Lumbar endoscopic postoperative chronic neuropathic pain early warning deep learning large model system

Abstract

The invention relates to the technical field of intelligent medical treatment, and discloses a lumbar endoscopic postoperative chronic neuropathic pain early warning deep learning large model system. According to the system, a postoperative multi-modal data initial pool of a patient is constructed, and original data streams are collected; performing space-time alignment and missing value interpolation on the original data stream by using the dynamic feature extraction network to form a standardized time sequence feature matrix; inputting the feature matrix into a hierarchical attention feature fusion network, calculating interaction weights of different modal features on a plurality of time scales, and generating enhanced feature representation fused with multi-scale context information; deducing an evolution path of a neural function state under different physiological parameter combinations through a risk trajectory simulation engine based on enhanced feature representation, and constructing an individualized risk evolution trajectory map; and according to the map, adopting a self-adaptive threshold decision model to identify an abnormal fluctuation interval deviating from a normal recovery mode, and outputting a chronic neuropathic pain risk early warning signal.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent medical treatment, in particular to a deep learning large model system for early warning of chronic neuropathic pain after lumbar endoscopy. BACKGROUND

[0002] The current early warning of chronic neuropathic pain after lumbar surgery mainly relies on the subjective experience of doctors or the threshold judgment of single clinical indicators. The existing technology lacks the ability to integrate postoperative multi-modal data, and the clinical information, surgical records and monitoring data are stored in scattered storage, lacking a unified spatio-temporal alignment mechanism. The feature extraction method is simple, and the dynamic correlation between multi-modal data has not been deeply mined. The risk prediction model is static, and it fails to simulate the dynamic evolution process of neural function under the interaction of individual physiological parameters. The early warning threshold setting is fixed, and it cannot adapt to the differences in recovery patterns of different patients. The existing method needs to solve the key technical problems of multi-source data fusion, dynamic feature extraction, individualized risk deduction and adaptive decision-making.

[0003] The traditional postoperative pain early warning system has obvious deficiencies in precision and forward-looking. The spatio-temporal alignment accuracy of multi-source data is low, and the key time sequence features are lost. The missing value imputation method is rough, which affects the data integrity. The feature fusion strategy is linear, which fails to capture the interaction effects of multiple time scales. The risk deduction model is idealized, and the complex coupling relationship between physiological parameters is not considered. The trajectory atlas construction relies on historical statistical laws, and the individualized adaptability is poor. The threshold decision mechanism is rigid, and the false alarm and missed alarm risks are high. The system lacks continuous optimization capability and cannot improve the early warning performance with case accumulation. SUMMARY

[0004] The purpose of the present application is to provide a deep learning large model system for early warning of chronic neuropathic pain after lumbar endoscopy to solve the problems raised in the background art.

[0005] To achieve the above-mentioned purpose, the present application provides a deep learning large model system for early warning of chronic neuropathic pain after lumbar endoscopy, which comprises: An initial pool of postoperative multi-modal data of a patient is constructed, and original data streams from a clinical information system, a surgical record system and a wearable device monitoring system are collected; Based on the initial pool of postoperative multi-modal data of the patient, a dynamic feature extraction network is used to perform spatio-temporal alignment and missing value imputation on the original data streams, forming a standardized time sequence feature matrix; The standardized time sequence feature matrix is input into a hierarchical attention feature fusion network to calculate the interaction weights of different modal features at multiple time scales, and an enhanced feature representation integrating multi-scale context information is generated; Based on the enhanced feature representation, the evolution path of the neural function state under different combinations of physiological parameters is deduced by a risk trajectory simulation engine, and an individualized risk evolution trajectory atlas is constructed. According to the individualized risk evolution trajectory atlas, an adaptive threshold decision model is used to identify abnormal fluctuation intervals deviating from the normal recovery mode, and a chronic neuropathic pain risk early warning signal is output.

[0006] Preferably, the specific steps of constructing the initial pool of postoperative multi-modal data of the patient include: Extracting patient demographic data, surgical operation details, perioperative medication records, and laboratory test results from a clinical information system; Synchronous operation recording system acquires the type of operating instruments, the duration of operation area, and the visibility score of neural structure in key frames of surgical video; Connect wearable device monitoring system to continuously collect postoperative heart rate variability, body movement frequency, posture maintenance duration, and other physiological behavior data; Assign a uniform timestamp to the above three types of data sources and establish a patient identifier association index to form the initial pool of postoperative multi-modal data of the patient.

[0007] Preferably, the specific steps of forming the standardized time sequence feature matrix include: Using a sliding window mechanism to divide the original data stream in the initial pool of postoperative multi-modal data of the patient into time segments; Using bidirectional long short-term memory network to learn forward and backward temporal dependencies within each time window to fill in missing data segments caused by device offline; Modeling the spatial correlation between different monitoring indicators through graph convolution network to unify the dimensions and normalize the values of heterogeneous data; Arrange the processed feature vectors in time sequence into a three-dimensional tensor structure to generate the standardized time sequence feature matrix.

[0008] Preferably, the specific steps of generating the enhanced feature representation fused with multi-scale context information include: Apply time attention mechanism on the standardized time sequence feature matrix to calculate feature importance scores at three time scales: hourly, daily, and weekly; Learn the nonlinear mapping relationship between physiological signals and behavior data through cross-modal interaction gate unit to dynamically adjust the contribution weights of different modalities; Use deep residual network to fuse multi-scale features, preserving the details of original features while enhancing long-range dependency modeling; Output the enhanced feature representation containing spatio-temporal correlation information.

