Seedness depth monitoring method based on graph convolutional neural network

Abstract

The invention discloses a sedation depth monitoring method based on a graph convolutional neural network, and relates to the technical field of sedation monitoring, and the method comprises the steps: collecting a multi-source biomarker of a patient, carrying out the filtering and artifact removal of an electroencephalogram signal, and generating preprocessed electroencephalogram data; performing attention pooling processing on the node-level feature tensor, calculating a sedation depth index, and generating a pain interference factor according to a connection edge weight of the dynamic function connection graph; and when the pain interference factor exceeds a preset pain threshold value and the sedation depth index meets a preset condition, generating a sedation optimization instruction, and when the virtual pharmacokinetic node displays metabolic abnormality, generating an infusion rate adjustment instruction. According to the method, through a dynamic function connection diagram construction link, the time-varying intensity of the skin conductance reaction signal and the rugosa electromyographic signal is used as a dynamic weight, and accurate quantification of the pain-sedation coupling effect is achieved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of sedation monitoring, and particularly relates to a sedation depth monitoring method based on a graph convolutional neural network. BACKGROUND

[0002] Sedation depth monitoring technology has developed from single electroencephalogram analysis to multi-modal physiological signal fusion. In recent years, graph neural networks have made progress in modeling brain functional connectivity, which can capture nonlinear interactions between brain regions. Meanwhile, the development of pharmacogenomics makes it possible to predict pharmacokinetics based on genotypes. In clinical practice, multi-parameter integrated monitoring improves monitoring accuracy by fusing electroencephalogram frequency band features and physiological response signals, and knowledge graphs are gradually applied to the standardized expression of drug metabolism rules.

[0003] However, the prior art still has deficiencies. Pain interference and insufficient sedation present similar electroencephalogram low-frequency suppression phenomena at the neurophysiological level, making it difficult for traditional models to distinguish between them. The fundamental reason is the lack of dynamic quantitative modeling of the pain-sedation coupling effect and real-time feedback mechanism for individual pharmacokinetic differences. SUMMARY

[0004] In view of the above existing problems, the present application is proposed.

[0005] Therefore, the present application provides a sedation depth monitoring method based on a graph convolutional neural network to solve the monitoring distortion problem caused by pain interference and individual differences in drug metabolism in sedation depth monitoring.

[0006] To solve the above technical problems, the present application provides the following technical solutions: The present application provides a sedation depth monitoring method based on a graph convolutional neural network, which includes collecting multiple source biomarkers of a patient, filtering and removing artifacts from electroencephalogram signals to generate preprocessed electroencephalogram data; extracting multiple frequency band power features as nodes from the preprocessed electroencephalogram data, using the time-varying intensity of the skin conductance response signal and the corrugator electromyography signal as dynamic weights to construct connection edges between nodes and generate a functional connectivity graph; analyzing drug metabolism sites in genotype data, querying a constructed drug metabolism knowledge graph to generate a pharmacokinetics parameter vector, and converting it into a virtual pharmacokinetics node embedded in the functional connectivity graph to generate a dynamic functional connectivity graph; inputting the dynamic functional connectivity graph into a graph convolutional neural network, combining the adjacent convolution of the virtual pharmacokinetics node through a hypergraph adaptation operator to generate a node-level feature tensor; performing attention pooling processing on the node-level feature tensor to calculate a sedation depth index, generating a pain interference factor according to the connection edge weights of the dynamic functional connectivity graph; generating a sedation optimization instruction when the pain interference factor exceeds a preset pain threshold and the sedation depth index meets a preset condition, and generating an infusion rate adjustment instruction when the virtual pharmacokinetics node shows abnormal metabolism.

[0007] As a preferred scheme of the sedation depth monitoring method based on the graph convolutional neural network, the multi-source biomarker comprises electroencephalogram signals, skin conductance response signals, corrugator electromyography signals and genotype data.

[0008] As a preferred scheme of the sedation depth monitoring method based on the graph convolutional neural network, the preprocessed electroencephalogram data is generated by the following specific steps, The electroencephalogram signals are amplified to generate amplified electroencephalogram signals, the amplified electroencephalogram signals are subjected to first band-pass filtering to generate wideband filtered signals, and second narrowband filtering is synchronously performed to generate sedation characteristic frequency band signals; The wideband filtered signals are subjected to empirical mode decomposition to separate high-frequency artifact components; The sedation characteristic frequency band signals and the high-frequency artifact components are subjected to coherence analysis to eliminate artifact components irrelevant to sedation characteristics, thereby generating artifact suppression signals; The artifact suppression signals are subjected to multi-scale wavelet decomposition and denoising, and are reconstructed into denoised artifact suppression signals, which are input into a sparse autoencoder network to generate preprocessed electroencephalogram data.

