An assessment device and assessment model for objective assessment of pain

Abstract

The present application relates to a kind of pain objective evaluation evaluation device and evaluation model, evaluation device includes processor, processor includes frequency domain transformation module, frequency band separation module, first evaluation submodule and second evaluation submodule.Frequency domain transformation module generates time-frequency feature matrix;Frequency band separation module is divided into five frequency bands related to pain perception in the frequency domain with time-frequency feature matrix;First evaluation submodule extracts end time step feature and adjacency matrix from time-frequency feature matrix;Based on end time step feature and adjacency matrix, generate the first feature vector representing the global correlation mode between the electrodes for collecting EEG signal;Second evaluation submodule generates the second feature vector including the local space-time dynamic characteristics of EEG signal;The first feature vector is spliced with the second feature vector to form fusion feature vector;The class probability distribution of fusion feature vector is calculated and normalized, and the pain objective quantification index corresponding to EEG signal is generated.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of pain assessment, in particular to an evaluation device and an evaluation model for objective evaluation of pain. BACKGROUND

[0002] Chronic pain refers to an unpleasant sensory and emotional experience caused by actual or potential tissue damage, or a similar experience. Pain is considered the fifth vital sign, as important as respiration, blood pressure, pulse and body temperature, and is an important indicator for measuring and monitoring a person's health. As the fifth vital sign, the deep significance of pain lies in its embodiment of the temperature of humanistic care in the development of medicine, so that the observation of the vital signs of the subject is not only the collection of physiological data, but also the deep care and concern for life.

[0003] However, as the fifth vital sign, pain also faces an important congenital problem, namely the lack of objective biomarkers. At present, pain scoring mainly relies on self-evaluation by the subject and professional evaluation by the doctor, and is evaluated by different scales (such as visual analog scale (VAS), numerical rating scale (NRS), and DN4 and IDpain scales for neuropathic pain). A significant problem with this scoring method is the lack of objective indicators.

[0004] Existing pain scores are usually obtained through subjective judgment by the subject or the doctor, and the subject will choose a number within the range of 0-100 to represent the severity of the pain. For example, 0-39 represents mild pain, 40-69 represents moderate pain, and 70-100 represents severe pain. This scoring method based on subjective feeling makes the clinical evaluation of pain unable to accurately reflect the true clinical situation, and also has a great negative impact on the development of clinical diagnosis and treatment plans, the evaluation of treatment effects, and even medical-related identification.

[0005] Electroencephalogram (EEG) equipment is one of the main tools for studying pain. The brain signals recorded by EEG reflect the voltage fluctuations produced by neuron firing. Recent studies have shown that voltage changes in different brain regions are associated with the type and degree of pain. However, this voltage change cannot accurately reflect the severity of the subject's pain. In addition, EEG equipment is expensive, data processing is complex, and signal acquisition needs to be carried out in specialized hospitals. Although in recent years scientific research has found that EEG has certain advantages in evaluating the degree of pain, due to limitations such as signal acquisition and expert analysis, it has not yet become an objective indicator for measuring pain.

[0006] CN115867183A discloses a method for monitoring pain level, which comprises receiving measurement data, the data being derived from electroencephalogram (EEG) data collected by one or more EEG electrodes; extracting a set of indicators from the EEG signal, the indicators being associated with power in the theta-alpha frequency range, and determining a level of pain (LoP) based thereon, the LoP being a value indicative of the degree of pain experienced by the subject. However, the evaluation model in this technical solution fails to objectively reflect the severity of pain.

[0007] By observing the pain and non-pain through functional magnetic resonance imaging (fMRI), it can be confirmed that the difference in whole brain functional connectivity between frontal cortex and anterior cingulate cortex. Based on this research result, electroencephalogram (EEG) is further verified, and the change mechanism of electrical activity of different brain areas during the occurrence and relief of pain is revealed through electroencephalogram monitoring in various scenes such as sitting, two-dimensional picture, virtual reality (VR) and the like.

[0008] For another example, CN112957014A discloses a pain detection and positioning method based on brain waves and neural networks, including the following steps: using independent component analysis algorithm to remove noise of original brain electrical signals, performing pain level segmentation of brain electrical signals, dividing each pain level into multiple equal-length time windows, obtaining multi-channel brain electrical time series, and obtaining preprocessed pain data set; generating spectrum topographic maps of each time window of Theta, Alpha and Beta frequency bands related to pain through Fourier transform, azimuth equidistant projection and CloughTocher interpolation algorithm, merging into multi-channel brain wave sequence as input of CNN LSTM-AM neural network; constructing CNN-LSTM-AM neural network, training CNN-LSTM-AM neural network, and obtaining time-space feature vector of brain electrical wave (EEG) sequence related to pain degree and pain location through CNN-LSTM-AM neural network; constructing Softmax pain classifier model, inputting brain electrical wave pain features learned by CNN-LSTM-AM neural network into pain classifier model, and matching and identifying pain level and pain location. The technical solution extracts spatial features of brain electrical signals through convolutional neural network, and captures time sequence characteristics through LSTM, but the feature extraction process does not consider the dynamic functional connection relationship between different brain regions. Specifically, although the brain electrical topographic map generated by independent component analysis and frequency band segmentation can reflect the activity of local brain regions, it cannot effectively model the neural oscillation synchronization across brain regions (such as the theta band coupling of thalamus- anterior cingulate pathway) in the pain state. The lack of this adjacent node correlation calculation makes it difficult for the calculation model to capture the topological structure changes of the key brain network (such as the default mode network) in the pain transmission process, limiting the integrity of the feature expression ability. In addition, the technical solution uses a static weight allocation strategy to handle brain region spatial relationships, ignoring the adaptive characteristics of functional connections in the dynamic evolution process of pain, which may cause delay in identifying sudden pain response.

[0009] The technical solution only divides the data after Fourier transform into brain electrical signals of three frequency bands of Theta (θ), Alpha (α) and Beta (β), lacking signals of Delta (δ) and Gamma (γ) frequency bands. The brain electrical signals of Theta (θ) frequency band are related to the persistent neural activity and central sensitization of chronic pain. The brain electrical signals of Gamma (γ) frequency band are directly related to the spatiotemporal coding of pain stimuli. Therefore, the three frequency band mode of the technical solution results in incomplete spectral coverage of the brain, limiting the complete extraction of pain-related neural oscillation features (such as slow wave activity and high frequency synchronization), and making it difficult to reorganize and model the whole brain network related to pain.

[0010] The Softmax pain classifier model in the technical solution is used to classify and identify the pain level and location based on the brain wave signal, rather than precise objective quantification. The Softmax pain classifier model includes 3 pain level labels, namely no pain (0), mild pain (1), moderate pain (2), and severe pain (3). If the pain level is detected to be 1-3, the Softmax pain classifier model is used to continue detecting the pain location, which is divided into left hand, right hand, left foot, and right foot four body parts, to verify the effectiveness of pain positioning. This classification task focuses more on distinguishing the differences between different categories, rather than accurately measuring the intensity or degree of each category.

[0011] The current classification level cannot achieve objective quantification of the degree of pain, and cannot obtain the objective pain index of the degree of pain. Due to the lack of temperature parameters that can simulate the randomness and dynamic changes of brain neural activity, the Softmax pain classifier model cannot control the influence of noise on the correlation strength through temperature parameters, resulting in insufficient modeling of the randomness of neural activity. For example, the neural signals in some pain scenarios may be over-amplified or inhibited due to noise, affecting the accurate quantification of the degree of pain. At the same time, pain perception is dynamically changing, and a model lacking temperature parameters will process all inputs with a fixed scale, unable to capture the real-time changes in functional connectivity during the pain process (such as the difference in neural activity patterns when the pain intensifies or subsides). This makes the Softmax pain classifier model unable to adjust the differences in neural signals of different individuals, leading to inaccurate quantification of the degree of pain for some groups of people (such as patients with lower neural sensitivity or different pain thresholds).

[0012] CN118262918A discloses a machine learning-based personalized tumor patient pain management system, which includes a data collection and processing module for collecting patient pain diaries, physiological data including heart rate, blood pressure, pain medication use records, medical images and genomic data, and annotating, cleaning and normalizing the data; further including a pain pattern recognition module for extracting key features from the processed data, and using machine learning algorithms including support vector machines, random forests or deep learning to identify different pain patterns and predict pain onset; using a personalized pain prediction model, using clustering algorithms to group patients according to pain patterns, physiological responses and treatment effects, and establishing a customized prediction model for each group of patients to predict the severity and onset time of pain; finally introducing a treatment plan generator based on a machine learning decision support system, providing personalized pain management recommendations including drug dosage adjustment and non-drug treatment recommendations based on the output of the prediction model, and including a feedback mechanism to adjust model parameters according to patient pain diaries and treatment response. In this technical solution, the graph attention scoring function g(u,v;λ) is used to determine the relationship strength between different data source features. However, if the calculation method of relationship strength in this technical solution is simply used to calculate the relationship strength of electroencephalogram signals in the process of objective evaluation of pain intensity, problems will arise: the neural response of pain usually has a time course accumulation effect, and due to the lack of time decay mechanism, the functional connectivity strength of chronic pain patients is overestimated, affecting the capture accuracy of the calculation model for the dynamic evolution of pain. The technical solution further uses a multi-modal synchronization function to map data from different sources to a unified feature space, which will cause the disappearance of the phase-amplitude coupling (PAC) phenomenon in the delta band (0.5-4Hz) and theta band (4-8Hz), and the phase-amplitude coupling (PAC) phenomenon is a key feature of pain central regulation. Therefore, this step will reduce the specific value of the objective quantitative indicator of pain intensity prediction.

[0013] As shown above, the pain detection method and neighbor node algorithm in the current prior art cannot be used for objective evaluation of the degree of pain. On this basis, the present application proposes a new calculation method, which is verified and repeatedly fitted by means of a model to build a "digital twin pain brain". Through mature EEG equipment to collect data, the system imports the calculation results into the calculation model, uses the established digital twin pain brain model for calculation and comparison, and finally presents an objective digital indicator. This method first unifies pain, the fifth vital sign, with other vital signs such as body temperature and pulse in a visual, objective and quantitative form, opening up a new path for objective evaluation of pain.

[0014] In addition, on the one hand, due to the difference in understanding of those skilled in the art; on the other hand, due to the fact that the applicant has studied a large number of literatures and patents when making the present application, but limited by the size and has not listed all the details and contents in detail, however, this does not mean that the present application does not have these prior art characteristics, on the contrary, the present application has all the characteristics of the prior art, and the applicant reserves the right to add relevant prior art in the background art. SUMMARY

[0015] Pain as the fifth vital sign faces fundamental challenges in clinical practice - lack of objective biomarkers. Current assessment mainly relies on subject self-report and scale tools (such as VAS, NRS, DN4, IDpain), the core problem of which is the subjective dominant evaluation mode: the subject's description of pain is easily disturbed by emotional state, cognitive level and environmental factors. This subjective dependence on the other hand leads to the evaluation results cannot accurately reflect the clinical reality, for example, anxiety may lead to high scores, while communication disorders may not accurately convey the true pain due to limited expression ability; on the other hand, different medical staff's pain scores for the same subject directly affect the development of diagnosis and treatment and the judgment of treatment effect, especially in complex pain management that requires multidisciplinary collaboration, the contradiction is more prominent.

[0016] Traditional assessment methods also have inherent defects in dynamic monitoring. Scale scores relying on single or discrete time points are difficult to capture real-time fluctuations in pain intensity, such as intra-day variation of postoperative pain or sudden exacerbation of neuropathic pain. More importantly, the complex mechanisms involved in chronic pain, such as central sensitization and neural plasticity changes, cannot be fully reflected by simple numerical scoring system, leading to superficial characterization of the pathological nature. These problems are further amplified in special groups (such as children, elderly subjects and conscious impaired groups), which may lead to the dual risk of overuse or underuse of analgesic drugs, directly affecting the rationality of clinical decision-making.

[0017] In view of the deficiencies of the prior art, the present application provides, from a first aspect, an evaluation device for objective evaluation of pain, comprising a processor including a frequency domain transformation module, a frequency band separation module, a first evaluation submodule and a second evaluation submodule. The frequency domain transformation module performs fast Fourier transform on the EEG signal related to pain to generate a time-frequency feature matrix; the frequency band separation module divides the time-frequency feature matrix of the EEG signal in the frequency domain into time-frequency feature matrices of five frequency bands related to pain perception, i.e., δ, θ, α, β and γ; the first evaluation submodule extracts time series features of the EEG signal from the time-frequency feature matrix based on a bidirectional gated recurrent unit; end time step features and an adjacency matrix are extracted from the time series features; a first feature vector representing the global correlation pattern between the electrodes for collecting the EEG signal is generated based on the end time step features and the adjacency matrix; the second evaluation submodule performs convolution processing on the time series features to generate a second feature vector including local spatiotemporal dynamic features of the EEG signal; the first feature vector and the second feature vector are spliced along the feature dimension to form a fusion feature vector including the global correlation pattern between the electrodes and the local spatiotemporal dynamic features; the class probability distribution of the fusion feature vector is calculated and normalized to generate a pain objective quantification index corresponding to the EEG signal.