[0009] Preferably, the specific steps of constructing the individualized risk evolution trajectory atlas include: Input the enhanced feature representation into the conditional generative adversarial network to simulate the change curve of pain sensitivity under different medication regimens; constructing a nerve conduction dynamics model based on a physical information neural network to predict the influence degree of the diffusion of inflammatory factors in the surgical area on the function of the nerve root; simulate the dynamic game process of the immune system and the nerve repair system through multi-agent reinforcement learning to generate a set of possible recovery paths; integrate all simulation results to form a three-dimensional risk evolution surface with a probability weight, i.e., the individualized risk evolution trajectory atlas.

[0010] Preferably, the specific steps of outputting the chronic neuropathic pain risk early warning signal include: defining a benchmark recovery corridor based on clinical guidelines on the individualized risk evolution trajectory atlas; using a variational autoencoder to detect the multidimensional distance between real-time monitoring data and the benchmark recovery corridor and calculate an abnormal deviation index; triggering a hierarchical early warning mechanism when the abnormal deviation index continuously exceeds the dynamic threshold value set by the adaptive threshold decision model; generating a structured early warning report containing the risk level, possible causes, and intervention suggestions, i.e., the chronic neuropathic pain risk early warning signal.

[0011] Preferably, the specific steps of using a dynamic feature extraction network to perform spatio-temporal alignment and missing value interpolation on the original data stream include: constructing a spatio-temporal alignment module for identifying timestamp offsets in multi-source data streams and synchronizing heterogeneous data to a unified time axis through interpolation algorithms; deploying a missing value detection unit for scanning discontinuities in the data stream and generating interpolation values based on time series prediction models; applying a graph attention network to model the spatial dependency between different monitoring indicators and adjusting the feature representation to reflect the collaborative changes between physiological parameters; using a multi-task learning framework to jointly optimize the spatio-temporal alignment and missing value interpolation process to minimize the weighted sum of reconstruction error and prediction bias.

[0012] Preferably, the specific steps of inputting the standardized time series feature matrix into a hierarchical attention feature fusion network and calculating the interaction weights of different modal features at multiple time scales include: using a local attention mechanism at the hour scale to capture short-term feature fluctuation patterns and calculate the local correlation scores of modal features at each time point; applying a recurrent attention unit at the daily scale to aggregate feature change trends within 24 hours and generate a daily feature importance distribution map; introducing a global attention pooling operation at the weekly scale to compress the time dimension and preserve long-term dependencies; The attention weight of three time scales is dynamically integrated through a gating fusion mechanism to output a modal feature interaction weight matrix.

[0013] Preferably, the specific steps of deducing the evolution path of the neurological function state under different combinations of physiological parameters by the risk trajectory simulation engine include: A neurophysiological parameter space mapping model is constructed to project multi-dimensional physiological parameters to a low-dimensional manifold to capture key feature combinations; A neural differential equation is used to simulate the dynamic change process of nerve conduction velocity with the concentration of inflammatory factors; A stochastic process is introduced to model the influence of external interference factors on the neurological function state, generating a set of probabilistic evolution paths; The transition probability of the neurological function state under different parameter combinations is calculated by the path integral method to form an evolution path atlas.

[0014] Preferably, the specific steps of simulating the dynamic game process of the immune system and the nerve repair system by multi-agent reinforcement learning include: An immune system agent proxy is designed, whose decision strategy is based on the concentration of inflammatory factors and the change of immune cell activity; A nerve repair system agent proxy is constructed, whose action space includes axon regeneration rate and myelin repair intensity; A game reward function is established to quantify the clinical benefit indicators of the balance state of immune response and nerve repair; A multi-agent policy gradient algorithm is used to train the agent proxy to simulate the equilibrium solution space in the long-term dynamic game.

[0015] Compared with the prior art, the present application has the following advantages: The dynamic feature extraction network is used to perform spatio-temporal alignment and missing value interpolation on the multi-modal raw data stream to form a standardized time series feature matrix. The network uses a graph neural network structure, where nodes represent data sources and edges represent spatio-temporal association relationships. The spatio-temporal alignment algorithm effectively compensates for the time stamp deviation between different sampling frequencies and devices, ensuring data synchronization. The missing value interpolation is realized through a generative adversarial network, which simulates the real data distribution characteristics and improves the interpolation accuracy. Standardization eliminates dimensional differences, unifies feature scales, and preserves data dynamic change rules.

[0016] The standardized time series feature matrix is input into the hierarchical attention feature fusion network to generate enhanced feature representation. The network includes two layers of intra-modal attention and cross-modal attention structures, which respectively learn the importance weights of features at different time points in a single modality and the interaction intensity of different modal features at the same time point. The time scale is divided into short-term, medium-term and long-term, corresponding to the feature changes at different stages after surgery. The interaction weights are normalized to highlight key feature contributions and effectively fuse multi-scale context information.

[0017] Based on the enhanced feature representation, the evolution path of the neurological function state is deduced by using a risk trajectory simulation engine to construct an individualized risk evolution trajectory atlas. The engine adopts a structure combining a recurrent neural network and a physical guidance mechanism to simulate the dynamic changes of neurological function under different physiological parameter combinations. Possible state change trajectories are generated through numerical integration and probability sampling, and the evolution path is stored in a graph structure, with nodes representing neurological function states and edges representing state transition probabilities.

[0018] According to the individualized risk evolution trajectory atlas, an adaptive threshold decision model is used to identify abnormal fluctuation intervals and output risk warning signals. The decision model is based on a reinforcement learning framework and dynamically adjusts the decision boundary according to the warning effect. Abnormal detection combines statistical process control and wavelet transform algorithms to accurately identify fluctuation intervals that deviate from normal recovery patterns. The warning signal contains multi-dimensional information such as risk level and occurrence probability, achieving precise and personalized risk warning.