[0009] As a preferred scheme of the sedation depth monitoring method based on the graph convolutional neural network, the connection edges between nodes are constructed to generate a functional connection graph by the following specific steps, First, second, third and fourth frequency band power features are extracted from the preprocessed electroencephalogram data as nodes; The skin conductance response signals are subjected to Hilbert-Huang transform decomposition to extract energy entropy values, and the corrugator electromyography signals are subjected to Teager energy analysis to extract transient energy peak values, and dynamic weight values are generated by matching time sequence energy fluctuations of the energy entropy values and the transient energy peak values; The time delay correlation strength between nodes in each frequency band power feature node is calculated, and the dynamic weight values are used to modulate the connection edge weights between nodes, thereby forming directed connection edges between nodes; The frequency band power feature nodes and the directed connection edges between nodes are used as connection relationships to generate a functional connection graph.

[0010] As a preferred scheme of the sedation depth monitoring method based on the graph convolutional neural network, the drug metabolism knowledge graph is constructed by mapping the corresponding relationship of genotype-metabolic phenotype-drug response parameters to structured rules by calling pre-stored pharmacogenomic data and clinical pharmacokinetic data.

[0011] As a preferred scheme of the sedation depth monitoring method based on the graph convolutional neural network, wherein: the drug metabolism knowledge graph constructed by the query generates a pharmacokinetic parameter vector, and the specific steps are as follows, The drug metabolism related sites in the genotype data are analyzed to generate a site embedding vector, and the genotype node in the drug metabolism knowledge graph is converted into a node embedding vector; The cosine similarity of the site embedding vector and the node embedding vector is calculated, and when the cosine similarity exceeds a preset similarity threshold, the genotype node and the associated metabolic phenotype node and drug response parameter node are activated to form an activated subgraph; The drug response parameter nodes in the activated subgraph are aggregated to generate a pharmacokinetic parameter vector.

[0012] As a preferred scheme of the sedation depth monitoring method based on the graph convolutional neural network, wherein: the embedding function connection graph, the specific steps are as follows, The pharmacokinetic parameter vector is converted into a pharmacokinetic parameter sequence, each frequency band power feature node is converted into an electroencephalogram frequency band power sequence, and the correlation coefficient of the pharmacokinetic parameter sequence and the electroencephalogram frequency band power sequence is calculated to generate the interaction strength between the brain region and the drug metabolism; According to the interaction strength between the brain region and the drug metabolism, all interaction strengths are aggregated to generate a virtual pharmacokinetic node and embed the functional connection graph to generate a dynamic functional connection graph.

[0013] As a preferred scheme of the sedation depth monitoring method based on the graph convolutional neural network, wherein: the generated node-level feature tensor, the specific steps are as follows, Each frequency band power feature node is taken as a base space, and a virtual pharmacokinetic node is taken as a fiber space to construct a fiber bundle geometry; According to the Riemann curvature of the fiber bundle geometry, an adjacent convolution operation is performed on the base space and the fiber space to generate a fiber bundle connection parameter; The fiber bundle connection parameter is optimized by the Chen-Weinberg gauge theory, and the Hodge decomposition is performed to generate a node-level feature tensor.

[0014] As a preferred scheme of the sedation depth monitoring method based on the graph convolutional neural network, wherein: the generated pain interference factor, the specific steps are as follows, The node-level feature tensor is processed by cross-modal attention pooling to generate a fusion feature vector; The fusion feature vector is input into a hierarchical regression network to calculate a sedation depth index; According to the connection edge weight of the dynamic functional connection graph, a dynamic graph entropy value is calculated; The dynamic graph entropy value is analyzed by multi-scale wavelet analysis to generate a pain interference factor.