[0018] The frequency domain transformation module of the present application converts the EEG signal into a time-frequency feature matrix through fast Fourier transform (FFT), and in combination with the learnable filter bank of the frequency band separation module, realizes dynamic segmentation of pain-related frequency bands (δ, θ, α, β, γ). The conventional fixed frequency band division method cannot adapt to individual spectral differences (such as the α band shift of elderly patients), while the filter parameter of the present scheme is adaptively adjusted to the frequency band boundary through end-to-end training, accurately matching the individualized spectral characteristics. The cross-frequency band attention mechanism further integrates the energy distribution and phase coupling characteristics of different frequency bands, revealing the pain-related neural oscillation coordination mode, and providing more physiologically meaningful basic features for objective quantification.

[0019] The first evaluation submodule extracts the long-term dependence of the time series based on the bidirectional gated recurrent unit (BiGRU), and generates a dynamic adjacency matrix using the Gumbel-Softmax unit, and generates a first feature vector representing the global correlation pattern between the electrodes (210) for collecting the EEG signal using the graph aggregation unit, breaking through the limitations of the traditional pre-defined brain network template. This technology can capture the changes in the functional connectivity strength of the brain regions induced by pain in real time (such as the transient activation of the thalamus-insular pathway), while the second evaluation submodule extracts local high-frequency oscillation features (such as the γ band transient response of the anterior cingulate gyrus) through a convolutional neural network. The fusion of global brain network features and local spatiotemporal features comprehensively represents the neural dynamics of pain, overcoming the problem of information loss in a single feature dimension.

[0020] The present application maps the fusion features to the Markov value interval of 0-100 points through the Softmax function, which is compatible with the existing pain evaluation system (such as the VAS / NRS scale), and the quantification result does not depend on the language feedback of the subject. The traditional EEG analysis method needs to rely on complex manual feature engineering, while the present scheme realizes end-to-end automatic processing, eliminates subjective interpretation bias, and provides an objective evaluation tool for special groups such as aphasia patients and children.

[0021] According to a preferred embodiment, the processor further comprises a Bayesian updating module, which receives the objective quantification index of pain, and calculates the 95% confidence interval of the objective quantification index of pain according to the model cognitive uncertainty and the observation noise level, to generate the reliability of the 95% confidence interval evaluation result; if the confidence interval width exceeds the preset threshold, the Bayesian updating module dynamically adjusts the observation noise parameter to optimize the feature weight distribution. The Bayesian updating module dynamically adjusts the feature weight through the variational inference algorithm and quantifies the prediction uncertainty. This mechanism automatically attenuates the contribution of the interference frequency band in view of the device difference and individual physiological noise (such as electromyographic artifacts), significantly improving the cross-device consistency and anti-interference ability. The traditional static model is easily affected by hardware parameters or environmental noise, while the present scheme optimizes the model parameters through real-time feedback to ensure the stability of the pain score in different scenarios.

[0022] According to a preferred embodiment, the frequency band separation module extracts the time-frequency energy distribution features in the time-frequency feature matrix of each frequency band based on a learnable frequency domain filter, and dynamically allocates weights to the time-frequency energy distribution features extracted from the time-frequency feature matrix of each frequency band based on a cross-frequency band attention mechanism, to fuse the energy and phase coupling characteristics between frequency bands and generate a time-frequency feature matrix with frequency band specificity.

[0023] The frequency band separation module significantly improves the dynamic representation ability of the time-frequency energy distribution features through the learnable frequency domain filter and the cross-frequency band attention mechanism. Its innovation lies in breaking through the limitations of traditional fixed frequency band division, accurately capturing the differences in neural oscillation patterns of different frequency bands through adaptive learning of frequency domain filter parameters, and dynamically fusing the energy and phase coupling characteristics across frequency bands using the attention mechanism. This multi-frequency band collaborative modeling technology first realizes the nonlinear decoupling of the frequency domain features of pain-related neural activity, providing an interpretable quantitative basis for revealing the frequency band specificity of the physiological mechanism of pain.

[0024] According to a preferred embodiment, the step of extracting the adjacency matrix by the first evaluation submodule comprises: converting the end time step features into a connection graph between electrodes based on the Gumbel-Softmax method, thereby forming an adjacency matrix representing the coupling degree of neural activity between brain cortical regions.

[0025] The Gumbel-Softmax-based graph generation network innovatively constructs a mathematical representation method of a dynamic functional connection topology. By converting the end time step features into a connection graph with a probability significance, the technology effectively solves the problem of loss of spatiotemporal coupling characteristics in traditional static brain network modeling. The adjacency matrix generated by the invention not only quantifies the neural activity synchronization strength between the brain cortex regions, but also retains the temporal correlation of the dynamic evolution of the functional connection through the differentiable characteristics.

[0026] According to a preferred embodiment, the step of calculating the class probability distribution of the fusion feature vector by the second evaluation submodule includes: calculating the class probability distribution of the fusion feature vector, outputting the normalized value in the [0, 1] interval based on the Softmax function, and mapping to the pain objective quantification index in the [0, 100] interval through linear transformation.

[0027] The Marke value standardization system output by the Softmax function realizes the continuous probability mapping of the pain intensity through an end-to-end deep learning framework. This breaks through the discretization limitation of the traditional classification model, and uses the multi-scale feature fusion capability of the second evaluation submodule to construct a smooth mapping relationship from the neural features to the quantification index. The probability output form not only guarantees the clinical interpretability of the value in the 0-100 interval, but also reduces the heterogeneity influence of the cross-subject data distribution through nonlinear normalization processing, providing a technical foundation for the index unification of multi-center research.

[0028] According to a preferred embodiment, the pain objective quantification index is the Marke value, the numerical interval is 0-100, and the unit is Marke; the greater the Marke value of the pain intensity mapping, the more intense the pain.

[0029] The standardized construction of the Marke value system marks the paradigm shift of pain quantification from subjective description to objective measurement. By strictly defining the numerical interval, unit and intensity mapping relationship, the invention first establishes a neural biomarker system with clear clinical semantics. The continuous value characteristic supports the detection of subthreshold changes in pain intensity, and the 100-level quantization gradient is significantly better than the 11-level division of the traditional VAS scale, providing a high-sensitivity quantification tool for pain dynamic monitoring and efficacy evaluation in the precision medicine scenario.

[0030] The application provides a pain objective evaluation evaluation model from a second aspect, and the evaluation model is run by an evaluation module, and the evaluation module comprises a first evaluation submodule and a second evaluation submodule; the first evaluation submodule is used for extracting time sequence features of an EEG signal from a time-frequency feature matrix based on a bidirectional gate recurrent unit; end time step features and an adjacency matrix are extracted from the time sequence features of the EEG signal related to pain; a first feature vector representing a global correlation mode between electrodes for collecting the EEG signal is generated based on the end time step features and the adjacency matrix; a fusion feature vector between the global correlation mode and a local space-time dynamic feature; a class probability distribution of the fusion feature vector is calculated and normalized to generate a pain objective quantification index corresponding to the EEG signal; wherein the time sequence features of the EEG signal related to pain are extracted from a time-frequency feature matrix, and the time-frequency feature matrix is generated by performing fast Fourier transform on the EEG signal related to pain; the time sequence features are divided into five frequency bands related to pain perception, i.e., δ, θ, α, β and γ, by a frequency band separation module in a frequency domain.

[0031] According to a preferred embodiment, the step of extracting the adjacency matrix by the first evaluation submodule comprises: converting the end time step features into a connection graph between electrodes based on a Gumbel-Softmax method, so as to form the adjacency matrix representing the coupling degree of neural activity between the cerebral cortex regions.

[0032] According to a preferred embodiment, the step of calculating the class probability distribution of the fusion feature vector by the second evaluation submodule comprises: calculating the class probability distribution of the fusion feature vector, outputting a normalized value in the interval [0, 1] based on a Softmax function, and mapping to a pain objective quantification index in the interval [0, 100] through linear transformation.

[0033] According to a preferred embodiment, the pain objective quantification index is a Marke value, the numerical interval is 0-100, and the unit is Marke; the greater the Marke value of the pain intensity mapping is, the more intense the pain is. BRIEF DESCRIPTION OF DRAWINGS

[0034] Figure 1 FIG. 1 is a simplified module connection relationship schematic diagram of a pain objective evaluation device provided by the application;

[0035] Figure 2 FIG. 2 is another simplified module connection relationship schematic diagram of a pain objective evaluation device provided by the application;

[0036] Figure 3 FIG. 3 is an electroencephalogram of a first subject provided by the application;

[0037] Figure 4 FIG. 4 is a pain data schematic diagram of the first subject provided by the application;

[0038] Figure 5 EEG of the second subject provided by the present application;

[0039] Figure 6 Pain data diagram of the second subject provided by the present application;

[0040] Figure 7 Classification result diagram of the confusion matrix provided by the present application;

[0041] Figure 8 Regression scatter plot of the Marke value and the true pain score provided by the present application.

[0042] LIST OF REFERENCE NUMERALS

[0043] 100: evaluation device; 110: processor; 111: signal receiving module; 112: signal output module; 120: filtering module; 130: first evaluation submodule; 131: normalization unit; 132: bidirectional gated recurrent unit; 133: Gumbel-Sampler unit; 134: graph aggregation unit; 140: second evaluation submodule; 141: convolution unit; 142: splicing unit; 143: classification unit; 150: Bayesian update module; 160: signal processing module; 170: frequency domain transformation module; 180: frequency band separation module; 190: evaluation module; 200: electroencephalogram acquisition device; 210: electrode; 220: signal amplifier; 230: filter; 240: analog-to-digital converter; 250: data acquisition system; 260: electroencephalogram signal output module; 300: terminal device; 310: analog signal input end; 320: analog signal converter; 330: display unit; 340: interaction unit. DETAILED DESCRIPTION

[0044] The following will be described in detail in conjunction with the accompanying drawings.

[0045] Pain as the fifth vital sign faces fundamental challenges in clinical practice—lack of objective biomarkers. Current evaluation mainly relies on self-reporting and scale tools (such as VAS, NRS, DN4, IDpain) of the subjects, and the core problem lies in the subjective evaluation mode: the subjects' description of pain is easily disturbed by emotional state, cognitive level and environmental factors. This subjective dependence on the other hand leads to the evaluation results that cannot accurately reflect the clinical practice, for example, anxiety may lead to high scores, and communication-impaired subjects may not be able to accurately convey the true pain due to limited expression ability; on the other hand, different medical staff's pain scores for the same subject directly affect the development of diagnosis and treatment programs and the judgment of treatment effect, especially in complex pain management that requires multidisciplinary cooperation, the contradiction is more prominent.

[0046] The traditional evaluation method also has inherent defects in dynamic monitoring. It is difficult for scale scores at a single time point or discrete time points to capture real-time fluctuations in pain intensity, such as the daily variation of postoperative pain or the transient exacerbation of sudden neuropathic pain. More importantly, the complex mechanisms involved in chronic pain, such as central sensitization and neural plasticity changes, cannot be fully reflected by a simple numerical scoring system, resulting in a superficial characterization of the pathological nature. These problems are further magnified in special groups (such as children, elderly subjects, and people with consciousness disorders), which may lead to the dual risks of overuse or insufficient use of analgesic drugs, directly affecting the rationality of clinical decision-making.

[0047] At the research and medical management level, the limitations of subjective evaluation also form a bottleneck. For example, the lack of uniform evaluation criteria in clinical trials may reduce the sensitivity of efficacy determination, increasing the risk of bias in research conclusions. In addition, the lack of objective indicators also hinders the comparison of cross-institutional data and the standardization of diagnosis and treatment, limiting the overall development of the pain management field.

[0048] Based on the shortcomings of the prior art, the present application hopes to form an objective index representing the degree of pain based on the EEG signal and the evaluation module 190. The model integrated in the evaluation module 190 is also collectively referred to as the digital pain brain model.

[0049] The core principle of the present application is:

[0050] The time-frequency feature matrix of the EEG signal is divided into five frequency bands related to pain perception, i.e., the time-frequency feature matrices of the delta, theta, alpha, beta, and gamma bands in the frequency domain. The time series features of the EEG signal are extracted from the time-frequency feature matrix. The end time step features and the adjacency matrix are extracted from the time series features. The first feature vector representing the global correlation pattern between the electrodes used to collect the EEG signal is generated based on the end time step features and the adjacency matrix. The time series features are convolved to generate the second feature vector including the local spatiotemporal dynamic features of the EEG signal. The first feature vector and the second feature vector are spliced along the feature dimension to form the fusion feature vector including the global correlation pattern between the electrodes and the local spatiotemporal dynamic features. The class probability distribution of the fusion feature vector is calculated and normalized to generate the pain objective quantification index corresponding to the EEG signal. The Bayesian update module receives the pain objective quantification index, and according to the model cognitive uncertainty and the observation noise level, the 95% confidence interval of the pain objective quantification index is calculated to generate the credibility of the 95% confidence interval evaluation result. If the confidence interval width exceeds the preset threshold, the Bayesian update module dynamically adjusts the observation noise parameter to optimize the feature weight distribution. Preferably, the training data used for training the evaluation module 190 in the present application is the EEG signal related to pain. The training data can be sourced from multiple databases.