[0019] Through the synergistic effect of the above technical solutions, the whole process from data preprocessing to risk warning is optimized. Dynamic feature extraction ensures data quality, hierarchical attention fusion improves feature representation ability, risk trajectory simulation realizes accurate prediction, and adaptive decision ensures warning accuracy. The system significantly improves the reliability and practicality of postoperative pain warning. BRIEF DESCRIPTION OF DRAWINGS

[0020] Figure 1 The working principle diagram of the lumbar endoscopic postoperative chronic neuropathic pain early warning deep learning large model system described in the present application; Figure 2 The flowchart for constructing the initial pool of postoperative multi-modal data for the patient; Figure 3 The flowchart for forming the standardized time series feature matrix; Figure 4 The performance index comparison chart of different feature fusion methods; Figure 5 The individualized chronic neuropathic pain risk evolution surface chart. DETAILED DESCRIPTION

[0021] The technical solutions in the embodiments of the present application will be described clearly and completely below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the scope of protection of the present application.

[0022] Please refer to Figure 1The application provides a lumbar endoscopy postoperative chronic neuropathic pain early warning deep learning large model system, which comprises: precise postoperative risk early warning is realized by integrating multi-modal data and analysis technology. A patient postoperative multi-modal data initial pool is constructed, and original data streams from a clinical information system, a surgical record system and a wearable device monitoring system are collected; based on the patient postoperative multi-modal data initial pool, a dynamic feature extraction network is used to perform time-space alignment and missing value interpolation on the original data streams, so that a standardized time sequence feature matrix is formed; the standardized time sequence feature matrix is input into a hierarchical attention feature fusion network, the interaction weight of different modal features on multiple time scales is calculated, and an enhanced feature representation of fused multi-scale context information is generated; based on the enhanced feature representation, a risk trajectory simulation engine is used to deduce the evolution path of the nerve function state under different physiological parameter combinations, and an individualized risk evolution trajectory atlas is constructed; according to the individualized risk evolution trajectory atlas, an adaptive threshold decision model is used to identify the abnormal fluctuation interval deviating from the normal recovery mode, and a chronic neuropathic pain risk early warning signal is output.

[0023] Embodiment 1: refer to Figure 2 In specific implementation, patient demographic data, surgical operation details, perioperative medication records and laboratory test results are extracted from the clinical information system, the clinical information system is usually deployed on a hospital server cluster, the structured table in the relational database is accessed through a secure data transmission protocol, the extraction process is performed by a structured query language script at a fixed time, and after filtering invalid or redundant items, the data is stored in an intermediate database. In some embodiments, the surgical operation details include fields such as surgical approach, foraminoplasty range and nerve decompression degree, the perioperative medication records cover the dosage and use time axis of non-steroidal anti-inflammatory drugs, opioid drugs and neurotrophic agents, and the laboratory test results mainly extract inflammation indicators such as C-reactive protein and white blood cell count values. Optionally, the data extraction module integrates a data quality verification submodule, detects abnormal values or logical contradictions through a rule engine, and ensures data consistency.

[0024] The synchronous surgical record system obtains the type of operating instrument, the duration of operating area and the visibility score of nerve structure of the key frame in the surgical video, the surgical record system is connected with a high-definition endoscope camera device, a video stream is extracted by a frame sampling algorithm to obtain a key frame sequence, and the key frame corresponds to a key step of surgery such as nerve root exposure or nucleus pulposus removal. In specific implementation, the type of operating instrument is classified by an image recognition convolutional neural network, the type of instrument such as a radio knife, a nucleus pulposus clamp or a working sleeve is identified, the duration of the operating area is calculated by time stamp difference, and the visibility score of the nerve structure is output by a pre-trained deep learning model, the score is quantified based on the nerve root contour sharpness and the tissue obstruction degree. In some embodiments, the visibility score of the nerve structure is expressed as: wherein: represents the visibility score, represents the number of features, represents the weight coefficient of the i-th feature, represents the i-th image feature value such as edge sharpness or contrast, and the weight coefficient is obtained by training a data set marked by a surgical expert.

[0025] The wearable device monitoring system continuously collects postoperative heart rate variability, body movement frequency, and posture maintenance duration. The wearable device, such as a smart waistband or patch sensor, communicates with the gateway through the Bluetooth Low Energy protocol, and the data is transmitted to the cloud platform buffer queue in a stream form. In specific implementation, the heart rate variability is calculated from the electrocardiogram signal to obtain frequency domain indicators such as low frequency to high frequency power ratio, the body movement frequency is monitored by a three-axis accelerometer to count the number of body movements per unit time, and the posture maintenance duration is calculated based on the posture estimation algorithm to count the continuous time interval of the patient maintaining a sitting or lying posture. Optionally, the data acquisition module realizes real-time data compression and encryption, reduces the transmission bandwidth occupation, and safeguards the privacy and security of the patient.

[0026] The above three types of data sources are assigned a unified timestamp and a patient identifier association index is established to form an initial pool of postoperative multi-modal data of the patient. The timestamp synchronization uses the Network Time Protocol to calibrate all device clocks, and the data association is based on a hash index table constructed based on the patient's unique hospital number. In specific implementation, the data pool is stored in a distributed columnar database, partitioned by time to optimize query efficiency, and the patient identifier association index supports fast association query and retrospective analysis of multi-modal data.

[0027] Embodiment 2: refer to Figure 3 The specific implementation of forming a standardized time series feature matrix involves time series processing and feature standardization of the initial pool of postoperative multi-modal data of the patient. In specific implementation, a sliding window mechanism is used to divide the original data stream in the initial pool of postoperative multi-modal data of the patient into time segments. The sliding window has a fixed step size and window size, for example, a one-hour window and a ten-minute step size sliding window. The continuous data stream is cut into sequence segments with overlapping regions for model processing. In some embodiments, the window size and step size are dynamically adjusted according to the data sampling frequency and clinical monitoring requirements. High-frequency physiological signals use smaller windows to capture instantaneous fluctuations, and low-frequency behavior data use larger windows to extract trend features.

