Search

CN-122004761-A - Pain detection device and digital pain brain model based on electroencephalogram

CN122004761ACN 122004761 ACN122004761 ACN 122004761ACN-122004761-A

Abstract

The invention relates to a pain detection device based on an electroencephalogram and a digital pain brain model. The apparatus performs pain quantification in cooperation with the first and second assessment sub-modules by a processor (110). The first evaluation sub-module (130) utilizes Gumbel-Sampling to construct a differentiable graph structure, quantifies functional connection between electrodes (210) to generate a first feature vector representing a global correlation mode, and the second evaluation sub-module (140) extracts local space-time dynamic features capturing transient characteristics of pain through two-dimensional convolution to generate a second feature vector. The device maps the global and local features into objective quantitative index Marke values in the interval of 0-100 through Softmax after fusing. In addition, a Bayesian updating module (150) is introduced to dynamically adjust the observed noise parameters and optimize the weights based on the 95% confidence interval, ensuring stability of cross-device evaluation. The invention eliminates subjective evaluation deviation and provides a high-sensitivity digital and visual objective quantification tool for the pain, namely the fifth vital sign.

Inventors

  • MA KE
  • ZHANG YUE
  • QIN CHUNHUI

Assignees

  • 上海交通大学
  • 上海睿酷医疗科技有限责任公司

Dates

Publication Date
20260512
Application Date
20250321
Priority Date
20240716

Claims (20)

  1. 1. An electroencephalogram-based pain detection device comprises a processor (110) and a terminal (300), wherein the terminal (300) is used for displaying pain objective quantification indexes representing pain degrees, the device is characterized in that the processor (110) comprises a first assessment submodule (130) for quantifying and analyzing functional connection relations between electrodes (210), a differentiable graph structure is built based on the functional connection relations between the electrodes (210) to obtain a global association mode between the electrodes (210), a second assessment submodule (140) is used for receiving time series characteristics generated by EEG signals, carrying out joint feature study along space and time dimensions, extracting local space-time dynamic characteristics capable of capturing transient characteristics of pain related nerve responses, normalizing fusion feature vectors comprising the global association mode between the electrodes (210) and the local space-time dynamic characteristics and outputting the pain objective quantification indexes.
  2. 2. The electroencephalogram-based pain detection apparatus according to claim 1, wherein the objective quantitative index of pain is Marke values, the numerical interval is 0-100, and the unit is Marke.
  3. 3. The electroencephalogram-based pain detection apparatus of claim 1, wherein the processor (110) performs a second filtering of the received EEG signal with a filtering module (120) to suppress residual interference prior to feature extraction, avoiding noise components affecting the accuracy of the extraction of the spatio-temporal features.
  4. 4. The electroencephalogram-based pain detection apparatus of claim 3, wherein the secondary filtering comprises eliminating baseline wander with a high pass filter having a cutoff frequency of 0.5Hz, and filtering myoelectric noise with a low pass filter having a cutoff frequency of 75 Hz.
  5. 5. The electroencephalogram-based pain detection apparatus of claim 1, wherein the processor (110) further comprises a signal processing module (160) for resampling the EEG signal, reducing the original sampling rate 500Hz to 250Hz, compressing the data volume while retaining pain-related frequency band information.
  6. 6. The electroencephalogram-based pain detection apparatus according to claim 1, wherein the processor (110) further comprises a frequency band separation module (180) for dividing a time-frequency characteristic matrix of an EEG signal in the frequency domain into time-frequency characteristic matrices of five frequency bands δ, θ, α, β and γ related to pain perception.
  7. 7. The electroencephalogram-based pain detection apparatus according to claim 6, wherein the frequency band separation module (180) extracts time-frequency energy distribution characteristics of each frequency band based on a learnable frequency domain filter, and fuses energy and phase coupling characteristics between frequency bands by a cross-frequency-band attention mechanism.
  8. 8. The electroencephalogram-based pain detection apparatus of claim 1, wherein the first evaluation sub-module (130) extracts an end time step feature HBiGRUlast to integrate a full-phase change in signal from stimulation start to end.
  9. 9. The electroencephalogram-based pain detection apparatus according to claim 1, wherein the differentiable graph structure adds gummel noise to an adjacency matrix by gummel-Sampler unit (133) and performs a micro-samplings to allow gradient back-pass of discrete edge connection decisions.
  10. 10. The electroencephalogram-based pain detection apparatus according to claim 9, wherein the gummel-Sampler unit (133) is provided with a temperature parameter τ to control the degree of softness of the sampling, so as to accommodate differences in functional connectivity of different individual brains in a painful state.
  11. 11. The digital pain brain model is characterized by comprising a graph generation network for generating a first characteristic vector representing a global association mode between electrodes (210) for acquiring EEG signals, and a convolution neural network for generating pain objective quantification indexes corresponding to the EEG signals according to the first characteristic vector and a second characteristic vector generated by the convolution neural network.
  12. 12. The digital pain brain model according to claim 11, wherein the graph generating network comprises a bi-directional gating loop unit (132) with hidden layer dimensions configured to D' =256 for processing inputs along forward and reverse time windows, respectively, extracting timing features with hidden state stitching to 512 dimensions.
  13. 13. The digital pain brain model according to claim 11, wherein the graph generation network converts the end time-step feature into an adjacency matrix a using gummel-Softmax method, wherein adjacency matrix element Ai, j represents the strength of association between electrodes i and j.
  14. 14. The digital pain brain model according to claim 13, wherein the construction of the adjacency matrix a takes into account a spatial decay term and a functional coupling term, the spatial decay term being determined by the electrode pair euclidean distance dij.
  15. 15. The digital pain brain model according to claim 11, wherein the convolutional neural network captures local space-time coupled features through a set of multi-scale convolutional kernels, outputting a second feature vector ZCNN.
  16. 16. The digital pain brain model according to claim 15, wherein the first eigenvector ZGNN and the second eigenvector ZCNN are stitched along a characteristic dimension to form a fused eigenvector Zconcat.
  17. 17. The digital pain brain model according to claim 16, wherein the fusion feature vector Zconcat is input to a classification unit, and the class probability distribution yclass is calculated using a Softmax function.
  18. 18. The digital pain brain model according to claim 11, wherein the training data of the model is derived from a Database comprising multi-center EEG signal data, including a digital twin pain brain model Database, PAIN Consortium Database or INCF Pain-EEG Database.
  19. 19. The digital pain brain model according to claim 11, wherein the model employs a weighted cross entropy loss function to assign differential weights to different pain classes to balance the problem of class maldistribution.
  20. 20. The digital pain brain model according to claim 11, wherein the model quantifies pain intensity prediction error using a Huber loss function, the loss function being controlled by a threshold parameter δ to switch from a quadratic form to a linear form to resist outlier interference.