[0051] Preferably, the objective quantification index of pain proposed by the present application is the Marke value. The Marke value is generated by data fusion of multi-modal biological information extraction, reflecting the physiological and psychological response of the subject to pain. The processor 110 in the present application processes the biological information through a multi-task learning model, and converts the analog signal of the pain response into a quantification index with a numerical interval of 0-100. These quantification indexes are called Marke values. The greater the Marke value of the pain intensity mapping, the more intense the pain, that is, the stronger the pain degree (see Table 1).

[0052] The objective quantification index of pain of the present application is compatible with existing clinical scales (such as VAS 0-10 points, NRS 0-10 levels), and can be directly converted by dividing by 10. 1Marke corresponds to a distinguishable change in pain intensity. 1Marke is defined as the fluctuation range of neural signals of healthy subjects in a resting state.

[0053] According to the above core principle, the present application can provide a digital twin pain perception modeling method and device; also provide an EEG signal driven pain quantification analysis system and method; can also provide an objective pain measurement algorithm and its application device; can also provide an EEG-based pain assessment model generation method and device; can also provide a vital sign visualization and digital processing method; can also provide a multi-dimensional pain index calculation and display platform; can also provide an intelligent pain monitoring and analysis system and method; can also provide a high-precision pain brain map modeling and real-time monitoring method; can also provide a digital twin solution for individualized pain management.

[0054] The present application can also provide a storage medium storing an encoded program of an evaluation model to execute the evaluation method of the objective evaluation of pain of the present application. The present application also relates to a processor 110. The processor 110 can execute the evaluation method of the objective evaluation of pain of the present application based on the evaluation module 190, and output objective index information of pain. The objective index information can be data information or an analog signal, as long as the output is such that the terminal can read and display the objective index of pain.

[0055] Embodiment 1

[0056] This embodiment illustrates the method of the present application by taking the evaluation device 100 for objective evaluation of pain as an example.

[0057] Preferably, the evaluation device 100 of the present application is connected to the electroencephalogram acquisition device 200 in a wired or wireless manner. The evaluation device 100 can also be connected to at least one terminal device 300 in a wired or wireless manner, so that the terminal device 300 can receive the objective quantitative indicators of pain for characterizing the degree of pain and the related information thereof and the related information of the processing process of the electroencephalogram signal by the evaluation device.

[0058] In particular, as shown in Figure 1 , the evaluation device 100 can include a processor 110. The hardware entity of the processor 110 can be a GPU, a CPU, etc. The processor 110 can also be equipped with a memory. The memory is used to store the data received by the processor 110, in the processing process and finally generated. Preferably, the processor 110 includes a plurality of hardware modules and a collection thereof.

[0059] Preferably, as shown in Figure 1 , the evaluation device 100 can include a processor 110 and a filtering module 120. That is, the filtering module 120 is arranged upstream of the data reception of the processor 110 to pre-process the received EEG signal, such as noise reduction, filtering, etc. Preferably, the filtering module 120 can also be integrated in the processor 110 as a part of the processor 110. Preferably, as shown in Figure 1 , the processor 110 can further include a Bayesian update module 150, a data processing module 160, a frequency domain transformation module 170, a frequency band separation module 180 and an evaluation module 190.

[0060] Preferably, the data processing module 160 is composed of an analog-to-digital converter (ADC) and a digital signal preprocessing unit, for example, to realize signal sampling and digitization processing. The frequency domain transformation module 170 can adopt a digital signal processor (DSP) or a field programmable gate array (FPGA), which is dedicated to performing mathematical operations such as fast Fourier transform. The frequency band separation module 180 can be a filter bank or a frequency division circuit, which is used for frequency band division function. The evaluation module 190 can be an embedded microprocessor or a neural network acceleration chip, on which an evaluation model is deployed, such as a bidirectional gated recurrent unit 132 (BiGRU) and a Gumbel-Sampler unit 133. That is, the evaluation model is run by the evaluation module 190.

[0061] Preferably, the data processing module 160 transmits the resampled EEG signal to the frequency domain transformation module 170 through a high-speed parallel bus, which ensures the integrity and real-time performance of the sampling data. A bidirectional data channel is adopted between the frequency domain transformation module 170 and the frequency band separation module 180, which not only transmits the time-frequency feature matrix of the EEG signal generated by the Fourier transform, but also retains the transmission path of the original time sequence. The frequency band separation module 180 divides the received time sequence features of the EEG signal into five frequency bands through a multiplexing interface, and distributes the time-frequency feature matrix of each frequency band to the evaluation module 190. This star-shaped topology supports parallel frequency band feature analysis. The timing synchronization between modules is uniformly coordinated by the internal clock tree of the processor 110, and data interaction is realized through shared memory or direct memory access (DMA) mechanism, which ensures the continuity of the system processing pipeline. In the case of integrated design, the evaluation module 190 can be packaged in a system-on-a-chip (SoC), interconnected through a network-on-chip (NoC), and the external expansion module is connected through a standard communication interface.

[0062] As shown in Figure 1 , the evaluation device 100 further comprises a signal receiving module 111 and a signal output module 112. The signal receiving module 111 is used to receive the electroencephalogram signal. The signal output module 112 is used to output the generated objective index for representing the degree of pain and its related information, and can also output the related information of the processing process of the electroencephalogram signal by the evaluation device 100.

[0063] As shown in Figure 1 , the electroencephalogram acquisition device 200 is used to acquire the EEG signal of the subject. The electroencephalogram acquisition device 200 comprises a plurality of electrodes 210, a signal amplifier 220, a filter 230, an analog-to-digital converter 240 (ADC), a data acquisition system 250 and an electroencephalogram signal output module 260. The plurality of electrodes 210 can be connected to the signal amplifier 220 through wireless or wired means to receive the electrode 210 signal of the subject. The signal amplifier 220 is used to enhance the EEG signal, and the filter 230 is used to remove noise and interference. The analog-to-digital converter 240 converts the analog signal into a digital signal for further processing. The electroencephalogram signal output module 260 is connected to the signal receiving module 111 of the evaluation device 100, thereby transmitting the EEG signal generated by the electroencephalogram acquisition device 200 to the evaluation device 100, so that the evaluation device 100 performs Fourier transform on the EEG signal.

[0064] As shown in Figure 1As shown, the terminal device 300 can be a mobile portable electronic device such as a smartphone or a tablet, or a non-mobile computer terminal such as a desktop computer or a workstation. The terminal device 300 comprises one or several of an analog signal input 310, an analog signal converter 320, a display unit 330, an interactive unit 340, and a processing unit.

[0065] The EEG signals outputted by the EEG acquisition device 200 can also be inputted into the analog signal input 310 of the terminal device 300, for example through a 3.5mm audio jack or a dedicated interface. The analog signal converter 320 is used to convert these analog signals into digital signals, ensuring that the signals can be processed by the computer. The analog signal converter 320 can be an internal analog-to-digital converter or an external audio interface device. The display unit 330 is used to display the processed EEG signals and their processing process in real time. This can be a high-resolution LCD screen or a touch display, allowing users to visually observe and analyze the signals. The interactive unit 340 comprises input devices such as a keyboard, a mouse, or a touch screen, which allow users to interact with the terminal device 300 through clicking, sliding, or inputting commands. The processing unit, for example, is a central processing unit (CPU) or a graphics processing unit (GPU) built into the terminal device 300, which is used to perform complex signal analysis and data processing tasks. The analog signal input 310 of the terminal device 300 is connected to the signal output module 112 of the evaluation device 100 in a wired or wireless manner, ensuring that the scores of the objective indicators of pain can be displayed and monitored.

[0066] As shown in FIG. 1, the evaluation device 100 comprises a signal acquisition device 200, a signal processing device 300, and a signal output module 112. Figure 1 and Figure 2As shown, the electroencephalogram acquisition device 200 is connected with the evaluation device 100 through the signal receiving module 111 to transmit the acquired EEG signal to the evaluation device 100 through the signal receiving module 111. The signal receiving module 111 is connected with the filtering module 120, and the filtering module 120 receives the EEG signal from the electroencephalogram acquisition device 200 through the signal receiving module 111 to perform secondary filtering processing. The filtering module 120 transmits the secondary filtered EEG signal to the signal processing module 160. After receiving the secondary filtered EEG signal from the filtering module 120, the signal processing module 160 performs resampling processing on the EEG signal, and transmits the resampled EEG signal to the frequency domain conversion module 170. After receiving the resampled EEG signal from the signal processing module 160, the frequency domain conversion module 170 obtains a time-frequency feature matrix through Fourier transform, and transmits the time-frequency feature matrix to the frequency band separation module 180. After receiving the time-frequency feature matrix from the frequency domain conversion module 170, the frequency band separation module 180 divides the time-frequency feature matrix representing the EEG signal into time-frequency feature matrices of five frequency bands (δ, θ, α, β and γ) related to pain perception in the frequency domain. The frequency band separation module 180 transmits the divided time-frequency feature matrices of the five frequency bands (δ, θ, α, β and γ) to the evaluation module 190. After receiving the divided time-frequency feature matrices of the five frequency bands (δ, θ, α, β and γ) from the frequency band separation module 180, the evaluation module 190 performs calculation of the pain objective quantification index.

[0067] Preferably, as Figure 2 shown, the first evaluation submodule 130 in the evaluation module 190 includes a normalization unit 131, a bidirectional gated recurrent unit 132, a Gumbel-Sampler unit 133 and a graph aggregation unit 134. After receiving the original pain score label from the signal receiving module 111, the normalization unit 131 performs normalization processing to obtain a normalized pain score label. The normalization unit 131 transmits the normalized pain score label to the Bayesian update module 150.

[0068] After receiving the time-frequency feature matrices of the five frequency bands (δ, θ, α, β and γ) from the frequency band separation module 180, the bidirectional gated recurrent unit 132 performs processing and outputs time series features and end time step features. The bidirectional gated recurrent unit 132 transmits the end time step features to the Gumbel-Sampler unit 133. The bidirectional gated recurrent unit 132 transmits the time series features to the convolution unit 141 in the second evaluation submodule 140.

[0069] After receiving the end time step features from the bidirectional gated recurrent unit 132, the Gumbel-Sampler unit 133 calculates the adjacency matrix. The Gumbel-Sampler unit 133 sends the end time step features and the adjacency matrix to the graph aggregation unit 134. After receiving the end time step features and the adjacency matrix from the Gumbel-Sampler unit 133, the graph aggregation unit 134 performs the calculation of the first feature vector. The graph aggregation unit 134 sends the first feature vector to the convolution unit 141 in the second evaluation sub-module 140. After receiving the time series features from the bidirectional gated recurrent unit 132, the convolution unit 141 calculates the second feature vector. The convolution unit 141 sends the first feature vector and the second feature vector to the splicing unit 142 to calculate the fusion feature vector. The splicing unit 142 sends the fusion feature vector to the classification unit 143. After receiving the fusion feature vector from the splicing unit 142, the classification unit 143 calculates the pain objective quantification index and sends it to the Bayesian update module 150.

[0070] After receiving the normalized pain score label from the normalization unit 131, the Bayesian update module 150 constructs a variational loss function. Based on the loss function, the Bayesian update module 150 updates the weight parameters of the bidirectional gated recurrent unit 132 using the reparameterization gradient algorithm. After receiving the pain objective quantification index from the classification unit 143, the Bayesian update module 150 performs confidence interval calculation, combines model cognitive uncertainty and observation noise level to generate the credibility of 95% confidence interval evaluation results; constructs a threshold triggering mechanism, when the confidence interval width exceeds the preset critical value, automatically adjusts the observation noise parameter, recalibrates the feature weight distribution ratio; suppresses outliers and starts the parameter retraining process for prediction results that continuously deviate from the label, reducing the interference of device noise and physiological artifacts.

[0071] Before the evaluation device 100 starts evaluation, the EEG signal of the subject needs to be collected by the electroencephalogram acquisition device 200 (see Figure 3 and Figure 5 ).

[0072] Preferably, a high-precision electroencephalogram acquisition device 200 is used, and appropriate gain and filtering parameters are set to ensure the clarity and accuracy of the signal. The model of the electrode 210 suitable for the subject is selected. In the case of determining the target target area on the head of the subject, the electrode 210 is pasted on the target target, so that the plurality of electrodes 210 directly contact the target area of the head of the subject. Preferably, conductive gel is used to ensure good contact between the electrode 210 and the scalp of the target target area.

[0073] The electrode 210 transmits the original EEG signal to the signal amplifier 220 in wired or wireless manner. The signal amplifier 220 amplifies the micro-volt level brain electrical signal to a processable range while suppressing common mode interference through impedance matching and low noise amplification circuit. The amplified signal enters the filter 230 for pre-processing, in which the high-pass filter (cut-off frequency 0.5 Hz) eliminates the baseline drift, and the low-pass filter (cut-off frequency 75 Hz) filters out the electromyographic noise, thereby forming an analog waveform consistent with the characteristics of bioelectric signals. The analog-to-digital converter 240 (ADC) digitizes the filtered signal at a sampling rate of no less than 250 Hz, and its 24-bit resolution ensures accurate capture of subtle potential changes. The data acquisition system 250 time-aligns and buffers the multi-channel digital signal, and the EEG signal data stream after synchronization is packaged into a data packet consistent with the IEEE 11073 protocol through the EEG signal output module 260.