[0028] Bidirectional long short-term memory network is used to learn forward and backward time dependencies in each time window to complete the data missing section caused by device offline. The bidirectional long short-term memory network includes two long short-term memory network layers in forward and backward directions, respectively scans from the beginning to the end of the sequence and from the end to the beginning of the sequence, and captures complete context information. In specific implementation, the data missing section is marked with a specific mask value before input, the hidden state of the bidirectional long short-term memory network is calculated based on the past and future data points, and an interpolation value is generated to replace the mask position. The calculation of the interpolation value fuses the past and future data patterns in the time window. Optionally, the training target of the bidirectional long short-term memory network includes minimizing the reconstruction error on the known data points to ensure the reliability of the interpolation value.

[0029] The spatial correlation between different monitoring indicators is modeled by a graph convolution network, and the heterogeneous data is dimensionally unified and numerically normalized. The graph structure takes the monitoring indicators as nodes, the node features are the time series data in the time window, and the edge weights are defined based on physiological knowledge or data correlation, reflecting the functional connection between the indicators. In specific implementation, the graph convolution operation aggregates the feature information of the neighbor nodes of each node, updates the node representation, so that the heterogeneous data from the clinical information system, the operation record system and the wearable device monitoring system are associated in the feature space. The numerical normalization adopts the min-max scaling method to linearly transform the values of each indicator to the interval [0, 1], eliminating the dimension effect. The output of the graph convolution network layer is a standardized feature vector enhanced by spatial correlation.

[0030] The processed feature vectors are arranged in time sequence as a three-dimensional tensor structure to generate a standardized time series feature matrix. The three dimensions of the three-dimensional tensor correspond to the number of samples, the time step and the feature dimension respectively. In specific implementation, each sample is the data in a time window, the time step corresponds to the number of time points in the window, and the feature dimension is the length of the unified feature vector after the graph convolution network processing. A space-time alignment module is constructed to identify the timestamp offset in the multi-source data stream, and synchronize the heterogeneous data to a unified time axis through an interpolation algorithm. The timestamp offset may be caused by the slight difference between the internal clocks of different acquisition devices. In specific implementation, the space-time alignment module first detects the timestamp offset of other data streams based on the time axis of the data stream with the highest sampling rate, and then resamples the data with lower sampling rate using a cubic spline interpolation algorithm to align it with the reference time axis.

[0031] A missing value detection unit is deployed to scan the discontinuities in the data stream and generate filling values based on a time series prediction model. The missing value detection unit identifies data discontinuities by comparing the interval of consecutive timestamps with the expected sampling interval. In specific implementation, the time series prediction model adopts an autoregressive integrated moving average model, and its formula is represented as: wherein: denotes the value of the time point t, is a constant term, is the order of the autoregressive term, is the i-th autoregressive coefficient, is the order of the moving average term, is the i-th moving average coefficient, is the error term of the time point t. The model is trained with valid data before and after the discontinuity and predicts the value at the missing position. A graph attention network is applied to model the spatial dependency between different monitoring indicators, adjusting the feature representation to reflect the collaborative changes between physiological parameters. The graph attention network calculates the attention weight for each edge in the graph, representing the strength of the influence between nodes. In specific implementation, when updating the node features, the contributions of neighboring nodes are weighted and summed according to the attention weights, thereby dynamically adjusting the feature representation and more accurately capturing physiological relationships such as heart rate variability and body motion frequency.

[0032] A multi-task learning framework is adopted to jointly optimize the spatio-temporal alignment and missing value imputation process, minimizing the weighted sum of reconstruction error and prediction bias. The multi-task learning framework shares the feature extraction layer and sets up a spatio-temporal alignment task branch and a missing value imputation task branch respectively. In specific implementation, the reconstruction error calculates the difference between the aligned data and the true benchmark data, and the prediction bias evaluates the gap between the imputed value and the actual recorded value. The total loss function is the weighted sum of the two, and the model parameters are optimized through the gradient descent algorithm. The specific implementation of the gradient descent algorithm is applied to the multi-task learning framework to jointly optimize the model parameters of the spatio-temporal alignment and missing value imputation process. The algorithm is calculated based on the total loss function, which is the weighted sum of the reconstruction error and the prediction bias. The reconstruction error measures the difference between the aligned data and the true benchmark data, and the prediction bias evaluates the gap between the imputed value and the actual recorded value. The gradient descent algorithm calculates the gradient of the loss function with respect to the model parameters and updates the parameter values in the opposite direction of the gradient, gradually reducing the value of the total loss function, thereby achieving joint optimization of spatio-temporal alignment and missing value imputation. Optionally, the weighting weight is set according to the importance of the two tasks in the specific application scenario.

[0033] In some embodiments, the time attention mechanism is applied on the standardized time-series feature matrix to calculate the feature importance scores at three time scales: hourly, daily, and weekly. The time attention mechanism assigns an independent attention weight calculation module for each time scale. The module evaluates the importance of feature vectors at different time points to the current context representation through a query-key-value model. The query vector is usually generated from the current context, while the key and value vectors are derived from the feature representations at different time points. In some embodiments, the hourly scale focuses on short-term physiological fluctuation patterns, the daily scale analyzes circadian rhythm changes, and the weekly scale tracks long-term recovery trends. The attention weight vectors calculated at each scale are used to weight and sum the original features to obtain the scale-specific feature representation.

