pre-trained spiking neural network, which not only preserves the spatial connectivity structure but also encodes the high-frequency variation characteristics of functional activity along the timeline.

[0064] This embodiment further illustrates that the calculation of entropy values is based on a multi-window sliding aggregation strategy. Within each sliding window, the pulse interval sequences of key nodes are statistically analyzed, their temporal distribution density is estimated using a probability distribution function, and their uncertainty index is calculated using the Shannon entropy form to assess the rhythmic stability of neural activity. This method not only provides a quantitative basis for determining synchronization enhancement but also provides temporal characteristics for subsequent physiological plausibility analysis of compensatory pathways. A decrease in entropy values typically indicates frequency concentration and enhanced neural network rhythmicity, a phenomenon particularly pronounced in the gamma frequency band. This can be used to identify effective compensatory pathways between brain regions. This embodiment further limits the sliding window width to between 100 and 200 ms, dynamically adjusting it based on the sampling frequency and neural discharge period to ensure statistical stability without sacrificing temporal resolution. When the pulse density exceeds the set threshold, the system will start the macro-pulse integration mechanism and generate a macro timestamp vector based on the principles of time continuity and phase retention to ensure that the analysis of frequency synchronization in subsequent analysis is not interfered by sparsification. In addition, the generation of sparse event streams in the compression process does not simply compress the data scale, but performs weight screening based on the structural characteristics of the neural network and the information entropy weight of the physiological signal. For example, in the deployment of the edge computing platform, if non-critical brain areas meet the conditions of low pulse density and low synaptic weight at the same time, their event streams will be actively truncated by the system to save computing resources; the event streams of critical brain areas enter the real-time processing channel through the priority queue mechanism to ensure the integrity of the decision-making characteristics.

[0065] Furthermore, the synaptic weight threshold is not fixed but dynamically adjusted based on the median of a weight distribution curve constructed from historical stroke patient rehabilitation data. This is combined with a reverse update of the current network prediction error to form a neural pathway screening strategy with adaptive learning capabilities. This helps avoid signal clipping that misses important compensatory features, while balancing computational efficiency with the credibility of physiological mechanisms. Regarding system deployment, the entire modeling and optimization process can be completed based on mainstream open-source platforms such as PyTorch (for SNN construction) and TensorFlow Lite for edge deployment. Data acquisition is recommended using a standardized fMRI acquisition protocol with a TR of 2000ms and DTI imaging (30 vectors, b=1000 s / mm²) to ensure consistency and reproducibility of model inputs. These are all extended implementations known to those skilled in the art.

[0066] Example 4: In this example, in the process of achieving spatiotemporal alignment, multimodal brain imaging data such as fMRI BOLD signals and DTI white matter fiber tracking data are first compensated using a neurovascular coupling model. Specifically, the hemodynamic delay is first calculated: and the conduction velocity difference of white matter fibers. In the formula, The calculation is as follows: ,in, represents the white matter fiber path length (unit: mm), represents the conduction velocity of white matter fibers (unit: mm / ms). is the neurovascular coupling coefficient, dimensionless, and is mainly used to compensate for neurovascular differences that are not explicitly modeled in the model. The initial value of can be set to 0.8 to 1.2, and dynamically optimized by minimizing the time axis residual of the BOLD signal and DTI data. In this process, the formula is used to correct the time offset of each BOLD signal sampling point, and the timestamp of each BOLD signal is corrected by associating it with the corresponding white matter fiber path in the DTI data, thereby ensuring the consistency of functional and structural data on the time axis and generating a spatiotemporally aligned data stream, as well as the formula, : Hemodynamic delay, used to compensate for the time offset caused by the difference in blood flow delay and white matter fiber conduction velocity; : white matter fiber path length, extracted from DTI data, in millimeters; : The conduction velocity of white matter fibers, usually based on the range set in the literature (0.2–0.6 mm / ms) or obtained through P300 evoked potential experiments; : Neurovascular coupling coefficient, used to correct the relationship between blood flow and neural activity, with a value range of 0.8 to 1.2.

[0067] The goal of spiking neural networks (SNNs) is to extract compensatory oscillation signatures from aligned multimodal data streams. In this step, nodes in the brain network are encoded through changes in functional connectivity strength, which are converted into spike trains. The membrane potential changes associated with spike trains reflect the temporal characteristics of neural activity. In particular, we focus on the phenomenon of increased synchronization in the gamma band (30-80 Hz), which can characterize the stability and resource efficiency of compensatory pathways. By analyzing the membrane potential oscillation characteristics of these spike trains, we can extract a temporal entropy change metric that represents the compensatory characteristics.

[0068] Based on the spatiotemporally aligned data streams, an energy-optimal transmission model is further constructed to constrain the rationality of brain network reconstruction. This model treats the dynamic reconstruction process of brain networks as an energy-optimal transmission problem and introduces the neurovascular coupling equation as a constraint. This model can effectively suppress the generation of spurious functional connections and ensure that the compensatory pathway conforms to the laws of neurovascular coupling. On this basis, data sparsification is performed using pulse temporal compression encoding (STCE) technology. Specifically, an adaptive sparsification strategy based on dynamic reconstruction pulse density is adopted. When the pulse density in certain key areas of the brain network exceeds a set threshold, a local pulse cluster merging mechanism is triggered. Through this mechanism, high-frequency micropulses are integrated into low-frequency macropulses while preserving their oscillation phase information. This operation not only effectively reduces the computational burden but also maintains the temporal synchronization of the compensatory pathway. For non-critical brain areas, if the pulse density is low and the information entropy is small, resource saving processing is performed through event-driven truncation coding, and only event streams with high information entropy and cross-brain area interactions are retained. This method can effectively reduce the consumption of computing resources while ensuring that information on key compensatory pathways is not lost. The temporal entropy change characteristics and compensation heat maps generated through the above steps will be output as real-time computing data streams and processed by edge embedded chips. The chip adjusts the parallel processing queue depth of the pulse cluster according to the real-time computing resources to ensure that the system can still stably output clinical-level evaluation results in a low computing power environment. These are all extended implementation methods that are known to ordinary technicians in this field.