[0015] As a preferred scheme of the sedation depth monitoring method based on the graph convolutional neural network, the specific steps of generating the infusion rate adjustment instruction are as follows, When the pain interference factor exceeds the preset pain threshold and the sedation depth index is lower than the first sedation threshold, a sedation depth improvement instruction is generated. When the pain interference factor exceeds the preset pain threshold and the sedation depth index is between the first and second sedation thresholds, a painkiller drug addition instruction is generated. The clearance rate parameter value is analyzed from the virtual pharmacokinetic node, and when the clearance rate parameter value exceeds the tolerance error, it is determined that the metabolism is abnormal. According to the severity of the metabolic abnormality, the infusion rate is adjusted, and an infusion rate adjustment instruction is generated.

[0016] The present application has the following advantages: through the dynamic functional connection graph construction link, the time-varying intensity of the skin conductance response signal and the frown muscle electrical signal is taken as the dynamic weight, and the precise quantification of the pain-sedation coupling effect is realized. The sympathetic nerve activation intensity and the body movement response intensity are fused into the dynamic edge weight, the connection strength between the nodes in the brain electrical frequency band is modulated, and the frequency domain disturbance characteristics of the pain interference on the brain function network are captured; the specificity of the pain-related brain network disorder can be improved, the false positive rate is reduced, the dynamic weight changes nonlinearly with the pain stimulation intensity, the pain contribution in the frequency band power attenuation is effectively separated, and high signal-to-noise ratio is provided for sedation optimization decision. BRIEF DESCRIPTION OF DRAWINGS

[0017] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following will briefly introduce the drawings needed to be used in the embodiment description. Obviously, the drawings in the following description are only some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor on the basis of these drawings.

[0018] Fig. 1 The flowchart of the sedation depth monitoring method based on the graph convolutional neural network.

[0019] Fig. 2 The flowchart for preprocessing the electroencephalogram data.

[0020] Fig. 3 The flowchart for constructing the functional connection graph.

[0021] Fig. 4 The flowchart for constructing the activation subgraph. DETAILED DESCRIPTION

[0022] In order to make the above objectives, characteristics and advantages of the present application more obvious and easy to understand, the specific embodiments of the present application will be described in detail below with reference to the accompanying drawings.

[0023] In the following description, a large number of specific details are set forth in order to facilitate a thorough understanding of the present application, but the present application can also be implemented in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the concept of the present application, therefore the present application is not limited to the specific embodiments disclosed below.

[0024] Secondly, the "one embodiment" or "embodiment" referred to herein means that the specific features, structures or characteristics can be included in at least one implementation of the present application. "In one embodiment" appearing in different places in the specification does not mean the same embodiment, nor is it an embodiment that is independent of or selected from other embodiments.

[0025] Reference Figs. 1-4 For one embodiment of the present application, the embodiment provides a sedation depth monitoring method based on graph convolutional neural network, comprising the following steps: S1, collecting patient multi-source biomarkers, filtering and removing artifacts of electroencephalogram signals to generate preprocessed electroencephalogram data.

[0026] S1.1: The multi-source biomarkers include electroencephalogram signals, skin conductance response signals, corrugator electromyography signals and genotype data.

[0027] It should be noted that the electroencephalogram signals of the patient are collected by electroencephalogram electrodes; the skin conductance response signals are collected by skin conductance sensors; the corrugator electromyography signals are collected by electromyography electrodes; and the genotype data is collected by gene sequencing equipment; The electroencephalogram signals refer to bioelectric signals reflecting the electrical activity of brain neurons; the skin conductance response signals refer to the changes in skin conductance caused by sweat gland activity, reflecting the arousal state of the autonomic nervous system; the corrugator electromyography signals refer to the electromyography signals generated by the contraction of the orbicularis oculi muscle, which are used to assess pain or discomfort response; and the genotype data refers to the nucleotide sequence information of the patient's drug metabolism related gene sites.

[0028] S1.2: The electroencephalogram signals are amplified to generate amplified electroencephalogram signals, the amplified electroencephalogram signals are first band-pass filtered to generate wideband filtered signals, and second narrowband filtering is performed synchronously to generate sedation characteristic frequency band signals; Specifically, the electroencephalogram signals are connected to the positive and negative input terminals through a differential amplifier to suppress common mode interference and amplify the voltage difference of weak electroencephalogram signals, thereby generating amplified electroencephalogram signals; A first band-pass filter is applied to the amplified EEG signal to allow frequency components in a low-to-high frequency range to pass and attenuate frequency components outside the low-to-high frequency range, generating a wide-band filtered signal; Meanwhile, a second narrow-band filter is applied to the amplified EEG signal to only allow narrow-band frequency components associated with the sedation state to pass and suppress narrow-band frequency components not associated with the sedation state, generating a sedation characteristic frequency band signal.