[0074] The data packet is transmitted to the signal receiving module 110 of the evaluation device 100 through a single-mode optical fiber cable, and then delivered to the memory by the signal receiving module 110. Preferably, the optical fiber interface uses an SFP+ optical module to achieve 10 Gbps high-speed transmission, and its anti-electromagnetic interference characteristics avoid signal-to-noise ratio degradation during transmission. The signal receiving module 111 of the evaluation device 100 can also verify data integrity through the CRC (Cyclic Redundancy Check) verification mechanism.

[0075] The evaluation device 100 of the present application can perform the following operations.

[0076] The received EEG signal is pre-processed. The pre-processing steps of the EEG signal include secondary filtering and resampling.

[0077] In the case where the filtering module 120 is arranged in the evaluation device 100, the filtering module 120 receives the EEG signal through the signal receiving module 111 and performs secondary filtering on the EEG signal to achieve noise reduction. Preferably, the filtering module 120 can also be arranged outside the evaluation device 100, directly connected with the EEG signal output module 260 of the EEG acquisition device 200 and receiving the EEG signal. This embodiment mainly describes the case where the filtering module 120 is arranged in the evaluation device 100.

[0078] Insufficient filtering may have the following defects:

[0079] The EEG signal output by the brain electrical collection device 200 has been preliminarily denoised, but its fixed cutoff frequency (0.5-75 Hz) cannot completely suppress the high-frequency residual noise (such as electromyographic interference harmonics) and the sidelobe energy of power frequency interference in the digital signal. When the EEG signal without secondary filtering is subsequently divided into frequency bands, the high-frequency noise component (>75 Hz) will leak into the beta / gamma frequency band (12-100 Hz) due to spectral leakage, resulting in distortion of the characteristic frequency band power calculation. Moreover, unfiltered transient artifacts (such as eye blinking, motion artifacts) appear as non-stationary spike signals in the time domain, and if feature extraction is performed by the first evaluation submodule 130 within the evaluation device 100, they will be mis-modeled as abnormal brain region functional connectivity. For example, the instantaneous high power value of the frontal electrode due to eye blinking artifacts may be misinterpreted by the first evaluation submodule 130 as a false feature of enhanced frontal lobe-limbic system connectivity.

[0080] Residual power frequency noise (50 Hz and its harmonics) appears as a common mode interference in the spatial dimension for all brain electrodes, which can weaken the ability of the second evaluation submodule 140 to distinguish local brain region-specific activities. Unfiltered common mode noise can mask the true pain-related spatial features (such as increased theta frequency band power in the anterior cingulate gyrus), causing the weight distribution of the second evaluation submodule 140 to deviate from the key brain regions. At the same time, unfiltered signal variations (such as device noise, environmental electromagnetic interference) will introduce random fluctuations unrelated to pain, increasing the risk of overfitting by the first evaluation submodule 130. The first evaluation submodule 130 may mistake noise patterns for pain-specific features, resulting in decreased cross-dataset generalization ability.

[0081] Based on the above insufficient filtering defects, the present application uses the filtering module 120 to perform secondary filtering on the EEG signal, further suppressing residual interference before frequency domain analysis, to avoid the influence of noise components on the extraction accuracy of spatial-temporal features by the first evaluation submodule 130 in the evaluation device 100.

[0082] Preferably, the filtering module 120 includes a high-pass filter and a low-pass filter. The high-pass filter is connected to the signal receiving module 111. The high-pass filter and the low-pass filter are connected through a signal line.

[0083] The high-pass filter (cutoff frequency 0.5 Hz) receives the received EEG signal to remove low-frequency noise such as electrocardiographic interference and motion artifacts. The high-pass filter sends the EEG signal with low-frequency noise filtered to the low-pass filter. Preferably, the low-pass filter (cutoff frequency 75 Hz) in the filtering module 120 removes high-frequency noise such as electromyographic interference and power supply noise, reducing low-frequency drift and high-frequency noise components. Preferably, the low-pass filter is connected to the signal processing module 160 to send the secondary filtered EEG signal to the signal processing module 160 within the evaluation device 100.

[0084] After receiving the filtered EEG signal from the low-pass filter in the filtering module 120, the signal processing module 160 resamples the EEG signal.

[0085] The twice-filtered EEG signal is a time-domain signal, and the sampling rate at this time is 500 Hz. If resampling is not performed, the time series length is doubled. For a 10-minute EEG signal recording process, the number of channels of the electrode 210 is 128 leads, and the data volume will increase from about 1.08 GB (250 Hz) to 2.16 GB (500 Hz). The time-consuming of the spectral feature extraction of the first evaluation submodule 130 is in a square relationship with the data volume, and the calculation time will be significantly prolonged.

[0086] The features related to pain are mainly concentrated in the five frequency bands of δ, θ, α, β, and γ, and the 250 Hz sampling rate has met the Nyquist theorem requirement (supporting the highest 125 Hz signal). High sampling rate data contains more individual-specific noise details (such as tiny myoelectric tremors), and the second evaluation submodule 140 of the present application may incorrectly associate them as pain features. From the perspective of data transmission, the memory bandwidth (about 25.6 GB / s) of the evaluation device 100 is utilized up to 93% under the 500 Hz data flow, which is easy to cause a data transmission bottleneck, resulting in a real-time processing delay exceeding the clinical tolerance threshold (> 200 ms).

[0087] Therefore, the signal processing module 160 is arranged in the processor 110, which is used to resample the EEG signal, and reduce the original sampling rate 500 Hz to 250 Hz, that is, to compress the data volume while retaining the information of the pain-related frequency bands, and to achieve an optimal balance between accuracy and efficiency.

[0088] The signal processing module 160 includes an integrated analog-to-digital converter (ADC) and a digital signal preprocessing unit. After the continuous EEG signal is digitized by the analog-to-digital converter (ADC) at a sampling rate of 500 Hz, the digital signal preprocessing unit performs anti-aliasing filtering through a low-pass filter (cutoff frequency 125 Hz) to eliminate high-frequency aliasing interference. Subsequently, 250 Hz integer times down-sampling is achieved by decimation. This sampling rate can completely cover the δ, θ, α, β, and γ frequency bands related to pain analysis. The signal processing module 160 sends the resampled EEG signal to the frequency domain transformation module 170.

[0089] The downsampling operation reduces the data size by 50%, and simultaneously realizes three core optimizations: first, by filtering out non-physiological signals exceeding 125 Hz such as electromyographic interference harmonics, the signal-to-noise ratio of the target frequency band is improved; second, the standardized sampling rate eliminates differences in acquisition parameters across devices, enhancing the model's generalization ability; finally, high-frequency transient noise such as motion artifacts is suppressed, reducing interference with subsequent time-frequency analysis and functional connectivity modeling. This design achieves the best balance between computational efficiency and feature fidelity while preserving pain biomarkers such as beta / gamma oscillation phase synchrony.

[0090] After receiving the resampled EEG signal from the signal processing module 160, the frequency domain transformation module 170 obtains a time-frequency feature matrix through Fourier transform, as shown in Figure 2 The EEG signal received from the signal processing module 160 is a time-domain signal where B represents the batch size, N represents the number of electrodes, and T represents the time series length. The data form of the time-domain signal can be represented by a tensor with dimensions {48, 128, 38}. Here, 48 represents the frequency component, 128 represents the number of electrodes, and 38 represents the time series length.

[0091] The frequency domain transformation module 170 divides the time series length T into K overlapping time windows, each with a length L and a sliding step H. T = L + (K-1)·H.

[0092] After windowing, the signal dimension is

[0093] The frequency domain transformation module 170 applies a Hanning window to each time window to suppress spectral leakage:

[0094]

[0095] The windowed signal is:

[0096] The frequency domain transformation module 170 performs a Discrete Fourier Transform (DFT) on each time window, with the calculation formula being:

[0097]

[0098] The frequency domain transformation module 170 outputs a complex time-frequency matrix where D = [L / 2]+1 is the number of single-sided spectral components.

[0099] The frequency domain transformation module 170 calculates the logarithmic power spectrum of the complex time-frequency matrix, and takes the calculation result as the time-frequency feature matrix:

[0100] X TF [b, n, k, d] = 10·log10 (|X STFT [b,n,k,d]| 2 +ε)。

[0101] The frequency domain transformation module 170 finally outputs a real number time-frequency feature tensor, that is, a time-frequency feature matrix Wherein, ε = 1e-6.

[0102] Through the above generation process of the time-frequency feature matrix (windowing→windowing→STFT→log power spectrum), the application realizes the quantitative characterization of the dynamic time-frequency characteristics of the EEG signal. First, the δ (1-4 Hz), θ (4-9 Hz), α (9-12 Hz), β (12-25 Hz), and γ (25-100 Hz) frequency bands of the human EEG signal correspond to different neural activity patterns, respectively. For example, acute pain often causes the power of the β / γ frequency band to increase, while chronic pain is related to the continuous oscillation of the θ frequency band. Through frequency channel division, the time-frequency feature matrix can accurately separate the power density changes of these frequency bands. Secondly, pain perception has a time evolution characteristic, latency (low-frequency power rise)→peak period (high-frequency transient burst)→decline period (synchronization across frequency bands). The window dimension K of the time-frequency feature matrix corresponds to the time evolution axis (the step length covered by each window), so that the time-frequency feature can be positioned to a specific time window and frequency band, and by analyzing the power trajectory along the dimension K (such as the power integration of the γ frequency band in the k=50-80 window), the dynamic response mode of the pain stimulus can be quantified.

[0103] After the Fourier transform, the frequency domain transformation module 170 performs standardization processing on the time-frequency feature matrix.

[0104] The calculation formula of the standardization processing is:

[0105]

[0106] In the above formula, μ X represents the mean matrix of each dimension; σ X represents the standard deviation matrix.

[0107] In the above steps, by performing discrete Fourier transform, the following advantages are brought for the subsequent determination of the biomarker of the degree of pain:

[0108] First, the neural coding mechanism of pain is revealed. After the EEG signal is processed by the discrete Fourier transform, the frequency band energy features closely related to pain can be extracted. These frequency band energy values are converted into a Mark objective pain index of 0-100 points through nonlinear mapping, so that the subjective feeling of “pain” has a standardized biomarker for the first time.

[0109] Second, the fusion analysis of multi-dimensional features is realized. The discrete Fourier transform output contains a three-dimensional feature matrix of 48 frequency components, 128 electrode channels, and 38 time windows, which can not only reflect the spatial specificity of brain regions such as the prefrontal cortex in pain regulation, but also capture the instantaneous gamma burst pattern of sudden pain. The integration of time-space-frequency features lays the foundation for building a digital twin pain brain model, making it possible to distinguish subtypes such as neuropathic pain and inflammatory pain.

[0110] Third, the precision of clinical decision-making is promoted. Through the discrete Fourier transform algorithm, the evaluation device 100 can update the pain evaluation results every 2-5 seconds, realizing dynamic monitoring of postoperative pain. Clinical studies have shown that a 30% decrease in gamma power before and after treatment is significantly related to analgesic effect (R 2 = 0.922, see Figure 8 ), which greatly reduces the risk of misjudgment caused by differences in subject expression in traditional methods.

[0111] The frequency domain transformation module 170 sends the standardized time-frequency feature matrix to the frequency band separation module 180, as shown in Figure 2 After receiving the time-frequency feature matrix from the frequency domain transformation module 170, the frequency band separation module 180 divides the time-frequency feature matrix representing the EEG signal into five frequency bands related to pain perception, namely δ, θ, α, β, and γ, in the frequency domain.

[0112] Specifically, the frequency band separation module 180 uses a parallel FIR filter bank to divide the standardized time-frequency feature matrix into corresponding δ (0-4 Hz), θ (4-9 Hz), α (9-12 Hz), β (12-25 Hz), and γ (25-100 Hz) frequency bands according to the energy distribution of different frequency intervals. For example, if the power spectral density estimate value shows energy concentration in a certain frequency interval, and the interval is within the β frequency band range, then this part of the signal will be divided into the β frequency band.

[0113] Preferably, the frequency band separation module 180 extracts the time-frequency energy distribution features in the time-frequency feature matrix of each frequency band through a learnable frequency domain filter, and dynamically allocates weights to the time-frequency energy distribution features extracted from the time-frequency feature matrix of each frequency band based on a cross-frequency attention mechanism, fuses the energy and phase coupling characteristics between frequency bands, and generates a time-frequency feature matrix with frequency band specificity.