[0034] The non-linear mapping relationship between physiological signals and behavioral data is learned through a cross-modal interaction gating unit, which dynamically adjusts the contribution weights of different modalities. The cross-modal interaction gating unit takes feature vectors from different data sources as input, such as the heart rate variability features collected by wearable monitoring systems and the neurostructural visibility score features obtained from surgical record systems. In some embodiments, the core calculation of the gating unit involves transformation and filtering of the input features, and its output is used to modulate the flow of different modal features in the final fused representation. Optionally, the gating unit can be implemented using a fully connected layer combined with an activation function.

[0035] A deep residual network is used to fuse multi-scale features, preserving the details of the original features while enhancing the modeling of long-range dependencies. The deep residual network is composed of multiple residual blocks, each containing a fully connected layer, a normalization layer, and an activation function, and the input and output are added through a shortcut connection. In some embodiments, the feature representations from the hourly, daily, and weekly time scales are first concatenated into a high-dimensional feature vector, which is then used as input to the deep residual network. The deep layers of the network alleviate the gradient vanishing problem through residual connections, enabling the model to effectively learn complex mappings from shallow multi-scale features to deep fused representations. It can be understood that the structure of the deep residual network facilitates cross-layer information transmission.

[0036] The output includes an enhanced feature representation containing spatio-temporal correlation information. This enhanced feature representation is a compact vector representation that integrates short-term details, medium-term trends, and long-term patterns, and is calibrated through cross-modal information. In some embodiments, the specific steps of inputting the standardized time-series feature matrix into the hierarchical attention feature fusion network and calculating the interaction weights of different modal features at multiple time scales include using a local attention mechanism at the hourly scale to capture short-term feature fluctuation patterns. The calculation range of the local attention mechanism is limited within a sliding time window, and the local correlation scores between different modal features at each time point are calculated. In some embodiments, the local attention weights are calculated using the softmax function, which is represented by the formula: The output includes an enhanced feature representation containing spatio-temporal correlation information. This enhanced feature representation is a compact vector representation that integrates short-term details, medium-term trends, and long-term patterns, and is calibrated through cross-modal information. In some embodiments, the specific steps of inputting the standardized time-series feature matrix into the hierarchical attention feature fusion network and calculating the interaction weights of different modal features at multiple time scales include using a local attention mechanism at the hourly scale to capture short-term feature fluctuation patterns. The calculation range of the local attention mechanism is limited within a sliding time window, and the local correlation scores between different modal features at each time point are calculated. In some embodiments, the local attention weights are calculated using the softmax function, which is represented by the formula: wherein: represents the weight of the j-th feature when computing the attention of the i-th feature, is a scalar attention score computed by the similarity between the query vector and the key vector, represents the number of features. This weight is used to aggregate feature information within a local window.

[0037] The recurrent attention unit is applied on a daily scale to aggregate the feature variation trend within 24 hours, generating a daily feature importance profile. The recurrent attention unit is usually constructed based on a gated recurrent unit or a long short-term memory network structure, whose hidden state recurrently processes the 24-hour scale feature sequence and outputs an attention weight at each time step, which reflects the contribution degree of the feature of that hour to the overall feature representation of the "day". In specific implementation, the daily feature importance profile is a 24-dimensional vector, and the value of each element in the vector corresponds to the feature importance score of a specific hour in a day, and the high or low score reveals the representative strength of the physiological behavior pattern in the daily cycle. Optionally, the parameters of the recurrent attention unit are learned on the training data.

[0038] The global attention pooling operation is introduced on a weekly scale to compress the time dimension and preserve long-term dependencies. The global attention pooling processes a time series lasting for several weeks and computes a global attention score for each time point in the entire sequence. In specific implementation, the computation of the global attention score relies on a learnable context vector, which interacts with the feature representation of each time point, and the attention distribution is normalized by the softmax function. The final feature representation is the weighted sum of all time point features according to their global attention scores, and this operation compresses the variable-length time series into a fixed-length vector.

[0039] The attention weights of the three time scales are dynamically integrated through a gating fusion mechanism to output a modal feature interaction weight matrix. The gating fusion mechanism receives the attention weight matrices from the hour, day, and week scales and learns a final fusion weight. In specific implementation, the fusion weight is usually generated by a small neural network or a linear transformation layer, whose input is the summary statistics or context information of the three-scale attention weight matrices, and the output modal feature interaction weight matrix is a multi-dimensional tensor that explicitly quantifies the interaction strength between different modal features at different time scales.

[0040] Referring to Figure 4FIG. 4 is a grouped column chart focusing on the technical effect verification of real hierarchical attention feature fusion, horizontally comparing the performance of single modality, simple splicing, and hierarchical attention fusion in three core performance indicators of accuracy, recall rate, and precision. The figure intuitively reflects the advantages of the technology: through the combination of time attention mechanism, cross-modal gating unit, and deep residual network, the spatio-temporal correlation and multi-scale context information of multi-modal data are effectively captured, significantly improving the quality of feature representation, and providing a more reliable feature basis for subsequent risk trajectory simulation and individualized early warning.

[0041] In specific implementations, the enhanced feature representation is input into a conditional generative adversarial network to simulate the change curve of pain sensitivity under different medication regimens. The generator of the conditional generative adversarial network takes the enhanced feature representation and a specific medication regimen encoding as conditional input, and outputs a predicted sequence of pain sensitivity in the future period of time. The discriminator distinguishes between the predicted sequence and the real clinical observation sequence. In some embodiments, the medication regimen encoding is a multi-dimensional vector containing drug type, dose, and administration frequency. The simulation process observes the changes in the pain sensitivity curve output by the generator by systematically changing the medication regimen encoding, thereby evaluating the potential effects of different drug interventions.