[0069] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0070] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not limiting. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. An intelligent assessment method for post-stroke cognitive impairment based on dynamic reconstruction of brain networks, characterized in that: The following steps are involved: Step 1: Acquire multimodal brain imaging data of a post-stroke patient, wherein the multimodal brain imaging data at least includes blood oxygen level-dependent BOLD signal data of functional magnetic resonance imaging (fMRI) and white matter fiber tracking data of diffusion tensor imaging (DTI); Step 2: For multimodal brain imaging data, a delay adaptive alignment algorithm is used to dynamically match BOLD signal data and white matter fiber tracking data on a non-uniform time axis. Specifically, the algorithm includes: constructing a neurovascular coupling model, dynamically estimating and correcting the hemodynamic delay based on the neurovascular coupling model. and the cross-modal time axis misalignment caused by differences in white matter fiber conduction velocity, generating a spatiotemporally aligned multimodal brain imaging data stream; Step 3: Based on the spatiotemporally aligned multimodal brain imaging data stream, a spiking neural network (SNN) dynamic entropy change model is constructed to encode the temporal oscillation characteristics of post-stroke brain network reconstruction. Specifically, the model encodes the changes in the dynamic functional connectivity strength between brain network nodes into SNN spike sequences. By analyzing the membrane potential oscillation characteristics of the spike sequences, the temporal entropy change characteristics that characterize the stability of the compensatory pathway and the efficiency of resource consumption are extracted. Step 4: Based on the spatiotemporally aligned multimodal brain imaging data stream, an energy-optimal transmission model is constructed to constrain the rationality of dynamic brain network reconstruction after stroke. Specifically, the dynamic brain network reconstruction process is modeled as an energy-optimal transmission problem, the neurovascular coupling equation is introduced as a constraint, the reconstruction efficiency of key hub nodes and the rationality of resource allocation are calculated, and a compensation heat map is generated to reflect the resource allocation of the compensation pathway. In step 5, pulse timing compression coding (STCE) technology is used to perform sparse processing on the pulse emission sequence of the SNN to generate a sparse event stream, and the temporal entropy change characteristics and compensation heat map are reconstructed in real time based on the sparse event stream on the edge embedded chip to achieve rapid and intelligent assessment of post-stroke cognitive impairment.

2. The intelligent assessment method for post-stroke cognitive impairment based on dynamic brain network reconstruction according to claim 1, characterized in that: In step 2, the delay adaptive alignment algorithm further includes: iteratively optimizing the parameters of the neurovascular coupling model to minimize the residual error between the BOLD signal data and the white matter fiber tracking data on the time axis, and, Estimated based on the following formula: ， in, is the neurovascular coupling coefficient, is the length of white matter fibers, is the white matter fiber conduction velocity.

3. The intelligent assessment method for post-stroke cognitive impairment based on dynamic brain network reconstruction according to claim 1, characterized in that: In step 4, the energy optimal transfer model further introduces a metabolic cost function, which is constructed based on the energy consumption relationship between neuronal activity and hemodynamics, and in step 3, the temporal entropy change characteristics include a synchronous enhancement pattern of the Gamma band 30-80Hz.

4. The intelligent assessment method for post-stroke cognitive impairment based on dynamic brain network reconstruction according to claim 1, characterized in that: In step 5, the pulse timing compression coding technology adopts an adaptive sparsification strategy based on dynamic reconstruction of pulse density, specifically including: real-time monitoring of the pulse computing density of each brain area in the brain network, when the pulse density of the key compensatory brain area exceeds the preset threshold, triggering the local pulse cluster merging mechanism, integrating the high-frequency micropulses into low-frequency macropulses and retaining their oscillation phase information; implementing event-driven truncation coding on the pulse sequences of non-critical brain areas, filtering low information entropy pulses through synaptic weight thresholds, and retaining only key conduction events of cross-hemispheric interactions; deploying a pulse buffer pool dynamic scheduling algorithm at the edge to dynamically adjust the parallel processing queue depth of the pulse cluster according to the real-time remaining computing resources.

5. The intelligent assessment method for post-stroke cognitive impairment based on dynamic brain network reconstruction according to claim 4, characterized in that: The local pulse cluster merging mechanism embeds a timestamp vector carrying the original oscillation phase when generating macropulses.

6. The intelligent assessment method for post-stroke cognitive impairment based on dynamic brain network reconstruction according to claim 5, characterized in that: In event-driven truncation coding, the synaptic weight threshold is dynamically adjusted based on the brain network reconstruction data and cognitive function recovery of historical stroke patients.

7. The intelligent assessment method for post-stroke cognitive impairment based on dynamic brain network reconstruction according to any one of claims 1 to 6, characterized in that: The method further includes the steps of: using a machine learning model to predict the cognitive function recovery trajectory of post-stroke patients based on the temporal entropy change characteristics and the compensatory heat map, and generating personalized rehabilitation suggestions.

8. The intelligent assessment method for post-stroke cognitive impairment based on dynamic brain network reconstruction according to claim 7, characterized in that: Machine learning models include support vector machines, neural networks, or regression models.

9. The intelligent assessment method for post-stroke cognitive impairment based on dynamic brain network reconstruction according to claim 1, characterized in that: After step 5, the method further includes the step of visually displaying the temporal entropy change characteristics and the compensation heat map.

Images