[0114] The frequency band separation module 180 significantly improves the dynamic representation ability of the time-frequency energy distribution features through a learnable frequency domain filter and a cross-frequency band attention mechanism. The innovation lies in breaking through the limitations of traditional fixed frequency band division, accurately capturing the differences in neural oscillation patterns in different frequency bands through adaptive learning of frequency domain filter parameters, and dynamically integrating the energy and phase coupling characteristics across frequency bands using the attention mechanism. This multi-frequency band collaborative modeling technique first realizes the nonlinear decoupling of the frequency domain features of pain-related neural activity, providing an interpretable quantitative basis for revealing the frequency-specificity of the physiological mechanism of pain.

[0115] The frequency band separation module 180 sends the time-frequency feature matrix of the five divided frequency bands (delta, theta, alpha, beta, and gamma) to the evaluation module 190, as shown in Figure 2 In the case of a communication connection between the processor 110 and the database, the time-frequency feature matrix of each frequency band can be stored in the database. Preferably, the database can also be a database of a digital twin brain connected with a third party.

[0116] From a neurology perspective, the physiological basis of the existing three-part method (theta: 4-8 Hz, alpha: 8-13 Hz, beta: 13-30 Hz) is rooted in classical research on neural oscillation mechanisms: alpha waves are related to resting state cortical synchronization, and beta waves are clearly associated with motor cortex activation and cognitive load. This division method has formed a paradigm in neuroelectrophysiological research, and its advantage is that the frequency band boundaries correspond to interpretable changes in neural activity patterns such as synaptic transmission efficiency and ion channel dynamics. The core difficulty in switching to five-part division lies in the continuity of neural oscillation mechanisms. For example, pain-induced gamma band (30-100 Hz) activity is related to local neural cluster synchronization, but its energy decays quickly and is easily disturbed by electromyographic artifacts, making it difficult to effectively extract using traditional Fourier transforms with fixed window lengths. However, the frequency domain transformation module 170 of the present invention improves the capture ability of high-frequency neural oscillations in three dimensions through the combined use of short-time Fourier transform (STFT) and Hanning window: first, the time-frequency resolution is optimized, with window length L and step size H set to improve the time resolution of the gamma band to milliseconds (theoretical minimum time unit ΔT = 1 / fs), while maintaining sufficient frequency resolution (Δf = 1 / L). This parameter configuration can effectively separate the gamma sub-band (such as low gamma 25-50 Hz and high gamma 50-100 Hz) in the 25-100 Hz range, overcoming the problem of high-frequency transient feature ambiguity caused by traditional FFT global spectral analysis; second, the artifact suppression is enhanced, with the Hanning window function significantly reducing the spectral sidelobe attenuation by tens of dB compared to the rectangular window, significantly reducing the harmonic interference of electromyographic artifacts (mainly distributed in 60-200 Hz) on the gamma band; finally, the dynamic features are preserved, with the logarithmic power spectrum conversion (x TFThe nonlinear compression and expansion of the low-amplitude γ oscillation is performed by the nonlinear compression and expansion function [b, n, k, d]. When the γ power produces a short-time burst of tens of ms magnitude under the pain stimulus, the transformation improves the weak signal detection sensitivity by more than 20%, effectively capturing the γ band phase amplitude coupling (PAC) characteristics between the prefrontal cortex and the anterior cingulate gyrus.

[0117] As shown in Figure 1 and Figure 2 After receiving the divided five frequency band (δ, θ, α, β and γ) time-frequency feature matrices from the frequency band separation module 180, the evaluation module 190 performs the calculation of the pain objective quantification index.

[0118] The evaluation module 190 includes a first evaluation submodule 130 and a second evaluation submodule 140. The first evaluation submodule 130 is composed of at least one graph generation network, which is used to generate a first feature vector representing the global correlation pattern between the electrodes 210 used to collect the EEG signal. The second evaluation submodule 140 is composed of at least one convolutional neural network, which is used to generate the pain objective quantification index corresponding to the EEG signal according to the first feature vector and a second feature vector generated by itself.

[0119] The first evaluation submodule 130 is provided with a normalization unit 131, a bidirectional gated recurrent unit 132 (BiGRU), a Gumbel-Sampler unit 133 and a graph aggregation unit 134. The normalization unit 131 receives the original pain rating label from the signal receiving module 111 and performs normalization processing on the original pain rating label. The bidirectional gated recurrent unit 132 (BiGRU) is used to extract the time sequence features of the EEG signal from the normalized time-frequency feature matrix. The Gumbel-Sampler unit 133 is used to perform the Gumbel-Softmax method. The graph aggregation unit 134 is used to generate the first feature vector representing the global correlation pattern between the electrodes 210 used to collect the EEG signal.

[0120] Specifically, the normalization unit 131 performs normalization processing on the received original pain rating label.

[0121] The range of the original pain rating label is y reg ∈ [0, 10].

[0122] The specific formula for the normalization step performed by the normalization unit 131 is:

[0123] The range of the pain rating label after the normalization of the original pain rating label by the normalization unit 131 is y norm ∈ [0, 1].

[0124] y min= 0, indicating the minimum value of the pain score; y max = 10, indicating the maximum value of the pain score.

[0125] In the regression task, the normalization unit 131 normalizes the original pain score label to the interval [0, 1] through a linear mapping. The linear normalization mapping of the original pain score label is consistent with the multidimensional neural coding mechanism of pain. The normalized pain score label is sent by the normalization unit 131 to the Bayesian update module 150 for loss calculation.

[0126] Preferably, the standardized time-frequency feature matrix After that, the bidirectional gated recurrent unit 132 (BiGRU) processes and outputs the time series feature: where B represents the batch size, N represents the number of EEG electrodes, that is, the number of channels, and T represents the length of the time series; D' represents the number of GRU hidden layer neurons. The dimension T is absorbed into the input dimension during the processing of the bidirectional gated recurrent unit 132 and does not appear in the output information. The bidirectional gated recurrent unit 132 sends the time series feature to the convolution unit 141 in the second evaluation sub-module 140.

[0127] The time series modeling mechanism of the bidirectional gated recurrent unit 132 (BiGRU) establishes a multiscale mapping relationship between the time series feature and the pain intensity through the synergistic action of the reset gate and the update gate. That is, the bidirectional gated recurrent unit 132 (BiGRU) extracts the end time step feature and the adjacency matrix from the time series feature H BiGRU .

[0128] To this end, the bidirectional gated recurrent unit 132 (BiGRU) extracts the end time step (the last time step) feature from the time series feature H BiGRU . Since the pain-related neural response has a time course characteristic, the end time step feature integrates the full-phase change of the signal from the start to the end of the stimulus.

[0129] Preferably, the hidden layer dimension of the bidirectional gated recurrent unit 132 (BiGRU) is configured as D' = 256. The bidirectional gated recurrent unit 132 (BiGRU) sequentially processes the input along the forward and reverse time windows respectively, concatenates the hidden states of each time window into 512 dimensions (2D' = 2x256), and outputs the complete time series feature: The end time step feature is:

[0130] The bidirectional gated recurrent unit 132 (BiGRU) sends the end time step feature to the Gumbel-Sampler unit 133, as shown in Figure 2 After receiving the end time step feature from the bidirectional gated recurrent unit 132 (BiGRU), the Gumbel-Sampler unit 133 can capture the cumulative dynamic pattern of the signal sequence. The neural response (such as gamma band oscillation) triggered by pain stimulation usually has a time course cumulative characteristic. By intercepting the end time step feature The first evaluation sub-module 130 integrates the dynamic changes of the signal full time window. For example, the gamma power time course integral of the postoperative pain patient is significantly higher than the instantaneous value. This cumulative feature is more in line with the long-term potentiation (LTP) mechanism of nociceptive central sensitization.

[0131] The Gumbel-Sampler unit 133 converts the end time step feature into a connection graph between the electrodes 210 based on the Gumbel-Softmax method, thereby forming an adjacency matrix representing the degree of neural activity coupling between the brain cortical regions. The formula of the connection graph (adjacency matrix) is: The adjacency matrix element A i，j of the adjacency matrix A represents the association strength between electrodes i and j, reflecting the functional connection pattern between brain regions.

[0132] Specifically, the adjacency matrix between the electrodes 210 is a square matrix, and each element represents the possibility of forming a connection between the corresponding electrode pair. Assuming that there is a network composed of multiple electrodes 210, there is a probability of a connection between each pair of electrodes 210, that is, the association strength, and the adjacency matrix element A i，j of the adjacency matrix represents the association strength between electrodes i and j. The diagonal element A i，i is usually 0 because the electrode 210 will not be connected to itself. For the adjacency matrix A, if the association strength (connection probability) is symmetric, then A i，j = A i,i .

[0133] Assuming there are 3 electrodes, the adjacency matrix A can be represented as:

[0134] where the association strength between the first electrode and the second electrode is 0.5, the association strength between the first electrode and the third electrode is 0.2, and the association strength between the second electrode and the third electrode is 0.3. Through the adjacency matrix A, the connection relationship between the electrodes 210 or nodes can be quantified and analyzed.

[0135] In the case of 128 channels of electrodes 210, let the electrode set ​The set of spatial coordinates is The adjacency matrix A is then The formula can be expressed as:

[0136]

[0137] The spatial decay term is: d ij = ||(x i , y i , z i ) - (x j , y j , z j )||2, representing the Euclidean distance between electrode pairs; σ controls the spatial correlation decay rate (typical values: 10-30 mm).

[0138] The functional coupling term is:

[0139] s i (t) represents the signal of electrode i in the time window [t-T, t]; T is determined by the dominant frequency of neural oscillations (e.g., T = 100 ms for the alpha band).

[0140] Preferably, the adjacency matrix A has symmetry, A i，j = A j，i .

[0141] Preferably, the adjacency matrix A implements a non-linear normalization: denotes compression to the interval [0, 1].

[0142] Preferably, the adjacency matrix A implements sparsity control: τ is a threshold value, for example τ = 0.3.

[0143] For example, when electrode i is located in Brodmann area 24 and electrode j is located in Brodmann area 32:

[0144] The spatial distance d ij = 15 mm; the functional correlation p ij = 0.62 (pain task-state data); σ = 20 mm, then:

[0145]

[0146] As shown in the data extraction part of several electrodes 210 as shown in Figure 4 and Figure 6 , in real tests, the adjacency matrix A obtained between different electrodes 210 is a multi-digit decimal number. It should be noted that Figure 4 and Figure 6 the adjacency matrix A shown is not symmetric, because Figure 4 andFigure 6 The data in the matrix have temporal causal relationship, i.e., the data have direction. The association strength from electrode E1 to electrode E2 is different from the association strength from electrode E2 to electrode E2.

[0147] The Gumbel-Sampler unit 133 in the first evaluation submodule 130 generates the adjacency matrix element A i , i∈[0, 1] quantifies the functional connection strength between electrode i and j, and its physical meaning corresponds to the coupling degree of neural activity between the brain cortex regions. For example, in the motor imagination task, A C3，C4 (corresponding to the left and right motor cortex) may have a significantly high value, reflecting the cross-brain coordination mechanism. This method balances “exploration” (random connection) and “exploitation” (determined connection) through a trainable parameter τ, and can more flexibly model dynamic functional connections than the traditional fixed threshold method.

[0148] Compared with the prior art which can only evaluate the pain position and cannot calculate the pain degree, the present application uses the Gumbel-Softmax method to convert the end time step features of the electroencephalogram signal into a differentiable adjacency matrix, and quantifies the dynamic functional connection strength between the brain cortex regions. This dynamic adjacency matrix has three technical advantages:

[0149] Firstly, the time sequence features extracted by the bidirectional gated recurrent unit 132 (BiGRU) can capture the reorganization process of the pain-related brain network, such as the transient enhancement of the prefrontal-limbic system connection under pain stimulation; secondly, the self-adaptive generated adjacency matrix breaks through the limitation of the fixed brain network template, and can accurately represent the spatio-temporal evolution law of individualized brain functional connection, and the generated connection graph shows that the reverse coupling strength between the central executive network and the default mode network under the pain state is improved by 3.2 times; finally, the local spatio-temporal features extracted by the Gumbel-Sampler unit 133 and the global connection features generated by the graph network are fused to form a fusion vector containing multi-scale pain features, so that the recognition accuracy of the evaluation module 190 for neuropathic pain is improved.

[0150] Preferably, the Gumbel-Sampler unit 133 constructs a differentiable graph structure based on the functional connection relationship between the electrodes 210 (quantified by the existence probability of the edge of the adjacency matrix). The Gumbel-Sampler unit 133 generates the adjacency matrix element A i，jGumbel noise is added and the decision of the reserved state of the edge is made by a differentiable sampling mechanism. By performing an argmax operation on the connection probability with added noise, it is determined whether to reserve the edge. The advantage of the Gumbel-Sampler unit 133 in sampling the adjacency matrix A is that it allows the gradient backpropagation of discrete choices (existence or nonexistence of edges), making the entire process differentiable; and it introduces randomness, which can better capture the uncertainty and dynamics of connections.

[0151] The Gumbel-Sampler unit 133 repeatedly samples and records the changes in the edges of the graph structure, thereby modeling the dynamic changes in connectivity over time, i.e., modeling the connectivity as a random variable that changes over time. The random variable that changes over time reflects the dynamic changes in functional connectivity of brain regions between pain onset and non-onset. The differential graph structure converts discrete brain network connections into analytically optimized pain biomarkers through mathematical continuity. The gradient information of the differential graph structure not only quantifies the activation degree of the pain pathway, but also reveals the neural mechanism of therapeutic intervention, providing a computational target for precise analgesia.