[0042] A neural conduction dynamics model is constructed based on a physical information neural network to predict the degree of influence of the diffusion of inflammatory factors in the surgical area on the function of the nerve root. The physical information neural network embeds the biophysical laws of neural conduction in the form of partial differential equation constraints in the network training loss function. In specific implementations, the concentration field of inflammatory factors is initialized by related indicators in the enhanced feature representation. The physical information neural network solves the coupled reaction-diffusion equation to simulate the spatio-temporal diffusion process of inflammatory factors around the nerve root and calculate their inhibitory effect on the conduction velocity. The degree of influence on nerve function is finally quantified as the percentage of conduction velocity reduction. Optionally, the boundary conditions of the physical information neural network are set according to the specific anatomical image data of the patient.

[0043] The dynamic game process between the immune system and the neural repair system is simulated through multi-agent reinforcement learning to generate a set of possible recovery paths. The immune system agent and the neural repair system agent interact in a shared environment, and the environment state is defined by variables such as inflammation level and tissue damage degree. In specific implementation, an immune system agent is designed, whose decision-making strategy is based on the concentration of inflammatory factors and the changes in immune cell activity. The action space of the immune system agent includes the amount of pro-inflammatory cytokine release and the intensity of anti-inflammatory response, and its goal is to balance between clearing potential threats and avoiding excessive inflammation. A neural repair system agent is constructed, whose action space includes the rate of axon regeneration and the intensity of myelin repair. The neural repair system agent adjusts its repair behavior according to the levels of neurotrophic factors and injury signals. A game reward function is established to quantify the clinical benefit indicators of the balance state between immune response and neural repair, and the reward function is used to evaluate the overall state of the system: wherein: represents the reward value, represents the normalized value of chronic inflammation, represents the degree of continuous tissue damage, represents the neural function recovery score, is the weight coefficient of each indicator. Multi-agent policy gradient algorithm is used to train the agent, simulate the equilibrium solution space in long-term dynamic game, and estimate the policy parameter gradient and update it. Finally, the agent learns the strategy that maximizes the cumulative reward in the long-term interaction, thereby generating a series of recovery paths reflecting different internal physiological balance states.

[0044] All simulation results are integrated to form a three-dimensional risk evolution surface with probability weights, i.e., an individualized risk evolution trajectory map. The two horizontal axes of the three-dimensional space represent different combinations of key physiological parameters, and the vertical axis represents the estimated risk probability of developing chronic neuropathic pain. In specific implementation, the specific steps of the risk trajectory simulation engine to deduce the evolution path of the neural function state under different physiological parameter combinations include constructing a neurophysiological parameter space mapping model to project multi-dimensional physiological parameters into a low-dimensional manifold to capture key feature combinations. This model usually uses autoencoders or principal component analysis to extract the most discriminative low-dimensional features from enhanced feature representations, which form the basis of the parameter space for subsequent deduction. Referring to Table 1, a simplified parameter space mapping is shown.

[0045] Table 1: Neurophysiological Key Parameter Space Mapping The neural differential equation is used to simulate the dynamic change process of the nerve conduction velocity with the concentration of inflammatory factors. The neural differential equation expresses the change rate of the nerve conduction velocity as a function of the current conduction velocity, the concentration of inflammatory factors and other related states, thereby realizing the modeling of the continuous time dynamic system. In specific implementation, the equation form can be preset based on physiological knowledge, and the parameters are calibrated through patient data. By numerically solving the neural differential equation, the continuous evolution trajectory of the nerve conduction velocity under a specific change path of the concentration of inflammatory factors can be obtained. The influence of external interference factors on the nerve function state is modeled by introducing a random process to generate a set of probabilistic evolution paths. The external interference factors include accidental infection, abnormal physiological stress and other events that are not directly captured in regular monitoring. In some embodiments, a Poisson process is used to simulate the occurrence of interference events, and the event intensity is related to the patient's basic state. Each interference event will produce a random impact on the nerve function state. By Monte Carlo simulation, different interference event sequences are sampled multiple times to generate a set of evolution paths with probability distribution. The transition probability of the nerve function state under different parameter combinations is calculated by the path integral method to form an evolution path atlas. The path integral method performs weighted summation on all possible paths from the initial state to the target state, and the weight is determined by the likelihood function and the risk function of the path. It can be understood that the path integral provides a rigorous mathematical framework for quantifying the state transition probability under uncertainty. The final three-dimensional risk evolution surface is constructed by densely sampling the parameter space and calculating the risk probability value corresponding to each sampling point. Optionally, the risk probability value can be determined by the proportion of adverse outcomes of similar paths in the historical data or the predicted score converted by the sigmoid function.

[0046] Referring to Figure 5 The figure is a three-dimensional surface graph, which is the core visualization result of constructing the individualized risk evolution trajectory atlas: the two independent variables in the figure are the inflammatory load and the repair potential, and the dependent variable is the risk probability of chronic neuropathic pain. The color gradient corresponds to the high and low risk probability. This figure directly presents the risk evolution rule under different physiological parameter combinations: when the inflammatory load increases and the repair potential decreases, the risk probability significantly increases; otherwise, the risk is in a lower interval. This result is the product of the integration of multiple technical links. After simulating the influence of drug use by the conditional generative adversarial network, constructing the nerve conduction dynamics model by the physical information neural network, and simulating the immune-repair game by the multi-agent reinforcement learning, the risk evolution surface with probability weight is formed, which provides individualized risk trajectory benchmarks for the subsequent adaptive threshold decision model to identify abnormal recovery intervals.

[0047] Example 5: In specific implementation, a baseline recovery corridor based on clinical guidelines is defined on an individualized risk evolution trajectory map. The construction of the baseline recovery corridor is based on the normal postoperative recovery range determined by large-scale clinical follow-up studies. For example, the upper and lower bounds of the normal evolution of multiple dimensions such as pain scores, inflammatory markers, and functional activity are delineated into a safe channel area on the risk evolution trajectory map. In some embodiments, the baseline recovery corridor is not fixed but is individually adjusted according to the patient's preoperative baseline condition and surgical type, forming a dynamic reference boundary. It can be understood that the baseline recovery corridor provides a quantitative standard for assessing whether the actual recovery process deviates from expectations.