Description

Pain detection device and digital pain brain model based on electroencephalogram The original basis of the divisional application is the patent application with the application number 202510350414.5, the application date 2025, the month 03 and the invention name of 'an assessment device and an assessment model for objective assessment of pain'. Technical Field The invention relates to the technical field of pain assessment, in particular to a pain detection device based on an electroencephalogram and a digital pain brain model. Background Chronic pain refers to an unpleasant sensory and emotional experience, or a similar experience, caused by actual or potential tissue damage. Pain is considered as the fifth leading sign, equally important as respiration, blood pressure, pulse and body temperature, and is an important indicator for measuring and monitoring a person's health. Pain is a fifth vital sign, and has a deep significance in that it represents the temperature of human care in the medical development process, so that the observation of the vital signs of a subject is not only the collection of physiological data, but also the deep care and attention of life. However, pain also faces an important congenital problem as a fifth vital sign, namely the lack of objective biomarkers. Currently, pain scores rely primarily on subject self-assessment and physician professional assessment by different scales, such as Visual Analog Scoring (VAS), digital grading scoring (NRS), and DN4 and IDpain scales for neuropathic pain. One significant problem with this scoring approach is the lack of objective indicators. The existing pain scores are usually determined by subjective judgment of the subject or physician who will choose a number in the range of 0-100 to indicate the severity of the pain. For example, 0-39 indicates mild pain, 40-69 indicates moderate pain, and 70-100 indicates severe pain. The subjective feeling-based scoring mode ensures that the clinical evaluation of pain cannot accurately reflect the actual clinical situation, and has great negative influence on the establishment of clinical diagnosis and treatment schemes, the evaluation of treatment effects and even the medical related identification. Electroencephalographic devices are one of the main tools for the current study of pain. The brain signals recorded by EEG reflect the voltage fluctuations generated by neuronal firing, and recent studies have shown that voltage changes in different brain regions are associated with a certain degree and type of pain. However, such voltage changes do not accurately represent the pain severity of the subject. Furthermore, EEG equipment is expensive, data processing is complex, and signal acquisition is required in a specialized hospital. Although scientific research has found that EEG has certain advantages in assessing pain levels in recent years, it has not been an objective measure of pain due to limitations such as signal acquisition and expert analysis. CN115867183a discloses a method for monitoring pain levels, the method comprising receiving measurement data derived from electroencephalogram (EEG) data collected by one or more EEG electrodes, extracting a set of indices from the EEG signal, the indices being associated with power in the θ - α frequency range, and determining the pain level (LoP) based thereon, i.e. a value indicative of the pain level in the subject. However, the assessment model in this solution fails to objectively reflect the severity of pain. The difference in the functional connection of the frontal cortex and the anterior cingulate cortex to the whole brain can be confirmed by functional magnetic resonance imaging (fMRI) for painful and non-painful persons. Based on the research result, electroencephalogram (EEG) is adopted for further verification, and by means of electroencephalogram monitoring under various scenes such as sitting, two-dimensional pictures, virtual Reality (VR) and the like, the electric activity change mechanisms of different brain areas during pain occurrence and alleviation are revealed. For example, CN112957014A discloses a pain detection positioning method based on brain waves and a neural network, which comprises the following steps of removing original brain wave noise by adopting an independent component analysis algorithm, dividing brain wave pain levels, dividing each pain level into a plurality of equal-length time windows, obtaining a multichannel brain wave time sequence, obtaining a preprocessed pain data set, respectively generating spectral topographic maps of Theta, alpha and Beta frequency bands related to pain by Fourier transformation, azimuth equidistant projection and CloughTocher interpolation algorithm, merging the spectral topographic maps into multichannel brain wave sequences serving as input of the CNN LSTM-AM neural network, constructing the CNN-LSTM-AM neural network, training the CNN-LSTM-AM neural network, obtaining time space f