[0152] During the onset of pain, certain brain regions may exhibit stronger interactions, leading to increased connectivity. During the non-onset of pain, connectivity decreases. The Gumbel-Sampler unit 133 models the connectivity as a random variable that changes over time, and by repeatedly sampling to record changes in graph connections, it can accurately capture the dynamic changes in functional connectivity of brain regions. This can intuitively observe the transition of pain state from the perspective of brain functional connectivity, providing a brain physiological basis for the determination of objective pain index. At the same time, changes in pain state will cause the reorganization of brain functional networks. The Gumbel-Sampler unit 133 models the randomness and dynamics of connectivity, which can reflect this reorganization process of brain functional networks in the pain onset state. This reorganization process is closely related to the generation and development of pain, and by monitoring and analyzing it, a more in-depth physiological mechanism explanation can be provided for the determination of objective pain index.

[0153] Preferably, the Gumbel-Sampler unit 133 can also set a temperature parameter to control the "softness" of sampling. The temperature parameter τ can be used to adjust the scaling of the correlation strength A ij , simulating the randomness and dynamic changes of brain neural activity, thereby adapting to the differences in functional connectivity of different individual brains in the pain state. For example, some patients may exhibit stronger or weaker correlation strength due to different neural sensitivities.

[0154] Suppose there is a graph consisting of N nodes (electrodes 210), the Gumbel-Sampler unit 133 calculates the correlation strength Aij .

[0155] Add noise: Add noise to the association strength A ij Add Gumbel noise g ij .

[0156]

[0157] The addition of noise simulates the randomness and uncertainty of brain neural activity, making the Gumbel-Sampler unit 133 closer to the actual complex working mechanism of the brain, and avoiding ignoring the dynamic changes of neural activity due to too deterministic probability.

[0158] Temperature adjustment: The Gumbel-Sampler unit 133 selects appropriate temperature parameters τ to adjust the association strength according to the current temperature.

[0159]

[0160] If the influence of the temperature parameter is ignored, the accuracy of the objective quantification index of pain will be reduced. For example, although the added Gumbel noise simulates uncertainty, if the temperature parameter τ is not adjusted, the influence of the noise cannot be dynamically adjusted. The neural signals in some pain scenarios may be over-amplified or suppressed by noise, thereby affecting the accurate quantification of pain intensity. If the temperature parameter is fixed, the Gumbel-Sampler unit 133 will process all inputs with a fixed scale, and cannot capture the real-time changes of functional connections during the pain process in the neural activity patterns when the pain is intensified or alleviated. In complex pain scenarios (such as mixed pain types or dynamically changing pain intensity), the Gumbel-Sampler unit 133 may not be able to capture diverse functional connection patterns by adjusting the temperature parameter τ, thereby reducing the reliability of the quantification results.

[0161] The Gumbel-Sampler unit 133 of the present application adjusts the temperature to adapt to the diversity and complexity of the functional connection changes of different individual brains in the pain state. This adjustment enables the Gumbel-Sampler unit 133 to better capture the characteristics in different pain scenarios. τ represents the temperature parameter. Preferably, the temperature parameter τ is initially set to 1.0 and gradually reduced to 0.1 in the training process according to an exponential decay strategy, to balance exploration and utilization.

[0162] The Gumbel-Sampler unit 133 determines whether there is a connection between each pair of nodes through the argmax operation.

[0163]

[0164] When τ is small, The maximum value in the sampling result is more prominent, and the sampling result is closer to the hard decision of 0 or 1; when τ is larger, more smooth, and the sampling result is more fuzzy.

[0165] In the sampling decision process, according to the adjusted connection probability, it is determined whether there is an actual functional connection between different brain regions of the brain. The decision mode at different temperatures can adapt to the clear or fuzzy stage of the change of the pain state, and accurately reflect the real-time state of the brain functional network.

[0166] The Gumbel-Sampler unit 133 records the differential graph structure obtained by the current sampling, and repeats the sampling.

[0167] The Gumbel-Sampler unit 133 repeatedly performs the above steps to generate multiple different graph structures, reflecting the dynamic changes of connectivity in the pain state. In the case of low temperature (small τ), the sampling result is closer to the hard decision, which can clearly distinguish the changes of connection between the onset and non-onset of pain. In the case of high temperature (large τ), the sampling result is more fuzzy, which can better capture the uncertainty and dynamics of the connection.

[0168] Low temperature will make the sampling closer to the hard decision. In the modeling of brain functional connection graph structure, this means that the Gumbel-Sampler unit 133 can more clearly determine the existence or non-existence of the connection between brain regions. In the pain state, the enhancement of the connection between certain brain regions may be a clear signal of the onset of pain. Low-temperature sampling enables the Gumbel-Sampler unit 133 to accurately capture such clear changes in connection, which can be used as an important basis for judging the onset and degree of pain. For example, in chronic pain patients, when the connection between certain specific brain regions is clearly determined to be enhanced under low-temperature sampling, it may correspond to a significant increase in the degree of pain, thereby providing a clearer and more accurate reference for the objective pain index.

[0169] High temperature will make the decision more fuzzy. During the occurrence and relief of pain, there are some transition stages, at which time the reorganization of the brain functional network is not achieved at one stroke, and the changes of the connection between brain regions are complex and uncertain. High-temperature sampling can better adapt to this situation, allowing the model to handle this fuzzy information in a more flexible way. For example, in the early stage of the onset or relief of pain, the changes of the connection between brain regions may not be very obvious, and high-temperature sampling can consider more possibilities, avoiding the deviation of the assessment of the degree of pain due to too absolute judgment, so that the objective pain index can more accurately reflect the dynamic changes of the pain state.

[0170] The Gumbel-Sampler unit 133 sends the end-time step features and the adjacency matrix to the graph aggregation unit 134. After receiving the adjacency matrix from the Gumbel-Sampler unit 133, the graph aggregation unit 134 proceeds with the calculation of the first feature vector.

[0171] Preferably, the graph aggregation unit 134 generates the first feature vector representing the global association pattern between the electrodes 210 used to collect the EEG signals based on the adjacency matrix.

[0172] In the graph generation network of the first evaluation sub-module 130, the nodes constantly exchange information (messages) with their neighbor nodes and update their own feature representations according to the information. The graph aggregation unit 134 aggregates the connection information of the electrodes 210 (nodes) of the neighbor matrix through the message passing mechanism. The principle of the graph aggregation unit 134 aggregating the connection information of the electrodes 210 (nodes) is that each electrode 210 (node) aggregates the feature information of its neighbor nodes by weighting, and the weight is determined by the adjacency matrix A.

[0173] The formula for the graph aggregation unit 134 to aggregate the connection information of the electrodes 210 (nodes) is:

[0174]

[0175] This process aggregates the neighborhood information of each electrode 210 to generate a first feature vector Z GNN with a dimension of d GNN , representing the global association pattern between the electrodes 210, that is, the dynamic brain functional connectivity network. The global association pattern refers to the large-scale functional connectivity network across brain regions, describing the statistical properties of signal synchronization, causality or information transmission efficiency between distributed brain regions.

[0176] The graph aggregation unit 134 sends the first feature vector Z GNN to the convolution unit 141 of the second evaluation sub-module 140, as shown in Figure 2 .

[0177] The association strength of all electrodes 210 collectively constitutes the dynamic brain functional connectivity network, and its topological features (such as node degree, clustering coefficient, and betweenness centrality) can represent the pain-related brain region coordination pattern. For example, the high-weight connection between the electrode 210 and the electrode 210 corresponding to the anterior cingulate cortex (ACC) may reflect the neural coding of the emotional component of pain. In a chronic pain model, the connection strength between the electrode 210 and the electrode 210 of the sensorimotor cortex increases over time, which is consistent with the rising trend of clinical pain scores.

[0178] After receiving the time series features H BiGRUConvolution processing is performed to generate a second feature vector that includes the local spatiotemporal dynamic features of the EEG signal.

[0179] Specifically, the second evaluation submodule 140 will evaluate the time series features H BiGRU The input is fed into a 2D convolutional layer, where two-dimensional convolution processing is applied. The 2D convolutional layer performs joint feature learning along the spatial (electrode arrangement) and temporal dimensions to extract local spatiotemporal dynamic features. Local spatiotemporal dynamic features refer to the coordinated change patterns of neural signals within a limited spatial range (the brain region covered by adjacent electrodes 210) and a short time window, emphasizing the rapid, localized activity characteristics of neuronal clusters. Local spatiotemporal dynamic features reflect the synchronous oscillation or inhibition phenomena of neuronal populations within specific brain regions. For example, pain stimulation induces an increase in γ-oscillation power in the S1 region, which is positively correlated with the intensity of acute mechanical pain. Local spatiotemporal dynamic features can capture the transient characteristics of pain-related neural responses. For example, the peak latency of evoked potentials appearing within 170-300 ms after laser pain stimulation is negatively correlated with subjective pain intensity (the shorter the latency, the higher the pain score).

[0180] Preferably, the information received by the convolutional unit 141 contains temporal features and topological features. The temporal features are time series features H. BiGRU The time dimension T and the bidirectional hidden layer features 2D′ are given. The topological features are represented by the first feature vector Z. GNN The global association pattern between electrodes 210 in the middle.

[0181] The calculation formula for extracting local spatiotemporal dynamic features by convolutional unit 141 in the second evaluation submodule 140 is as follows:

[0182]

[0183] By using multi-scale convolutional kernel groups, spatiotemporal coupling features of different ranges are effectively captured, preserving the local time-varying characteristics of the signal, with an output dimension of d. GNN The second eigenvector Z CNN .

[0184] like Figure 2 As shown, after receiving the first feature vector from the graph aggregation unit 134, the convolution unit 141 in the second evaluation submodule 140 sends the first feature vector and the second feature vector along the feature dimension to the splicing unit 142 for splicing, so as to form a fused feature vector including the global correlation pattern and local spatiotemporal dynamic features between the electrodes 210.

[0185] After receiving the first feature vector and the second feature vector from the convolution unit 141, the concatenation unit 142 concatenates the first feature vector and the second feature vector along the feature dimension. The concatenation formula is as follows: Among them, Z concat Z represents the fused feature vector.GNN denotes the first feature vector; Z CNN denotes the second feature vector.

[0186] The fusion feature vector Z concat contains both global correlation patterns between electrodes 210 and local spatiotemporal dynamic features. The concatenation unit 142 sends the fusion feature vector Z concat to the classification unit 143 in the second evaluation submodule 140.

[0187] The fusion feature vector Z concat is received by the classification unit 143 from the concatenation unit 142. The classification unit 143 then outputs a normalized objective quantification of pain based on a Softmax function.

[0188] Preferably, the classification unit 143 computes a class probability distribution of the fusion feature vector and normalizes it to generate an objective quantification of pain corresponding to the EEG signal. The step of computing a class probability distribution of the fusion feature vector by the classification unit 143 in the second evaluation submodule 140 includes:

[0189] The classification unit 143 computes a class probability distribution of the fusion feature vector Z concat using a Softmax function.

[0190]

[0191] denotes the predicted output of the Softmax function, and denotes the pain class probability distribution. B denotes the batch size, i.e., the number of samples input in one training / inference. C denotes the number of pain classes (e.g., healthy, non-neurological, and neurological). W class denotes the weight matrix of the fully connected layer. W class has a dimension of C x (d GNN +d CNN ) denotes the linear transformation parameter that maps the fusion feature to the class space. b class denotes the bias term of the fully connected layer.

[0192] The class probability distribution of the fusion feature vector Z concat reflects the differences in the combination weights of the neural features under different pain levels. The probability distribution encodes both local and global features in the perception of pain. The local features characterize the pain sensory input. The global features characterize the pain emotion and cognitive regulation. The probability value of the class probability distribution is directly related to the pain level.

[0193] The formula for the classification unit 143 to output a normalized objective quantification of pain based on a Softmax function is:

[0194]

[0195] represents the normalized pain objective quantification index, representing the continuous prediction value of the pain intensity by the second evaluation sub-module 140. The output value of the pain objective quantification index ranges between [0, 1], which can be linearly scaled to the actual pain score, i.e. the Marke value. The Marke value ranges between [0, 100].

[0196]

[0197] For example, Then the corresponding pain objective quantification index is 75 Marke. Then the corresponding pain objective quantification index is 20 Marke.

[0198] Preferably, different values of the Marke value output by the evaluation module 190 of the present application represent different degrees of pain.

[0199] If the pain intensity is mapped to 100 Marke, it means the most severe pain, and the EEG signal shows significant changes.

[0200] If the pain intensity is mapped to 80 Marke, it means severe pain, and the EEG signal shows obvious changes.

[0201] If the pain intensity is mapped to 50 Marke, it means moderate pain, and the EEG signal shows moderate changes.

[0202] If the pain intensity is mapped to 30 Marke, it means mild pain, and the EEG signal shows small changes.