[0048] A variational autoencoder (VAE) is used to detect the multidimensional distance between real-time monitoring data and the baseline recovery corridor, and to calculate the anomaly deviation index. The VAE consists of an encoder and a decoder. The encoder maps the real-time input multidimensional physiological data to a latent variable space, and the decoder reconstructs the data from this space. In practice, the anomaly deviation index is calculated based on the statistical difference between the representation of real-time data in the latent space and the latent representation of normal data corresponding to the baseline recovery corridor. A commonly used metric is the Mahalanobis distance, expressed by the following formula: in: Representing data points The distance to Maharanobis This represents the mean vector of the underlying representation of the baseline recovery corridor normal data. This indicates the transpose operation. This represents the covariance matrix of the underlying representation of the baseline recovery corridor normal data. The distance represents the inverse of the matrix, taking into account the correlation between different dimensions of the data, and can more sensitively detect multivariate anomalies. Variational autoencoders, by adding a regularization term to their training objective, make the distribution of the latent space approximate a standard normal distribution, which helps improve the robustness of anomaly detection.

[0049] When the abnormal deviation index continuously exceeds the dynamic threshold set by the adaptive threshold decision model, a tiered early warning mechanism is triggered. The adaptive threshold decision model dynamically adjusts the threshold for triggering the early warning based on the patient's historical data, current overall physiological state, and environmental factors. In practice, the dynamic threshold is not a single fixed value but is generated by a function. The inputs to this function include recent sliding window statistics of the abnormal deviation index, circadian rhythm phase, and medication records. For example, the threshold may be appropriately relaxed during nighttime rest periods or after taking analgesics to reduce false alarms, while remaining sensitive during daytime activity periods. The tiered early warning mechanism typically sets multiple threshold levels, corresponding to different colored warning signals. The duration and magnitude of the continuous exceedance jointly determine the final warning level triggered.

[0050] A structured early warning report containing risk level, possible causes and intervention suggestions, i.e. a chronic neuropathic pain risk early warning signal, is generated, which is organized in a predefined JSON or XML format to ensure machine readability and easy integration into clinical information systems. In specific implementations, the risk level is directly determined by the grading result that triggers the early warning mechanism, the analysis of possible causes combines the highest abnormal deviation indices of several data dimensions, the recent simulated risk evolution path and medical rules in the knowledge base, for example, if the abnormal deviation is mainly driven by a significant decrease in heart rate variability and a decrease in body motion frequency, the possible cause is inferred as "autonomic dysfunction or pain exacerbation leading to limited activity". Intervention suggestions are then generated from the pre-stored clinical decision support knowledge base according to the risk level and inferred causes, which can include adjusting the analgesic regimen, suggesting specific examinations or strengthening rehabilitation training. Optionally, the structured early warning report can generate trend charts through a visualization component to assist clinicians in quickly understanding the risk situation. In some embodiments, the early warning signal is automatically pushed to the attending physician's mobile terminal or hospital monitoring platform after generation.

[0051] It should be noted that the terms such as first and second are used herein merely to distinguish one entity or action from another, without necessarily requiring or implying any such actual relationship or order between such entities or actions. Moreover, the terms "comprises", "comprising", or any other variations thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but can include other elements not expressly listed or inherent to such process, method, article, or apparatus.

[0052] Although embodiments of the present application have been shown and described, it is to be understood that various modifications, substitutions, alternatives, and variations can be made in the embodiments without departing from the spirit and scope of the present application as defined by the appended claims and their equivalents.

Claims

1. A deep learning-based large-scale model system for early warning of chronic neuropathic pain after lumbar endoscopic surgery, characterized in that, The system operation includes the following steps: An initial pool of postoperative multimodal patient data was constructed, aggregating raw data streams from clinical information systems, surgical record systems, and wearable device monitoring systems; Based on the initial pool of postoperative multimodal data of the patients, a dynamic feature extraction network is used to perform spatiotemporal alignment and missing value imputation on the original data stream to form a standardized temporal feature matrix. The standardized temporal feature matrix is ​​input into the hierarchical attention feature fusion network to calculate the interaction weights of different modal features at multiple time scales and generate an enhanced feature representation that integrates multi-scale contextual information. Based on the enhanced feature representation, the evolution path of neural functional state under different combinations of physiological parameters is deduced through the risk trajectory simulation engine, and an individualized risk evolution trajectory map is constructed. Based on the individualized risk evolution trajectory map, an adaptive threshold decision model is used to identify abnormal fluctuation ranges that deviate from the normal recovery pattern and output a chronic neuropathic pain risk warning signal.

2. The deep learning-based large-scale model system for early warning of chronic neuropathic pain after lumbar endoscopic surgery as described in claim 1, characterized in that, The specific steps for constructing the initial pool of postoperative multimodal patient data include: Patient demographic data, surgical procedure details, perioperative medication records, and laboratory test results were extracted from the clinical information system. The synchronous surgical recording system acquires the type of operating instrument, duration of the operating area, and visual score of neural structures in key frames of surgical videos. The system connects to wearable monitoring devices to continuously collect physiological data such as postoperative heart rate variability, body movement frequency, and posture maintenance duration. The three types of data sources mentioned above are assigned a unified timestamp and a patient identifier association index is established to form the initial pool of postoperative multimodal data of the patients.