[0203] If the pain intensity is mapped to 0 Marke, it means no pain, and the EEG signal shows no changes.

[0204] Different scores reflect different degrees of pain and are closely related to the changes of the EEG signal in the corresponding frequency band. Table 1 shows examples of specific changes of the EEG signal under different scores.

[0205] Table 1:

[0206]

[0207] From the above table, it can be more intuitively seen that the specific changes of the EEG signal in each frequency band under different pain intensities. These changes can be quantified and analyzed by the pain objective quantification index for further pain assessment.

[0208] Figure 3 The electroencephalogram of the first subject is shown. Figure 4Extraction data of the first subject is shown. After the calculation of the graph network model, part of the data of the probability matrix of the connection between electrodes 210 is shown as Figure 4 The pain score of the first subject is: 0.9823 Marke.

[0209] Figure 5 Electroencephalogram of the second subject is shown. Figure 6 Extraction data of the second subject is shown. After the calculation of the graph network model, part of the data of the probability matrix of the connection between electrodes 210 is shown as Figure 6 The pain score of the second subject is: 1.91E-8 Marke.

[0210] After receiving the normalized pain score label from the normalization unit 131, the Bayesian update module 150 constructs a variational inference loss function to update the weight parameters of the bidirectional gated recurrent unit 132 through the reparameterization gradient algorithm.

[0211] The classification unit 143 sends the pain objective quantification index to the Bayesian update module 150, as shown in Figure 2 After receiving the pain objective quantification index (Marke value) from the classification unit 143, the Bayesian update module 150 calculates the 95% confidence interval of the Marke value according to the model cognitive uncertainty (parameter posterior distribution variance) and the observation noise level. If the confidence interval width exceeds the preset threshold, the Bayesian update module 150 dynamically adjusts the observation noise parameter to optimize the feature weight distribution. If the Marke value deviates from the normalized pain score label by more than 3 times the total standard deviation for 3 consecutive times, the parameter retraining process is triggered to suppress device noise and physiological artifact interference.

[0212] The Bayesian update module 150 introduces individualized physiological baseline data (such as resting-state EEG features) to establish a personalized prior distribution. The output of the Bayesian update module 150 is the calibrated Marke value and its confidence interval, i.e., the posterior distribution.

[0213] The Bayesian update module 150 establishes an individual-specific prior distribution based on the first 10 measurements of the Marke value. The Bayesian update module 150 calculates the posterior distribution through Markov Chain Monte Carlo (MCMC) sampling. The Bayesian update module 150 performs a parameter retraining process every 24 hours to avoid concept drift.

[0214] After receiving the pain quantification index (Marke value) from the classification unit 143, the Bayesian update module 150 performs the following operations.

[0215] Calculate the 95% confidence interval:

[0216] If the confidence interval width exceeds the threshold, adjust the observation noise parameter according to The noise parameter is dynamically adjusted. If the Marke value deviates from the normalized label by more than 3σ for 3 consecutive times, the parameter retraining is triggered. The output of the Bayesian update module 150 is the calibrated Marke value and its confidence interval (posterior distribution).

[0217] In the pain classification task, the training data may have an uneven distribution of pain categories (e.g., there are much fewer severe pain samples than mild pain samples), causing the model to be biased towards the majority class. The weighted cross-entropy loss function assigns different weights to different pain categories, balancing the attention of the evaluation module 190 to each pain category.

[0218] The Bayesian update module 150 uses a weighted cross-entropy loss function to solve the class imbalance problem.

[0219]

[0220] where, represents the weighted cross-entropy loss function of the classification task, represents the weight of the i-th sample true pain category (normalized pain rating label) y i . represents the evaluation module 190's predicted score for the i-th sample in pain category C (unnormalized original pain rating label). Preferably, C = 5, indicating 5 different levels of pain. represents the evaluation module 190's predicted score for the i-th sample in its true pain category y i .

[0221] If the weight of a certain level of pain in the pain sample is higher, the Bayesian update module 150 will enhance the sensitivity to that level of pain (e.g., ACC-insular connection strength abnormality), avoiding misdiagnosis of severe cases.

[0222] The Bayesian update module 150 uses a weighted cross-entropy loss function to dynamically adjust the pain category weight, which can alleviate the prediction bias caused by the imbalance of pain categories, improve the recognition ability of the evaluation module 190 to the minority pain categories (e.g., severe pain), avoid misdiagnosis, and also enhance the sensitivity of the evaluation module 190 to key biomarkers (e.g., high-frequency oscillation signals).

[0223] The Bayesian update module 150 uses the Huber loss function to implement the regression prediction task of pain intensity.

[0224]

[0225] represents the Huber loss function of the regression task, used to quantify the prediction error of pain intensity. y iyi represents the normalized pain score label of the i-th sample by the evaluation module 190. yi represents the predicted pain intensity value of the i-th sample by the evaluation module 190. δ represents a threshold parameter, controlling the switching threshold of the loss function from quadratic form to linear form. For example, δ = 1. The value of δ can be adjusted according to the data distribution.

[0226] The Huber loss function significantly improves the robustness of the Bayesian update module 150 in the pain intensity prediction task through the following mechanism. The Huber loss function combines the characteristics of mean square error (MSE) and mean absolute error (MAE), effectively resisting outliers. When the absolute value of the prediction error is less than the preset threshold δ, the MSE (quadratic function form) is adopted to maintain the smoothness of the gradient; when the error exceeds the threshold δ, it is switched to MAE (linear function form), which can suppress the gradient contribution of abnormal values.

[0227] The value of the Huber loss function presents a continuous and smooth transition characteristic in the gradient when the error approaches the threshold δ, solving the problem of non-differentiability of MAE at zero.

[0228] Preferably, for a population of individuals with significant differences in pain perception (such as chronic pain patients), the value of δ can be appropriately increased (such as from 1.0 to 1.5) to accommodate the natural volatility of their pain scores.

[0229] The Bayesian update module 150 uses the Adam optimization function to update the model parameters, with an initial learning rate of η = 1 × 10 -3 and introduces the following strategies to improve generalization ability. Here, the model parameters include trainable weight matrices and bias terms.

[0230] The formula for adjusting the dynamic learning rate is:

[0231]

[0232] where η t represents the decay learning rate, η0represents the initial learning rate (default 1 × 10 -3 ), γ represents the decay coefficient (default 0.95), and t represents the current Epoch number.

[0233] This formula indicates that in the initial stable stage (the first 20 Epochs), the initial learning rate η0= 1 × 10 -3, which allows the Bayesian update module 150 to quickly converge to a reasonable parameter space region. In the later fine-tuning stage (Epoch > 20), the learning rate is exponentially decayed, gradually reducing the parameter update step, so that the Bayesian update module 150 can more accurately approach the optimal solution. By adjusting the dynamic learning rate, the balance between training speed and stability is achieved, avoiding the Bayesian update module 150 from falling into a local optimal solution.

[0234] The formula for calculating the L2 weight decay of the Bayesian update module 150 is:

[0235]

[0236] represents the total loss function, represents the Huber loss of the regression task, λ = 1 × 10 -4 , which represents the decay coefficient. represents the L2 norm square of the model trainable parameters (including weight matrix and bias term).

[0237] The present application evaluates the process of identifying and classifying pain, i.e., the process of evaluating the objective quantification index of pain (Marke value) output by the evaluation module 190.

[0238] First, the category probability distribution output by the first evaluation submodule 130 is (C is the number of pain categories), and the predicted label is obtained by taking the maximum value index:

[0239]

[0240] represents the predicted label vector (length B), and the element value is an integer from 0 to C-1, indicating the predicted category of each sample. The index position indicates the predicted pain category label number. Each index corresponds to a predefined pain category (such as 0-no pain, 1-mild, 2-moderate, and 3-severe). By finding the position of the maximum value (argmax) in the probability vector, the pain category to which the current sample is most likely to belong can be determined. The predicted labels are, for example, neuropathic pain, inflammatory pain, and psychogenic pain. Each predicted label corresponds to a different range of Marke values.

[0241] Through this step, the pain intensity category and level are obtained. For example, mild pain (30-50 Marke), the energy in the beta band (12-25 Hz) is locally enhanced in the somatosensory cortex, but does not break through the filtering threshold of the thalamic gate, showing intermittent synchronous pulses.

[0242] The present application evaluates the objective quantification index of pain (Marke value) through the confusion matrix.

[0243] Confusion Matrix: used to count the distribution of true labels and predicted labels, formula as follows:

[0244]

[0245] m ii m represents the number of pain samples with true class i that are correctly classified. ij m represents the number of samples with true class i that are misclassified as class j.

[0246] In this step, the classification results of all samples are counted, and a matrix is established according to the combination of true class and predicted class. Each element cell in the matrix records the number of samples with true class i and predicted class j. The calculation output of this confusion matrix is the correct classification proportion of all pain samples. The calculation result of the confusion matrix is CxC. C represents the pain class.

[0247] As Figure 7 shown in the confusion matrix classification result diagram of the present application is used to detect the classification performance of the evaluation module 190. Figure 7 The classes of the classification in the above table are healthy, non-neurological, and neurological. The horizontal is the true label, and the vertical is the predicted label. The number in each cell represents the matching of the model prediction result and the true result, for example: the number in the middle cell is 0.99, indicating that the proportion of samples with true label as non-neurological and predicted label as non-neurological is very high. The proportion of samples with true label as neurological and predicted label as non-neurological is 0. The gray bar represents the proportion value, black corresponds to a proportion of v, and white corresponds to a proportion of 0. Overall, the evaluation module 190 performs better in distinguishing non-neurological samples.

[0248] Through the calculation of the confusion matrix, it can be confirmed whether the evaluation module 190 of the present application meets the clinical requirements in distinguishing the pain types (such as neurological, inflammatory, and psychogenic) and the pain degree (Marke value). For example, if the Marke value in the “moderate range” is often calculated as the Marke value in the “severe range”, it may indicate that the evaluation module 190 is insufficient in capturing the characteristics of moderate pain (such as the beta wave threshold of a specific brain region).

[0249] Preferably, the F1 score is calculated.

[0250] TP / FP / FN in the confusion matrix is calculated.

[0251] TP (True Positive) represents the number of samples that are actually positive (such as “pain exists”) and are correctly predicted as positive by the model.

[0252] FP (False Positive) represents the number of samples that are actually negative (e.g., "no pain") but are incorrectly predicted as positive by the model.

[0253] FN (False Negative) represents the number of samples that are actually positive but are incorrectly predicted as negative by the model.

[0254] F1 score is used to evaluate the classification accuracy (Precision) and coverage (Recall) of the module comprehensively, avoiding the bias caused by class imbalance when using accuracy alone.

[0255] According to the calculation granularity, it can be divided into:

[0256] Micro-F1: TP / FP / FN in the global statistical confusion matrix, ignoring class imbalance:

[0257]

[0258] F1 micro Micro-F1 is the micro-average F1, which represents the total number of correct and incorrect samples of all classes. TP c represents the number of samples that are actually c class and predicted as c class; FP c represents the number of samples that are actually non-c class but predicted as c class; FN c represents the number of samples that are actually c class but predicted as non-c class.

[0259] Micro-F1 can be calculated by calculating TP (true positive), FP (false positive), and FN (false negative) for all samples globally, ignoring class imbalance.

[0260] Macro-F1: Unweighted average of F1 for each class.

[0261]

[0262] F1 macro Macro-F1 is the macro-average F1, which represents the arithmetic mean value obtained after calculating the F1 value of each class independently.

[0263] Weighted-F1: Weighted average according to the number of samples in each class.

[0264]

[0265] F1 weighted Weighted-F1 is the weighted F1, which represents the final weighted average according to the weight assigned according to the proportion of the number of samples in each class. N c N represents the number of true samples in class c.

[0266] The data obtained by this step is a value between 0 and 1, the closer to 1, the better the model classification performance. The present application judges the ability of the evaluation module 190 to stably capture pain-related cross-brain region features (such as the association between prefrontal regulation signals and emotional responses of the limbic system) by calculating the F1 score.

[0267] Calculate the classification accuracy.

[0268] Get the total number of all correct classifications in the confusion matrix.

[0269] Classification accuracy (Accuracy): the proportion of predictions that are correct among all samples.

[0270]

[0271] Accuracy represents the classification accuracy, m cc represents the diagonal elements of the confusion matrix (the number of samples correctly classified).

[0272] As Figure 7 shown in the confusion matrix classification result diagram of the present application is used to detect the classification performance of the evaluation module 190. Figure 7 The categories of the classification in the confusion matrix are healthy, non-neurological, and neurological. The horizontal is the true label, and the vertical is the predicted label. The number in each square represents the matching of the model prediction result and the true result, for example: the middle square number is 0.99, indicating that the proportion of samples with true label non-neurological and predicted label non-neurological is very high. The proportion of samples with true label neurological and predicted label non-neurological is 0. The gray bar represents the proportion value, black corresponds to a proportion of 1, and white corresponds to a proportion of 0. Overall, the evaluation module 190 performs better in distinguishing non-neurological samples.

[0273] Through the calculation of the confusion matrix, it can be confirmed that the ability of the evaluation module 190 of the present application to distinguish between pain types (such as neurological, inflammatory, and psychogenic) and degrees (mild, moderate, and severe) meets the clinical requirements. For example, if the "moderate" F is often misjudged as "severe", it may indicate that the evaluation module 190 is insufficient in capturing the characteristics of moderate pain (such as the beta wave threshold of a specific brain region).

[0274] The scoring process of the pain objective quantification index of the present application is as follows.

[0275] The present application reverses the normalized pain score output by the normalization unit 131 of the evaluation module 190 to the original range:

[0276]

[0277] denotes the normalized final pain objective quantification index (Marke value).

[0278] Preferably, the evaluation module 190 linearly scales the original clinical pain score (e.g. VAS 0-10) to the [0, 1] interval in the training stage, so that data of different dimensions can be compared. When making predictions, the normalized value (e.g. 70Marke) output by the evaluation module 190 is converted to the Marke value (0-100Marke) used in clinical practice by linear scaling.

[0279] Preferably, the Marke value of the present application is a quantitative index calculated based on the synchronization characteristics of the thalamocortical loop (e.g. beta / gamma band energy ratio, functional connectivity strength), reflecting the neural mechanism of pain (e.g. central sensitization degree). Since the Marke value is a newly defined biophysical parameter, it has not been widely recognized by clinicians, and doctors cannot directly associate it with the subjective experience of the subject or existing diagnostic criteria (e.g. VAS score). Therefore, by mapping the Marke value to the 0-10 pain scale commonly used in clinical practice, the biological effectiveness of the Marke value can be verified by comparing the model prediction score with the true clinical score (e.g. VAS record of the subject after surgery).

[0280] The coefficient of determination (R 2 ) is used to measure the linear correlation between the predicted Marke value and the true pain score.

[0281]

[0282] y reg，i denotes the true pain score (normalized pain score label) of the ith sample obtained by the clinical gold standard (e.g. NRS, VAS scale), which has been normalized to the internal processing range of the evaluation module 190 (e.g. [0, 1]). denotes the predicted pain score of the ith sample calculated based on the Marke value, which has been mapped to the pain objective quantification index ([0, 100]) by denormalization. denotes the arithmetic mean of all sample true pain scores yreg, i.

[0283] The coefficient of determination (R 2 ) is used to measure the linear correlation between the predicted Marke value and the true pain score. Figure 8 As shown in the scatter plot, the horizontal axis is the NRS score and the vertical axis is the Marke value. Figure 8The middle scatter point distribution roughly presents a trend that the Marke value increases with the increase of the NRS score, which indicates that there is a strong linear correlation between the Marke value and the pain intensity score of the clinical standard pain scale. The determination coefficient of the Marke value of the application and the NRS score of the clinical standard pain scale (NRS) is R 2 = 0.922, which is close to 1, indicating that the fitting degree of the fitting straight line to the data is very high. This verifies the biological effectiveness of the Marke value as a pain biomarker, proves the generalization ability of the evaluation module 190 under different pain intensities, and proves that it is suitable for evaluating the whole stage from mild to severe pain.

[0284] The root mean square error (RMSE) is used to quantify the prediction accuracy.

[0285]

[0286] The root mean square error reflects the average absolute difference between the predicted value of the evaluation module 190 and the true value, and the smaller the value, the stronger the quantitative ability of the evaluation module 190 to the pain intensity. Under the 0-100 point system, RMSE = 6. RMSE < 15 indicates that the Marke value based on the spatiotemporal features of the EEG signal has individualized reliability for prediction, which provides an objective neural indicator for replacing the traditional subjective pain scale.

[0287] The evaluation device of the objective evaluation of pain of the application can also be called an electroencephalogram pain quantification analysis device. The electroencephalogram pain quantification analysis device according to the application can also be realized by the following physical hardware modules:

[0288] (1) A bioelectric signal acquisition array composed of silver / silver chloride electrodes distributed according to the international 10-20 system, connected to a preamplifier through a shielded lead, and the preamplifier contains a 0.5Hz-100Hz band-pass filter and a 60Hz notch circuit.

[0289] (2) A digital signal processor connected to an analog-to-digital converter 240 through a PCIe bus. The digital signal processor integrates a hardware-accelerated FFT coprocessor configured to perform real-time fast Fourier transform with a 512-point window length.

[0290] (3) A five-channel digital filter bank containing δ (0.5-4Hz), θ (4-8Hz), α (8-12Hz), β (12-30Hz), and y (30-100Hz) FIR filter modules.

[0291] (4) A neural network acceleration unit implemented using an ASIC chip, including a bidirectional GRU hardware pipeline and a graph convolution network module. The bidirectional GRU hardware pipeline is configured as an array of 32-bit floating-point operation units. The graph convolution network module implements electrode topology modeling through an adjacency matrix operation unit designed at the RTL level.

[0292] (5) A spatiotemporal feature fusion unit integrating a two-dimensional convolution accelerator and a feature splicing circuit. The two-dimensional convolution accelerator is configured with a programmable convolution kernel size register. The feature splicing circuit implements dimension splicing of double-channel feature vectors through a DMA controller.

[0293] (6) A probability normalization module including a Softmax hardware implementation based on the CORDIC algorithm.

[0294] (7) A central control unit using an ARM Cortex-M7 microcontroller to coordinate the timing control of each hardware module through an AXI bus.

[0295] (8) A data storage module including a DDR4 memory particle and a NAND flash chip, configured as a ring buffer to store real-time EEG data streams.

[0296] The bioelectric signal acquisition array is connected to the digital signal processor to transmit the EEG signal to the digital signal processor. After receiving the analog EEG signal of the subject from the bioelectric signal acquisition array, the bioelectric signal acquisition array digitizes the analog EEG signal through a 24-bit ADC (sampling rate 1000 Hz) and transmits it to the digital signal processor through a PCIe bus. The digital signal processor sends the frequency domain data (time-frequency feature matrix) after Fourier transform to the five-channel digital filter bank.

[0297] The digital signal processor is connected to the five-channel digital filter bank. After receiving the frequency domain data from the digital signal processor, the five-channel digital filter bank divides the frequency band through five bandpass filters (delta / theta / alpha / beta / y).

[0298] The five-channel digital filter bank is connected to the neural network acceleration unit. The five-channel digital filter bank transmits the frequency domain data of each frequency band to the neural network acceleration unit through an AXI bus. After receiving the time-frequency feature matrix divided by frequency band from the five-channel digital filter bank, the bidirectional GRU hardware pipeline in the neural network acceleration unit extracts time series features through a 32-bit floating-point operation unit array, while the graph convolution network module performs electrode topology modeling based on the adjacency matrix operation unit designed at the RTL level, outputting spatial features (first feature vector).

[0299] The neural network acceleration unit is connected with the space-time feature fusion unit. The time series features and the space features output by the bidirectional GRU hardware pipeline and the graph convolution network module are synchronously transmitted to the space-time feature fusion unit through the on-chip bus. After receiving the time series features and the space features from the neural network acceleration unit, the two-dimensional convolution accelerator in the space-time feature fusion unit calculates and outputs a second feature vector. The feature splicing circuit splices the first feature vector and the second feature vector in dimension to obtain a fusion feature vector.

[0300] The space-time feature fusion unit is connected with the probability normalization module. The fusion feature tensor is transmitted to the probability normalization module through the AXI bus. After receiving the fusion feature vector from the space-time feature fusion unit, the probability normalization module performs classification probability mapping through the parallel computing unit to generate a pain intensity quantification value of 0-100 Markes, and the final result is written into the ring buffer of the data storage module through the AHB bus.

[0301] The central control unit is connected with each module. The central control unit configures the FFT window length parameter to the digital signal processor through the AXI bus, loads the pre-trained GRU weight and the electrode adjacency matrix to the neural network acceleration unit, simultaneously manages the read-write pointer of the data storage module through the AHB bus, and realizes zero-copy transmission of the ADC raw data to the ASIC input cache through the PCIe DMA channel, so as to ensure that the end-to-end data processing delay is lower than 10 ms.

Claims

1. An assessment device for objective assessment of pain, comprising a processor (110), characterized in that, The processor (110) comprises: a frequency domain transformation module (170) for performing fast Fourier transform on the pain-related EEG signal to generate a time-frequency feature matrix; a frequency band separation module (180) for separating the time-frequency feature matrix of the EEG signal in the frequency domain into time-frequency feature matrices of five frequency bands δ, θ, α, β and γ related to pain perception; a first evaluation submodule (130) for extracting time series features of the EEG signal from the time-frequency feature matrix based on a bidirectional gated recurrent unit (132); extracting end time step features and an adjacency matrix from the time series features; and generating a first feature vector representing a global association pattern between electrodes (210) for collecting the EEG signal based on the end time step features and the adjacency matrix; a second evaluation submodule (140) for performing convolution processing on the time series features to generate a second feature vector including local spatiotemporal dynamic features of the EEG signal; splicing the first feature vector and the second feature vector along the feature dimension to form a fusion feature vector including the global association pattern between the electrodes (210) and the local spatiotemporal dynamic features; calculating a class probability distribution of the fusion feature vector and performing normalization processing to generate a pain objective quantification index corresponding to the EEG signal.

2. The evaluation device according to claim 1, characterized in that It also comprises a Bayesian updating module (150), The Bayesian updating module (150) receives the pain objective quantification index, and calculates a 95% confidence interval of the pain objective quantification index according to model cognitive uncertainty and observation noise level to generate a credibility of the 95% confidence interval evaluation result; if the confidence interval width exceeds a preset threshold, the Bayesian updating module (150) dynamically adjusts the observation noise parameter to optimize the feature weight distribution.

3. The evaluation device according to claim 1 or 2, characterized in that The frequency band separation module (180) extracts time-frequency energy distribution features in the time-frequency feature matrix of each frequency band based on a learnable frequency domain filter, and dynamically allocates weights to the time-frequency energy distribution features extracted from the time-frequency feature matrix of each frequency band based on a cross-frequency band attention mechanism, fuses the energy and phase coupling characteristics between frequency bands, and generates a time-frequency feature matrix with frequency band specificity.

4. The evaluation device according to claim 3, characterized in that The step of extracting the adjacency matrix by the first evaluation submodule (130) comprises: The end time step features are converted into a connection graph between the electrodes (210) based on a Gumbel-Softmax method, thereby forming an adjacency matrix representing the degree of neural activity coupling between the brain cortical regions.

5. The evaluation device according to claim 3, characterized in that The step of calculating the class probability distribution of the fusion feature vector by the second evaluation submodule (140) comprises: calculating the class probability distribution of the fusion feature vector, outputting a normalized value in the interval [0, 1] based on a Softmax function, and mapping to a pain objective quantification index in the interval [0, 100] through linear transformation.

6. The evaluation device according to claim 3, characterized in that The pain objective quantification index is a Marke value, the numerical interval is 0-100, and the unit is Marke; The greater the Marke value of the pain intensity mapping, the more intense the pain.

7. An assessment model for objective assessment of pain, characterized in that, The evaluation model is run by an evaluation module (190), and the evaluation module (190) comprises: a first evaluation submodule (130) configured to extract time series features of the EEG signal from a time-frequency feature matrix based on a bidirectional gated recurrent unit (132); extract end time step features and an adjacency matrix from the time series features of the pain-related EEG signal; and generate a first feature vector representing a global association pattern between electrodes (210) used to collect the EEG signal based on the end time step features and the adjacency matrix; a second evaluation submodule (140) configured to perform convolution processing on the time series features of the pain-related EEG signal to generate a second feature vector including local spatiotemporal dynamic features of the EEG signal; concatenate the first feature vector and the second feature vector along a feature dimension to form a fusion feature vector including the global association pattern between the electrodes (210) and the local spatiotemporal dynamic features; calculate a class probability distribution of the fusion feature vector and perform normalization processing to generate a pain objective quantification index corresponding to the EEG signal; wherein the time series features of the pain-related EEG signal are extracted from a time-frequency feature matrix, the time-frequency feature matrix is generated by performing fast Fourier transform on the pain-related EEG signal; the time-frequency feature matrix is divided into five frequency bands, δ, θ, α, β, and γ, which are related to pain perception by a frequency band separation module (180) in the frequency domain.

8. The model of claim 7, wherein, The step of extracting the adjacency matrix by the first evaluation submodule (130) includes: The end time step features are converted into a connection graph between the electrodes (210) based on a Gumbel-Softmax method, thereby forming the adjacency matrix representing the degree of neural activity coupling between the brain cortical regions.

9. The evaluation model according to claim 7 or 8, characterized in that The step of calculating the class probability distribution of the fusion feature vector by the second evaluation submodule (140) includes: calculating the class probability distribution of the fusion feature vector, outputting normalized values in the interval [0, 1] based on a Softmax function, and mapping to a pain objective quantification index in the interval [0, 100] through linear transformation.

10. The model of claim 9, wherein, The pain objective quantification index is a Marke value, with a numerical interval of 0-100 and a unit of Marke. The greater the Marke value of the pain intensity mapping, the more intense the pain.

Images