3. The deep learning-based large-scale model system for early warning of chronic neuropathic pain after lumbar endoscopic surgery as described in claim 2, characterized in that, The specific steps for forming the standardized time-series feature matrix include: A sliding window mechanism was used to divide the raw data stream in the initial pool of the patient's postoperative multimodal data into time segments; Within each time window, a bidirectional long short-term memory network is used to learn forward and backward temporal dependencies and fill in data gaps caused by device offline. By modeling the spatial correlation between different monitoring indicators using graph convolutional networks, dimensional unification and numerical normalization are achieved for heterogeneous data. The processed feature vectors are arranged in chronological order into a three-dimensional tensor structure to generate the standardized temporal feature matrix.

4. The deep learning-based large-scale model system for early warning of chronic neuropathic pain after lumbar endoscopic surgery according to claim 3, characterized in that, The specific steps for generating the enhanced feature representation that incorporates multi-scale contextual information include: A time attention mechanism is applied to the standardized time-series feature matrix to calculate feature importance scores at three time scales: hourly, daily, and weekly. By learning the nonlinear mapping relationship between physiological signals and behavioral data through cross-modal interactive gating units, the contribution weights of different modalities can be dynamically adjusted. Deep residual networks are used to fuse multi-scale features, which enhances long-range dependency modeling while preserving the details of the original features. The output includes the enhanced feature representation containing spatiotemporal correlation information.

5. The deep learning-based large-scale model system for early warning of chronic neuropathic pain after lumbar endoscopic surgery according to claim 4, characterized in that, The specific steps for constructing an individualized risk evolution trajectory map include: The enhanced feature representation is input into a conditional generative adversarial network to simulate the change curve of pain sensitivity under different medication regimens; A neural conduction dynamics model was constructed based on a physical information neural network to predict the extent to which the spread of inflammatory factors in the surgical area affects nerve root function. By simulating the dynamic game process between the immune system and the neural repair system through multi-agent reinforcement learning, a set of possible recovery paths is generated. All simulation results are integrated to form a three-dimensional risk evolution surface with probability weights, namely the individualized risk evolution trajectory map.

6. The deep learning-based large-scale model system for early warning of chronic neuropathic pain after lumbar endoscopic surgery according to claim 5, characterized in that, The specific steps for outputting the early warning signal for chronic neuropathic pain include: Define a baseline recovery corridor based on clinical guidelines on the individualized risk evolution trajectory map; A variational autoencoder is used to detect the multidimensional distance between real-time monitoring data and the baseline recovery corridor, and the abnormal deviation index is calculated. When the abnormal deviation index continuously exceeds the dynamic critical value set by the adaptive threshold decision model, a graded early warning mechanism is triggered. Generate a structured early warning report containing risk level, possible triggers, and intervention recommendations, which is the aforementioned chronic neuropathic pain risk warning signal.

7. The deep learning-based large-scale model system for early warning of chronic neuropathic pain after lumbar endoscopic surgery according to claim 3, characterized in that, The specific steps for spatiotemporal alignment and missing value imputation of the original data stream using a dynamic feature extraction network include: A spatiotemporal alignment module is constructed to identify timestamp offsets in multi-source data streams and synchronize heterogeneous data to a unified time axis through an interpolation algorithm; Deploy a missing value detection unit to scan for discontinuities in the data stream and generate imputation values ​​based on a time series prediction model; Graph attention networks are used to model the spatial dependencies between different monitoring indicators, and feature representations are adjusted to reflect the synergistic changes among physiological parameters. A multi-task learning framework is adopted to jointly optimize the spatiotemporal alignment and missing value imputation process, minimizing the weighted sum of reconstruction error and prediction bias.

8. The deep learning-based large-scale model system for early warning of chronic neuropathic pain after lumbar endoscopic surgery according to claim 4, characterized in that, The specific steps for inputting the standardized temporal feature matrix into the hierarchical attention feature fusion network and calculating the interaction weights of different modal features at multiple time scales include: Local attention mechanisms are used on an hourly scale to capture short-term feature fluctuation patterns and calculate the local correlation score of modal features at each time point. A recurrent attention unit is applied on a daily scale to aggregate the feature change trends over 24 hours, generating a daily feature importance distribution map; Introduce a global attention pooling operation on a weekly scale to compress the time dimension and preserve long-term dependencies; Attention weights across three timescales are dynamically integrated through a gating fusion mechanism to output a modal feature interaction weight matrix.

9. The deep learning-based large-scale model system for early warning of chronic neuropathic pain after lumbar endoscopic surgery according to claim 5, characterized in that, The specific steps for using the risk trajectory simulation engine to deduce the evolutionary path of neural functional states under different combinations of physiological parameters include: A spatial mapping model for neurophysiological parameters is constructed to project multidimensional physiological parameters onto a low-dimensional manifold to capture key feature combinations. The dynamic changes in nerve conduction velocity with the concentration of inflammatory factors are simulated using neural differential equations. We introduce stochastic processes to model the impact of external disturbances on neural functional states, generating a set of probabilistic evolutionary paths; The path integral method is used to calculate the transition probability of neural functional states under different parameter combinations, forming an evolutionary path map.

10. The deep learning-based large-scale model system for early warning of chronic neuropathic pain after lumbar endoscopic surgery according to claim 5, characterized in that, The specific steps of simulating the dynamic game process between the immune system and the neural repair system through multi-agent reinforcement learning include: Design an intelligent agent for the immune system whose decision-making strategy is based on changes in inflammatory factor concentration and immune cell activity; Construct an intelligent agent for a neural repair system, whose action space includes axonal regeneration rate and myelin repair intensity; Establish a game-theoretic reward function to quantify clinical benefit indicators of the balance between immune response and neural repair; A multi-agent policy gradient algorithm is used to train agent agents to simulate the equilibrium solution space in a long-term dynamic game.

Images