[0029] It should be noted that the frequency components in the low-to-high frequency range refer to the complete frequency spectrum band in the EEG signal that covers the main physiological activity characteristics, including continuous frequency distribution from slow waves to fast waves; The narrow-band frequency components associated with the sedation state refer to specific slow wave frequency bands in the EEG signal that are directly related to the depth of sedation, and can reflect the drug-induced neural inhibition state.

[0030] S1.3: Empirical mode decomposition is performed on the wide-band filtered signal to separate high-frequency artifact components; Specifically, all local maximum points and minimum points of the wide-band filtered signal are identified, the maximum points are connected by cubic spline interpolation to form an upper envelope line, and the minimum points are connected to form a lower envelope line. The mean value of the upper and lower envelope lines is calculated to form a mean envelope line. The difference between the wide-band filtered signal and the mean envelope line is taken as a candidate intrinsic mode, and the wide-band filtered signal is decomposed into multiple intrinsic modes arranged from high frequency to low frequency through iterative screening. The intrinsic mode with a frequency higher than the upper limit of the sedation characteristic frequency band signal is identified as a high-frequency artifact component.

[0031] S1.4: Coherence analysis is performed on the sedation characteristic frequency band signal and the high-frequency artifact component to eliminate artifact components unrelated to the sedation characteristic, generating an artifact-suppressed signal; Specifically, frequency domain transformation is performed on the sedation characteristic frequency band signal to extract the sedation energy distribution value of the sedation characteristic frequency band signal at each frequency point, and frequency domain transformation is performed on the high-frequency artifact component to extract the high-frequency energy distribution value of the high-frequency artifact component at each frequency point. According to the energy distribution intensity of the sedation characteristic frequency band signal and the high-frequency artifact component at the same frequency point, a joint energy distribution intensity is generated. The ratio of the square of the joint energy distribution intensity to the product of the sedation energy distribution value and the high-frequency energy distribution value is taken as the coherence coefficient at the frequency point; If the coherence coefficient at the frequency point is lower than the preset coherence threshold, it is determined that the high-frequency artifact component at the frequency point is not related to the sedation characteristic and is directly eliminated. The high-frequency artifact component with a coherence coefficient at the frequency point exceeding the preset coherence threshold is retained, and the high-frequency artifact component is superimposed with the sedation characteristic frequency band signal to generate an artifact-suppressed signal.

[0032] It should be noted that the preset coherence threshold is set based on the conventional experience standard for distinguishing relevant and irrelevant artifacts in electroencephalogram signal processing, and an example value is 0.5.

[0033] S1.5: The artifact suppression signal is subjected to multi-scale wavelet decomposition denoising, and is reconstructed into a denoised artifact suppression signal, which is input into a sparse autoencoder network to generate preprocessed electroencephalogram data.

[0034] It should be noted that the pre-training process of the sparse autoencoder network is as follows: historical patient electroencephalogram signals are collected, and historical preprocessed electroencephalogram data are used as training data sets at the same time; a fully connected neural network with a symmetric structure of an encoder and a decoder is used, wherein the number of neurons in the encoder layer is gradually reduced to a low-dimensional representation, and the decoder layer is symmetrically increased to the original signal dimension; a sparse constraint condition is applied to the output layer of the encoder during the training process; the network weight parameters are iteratively adjusted by an optimization algorithm, and the difference error between the input preprocessed electroencephalogram data and the decoder reconstructed signal is minimized while satisfying the sparse constraint condition (an optimization restriction mechanism that forces some neurons in the encoder part to remain in a silent state and only allows a small number of neurons to be activated), until the network reconstruction performance reaches a stable convergence state, the pre-training is completed, and the trained sparse autoencoder network is obtained.

[0035] Specifically, the artifact suppression signal is subjected to multi-scale decomposition using a Daubechies wavelet basis function to generate wavelet coefficients of different scales; According to the correlation strength of the wavelet coefficients of adjacent scales, if the correlation strength is lower than a preset correlation threshold, it is determined that it is a noise component and is set to zero, and an inverse wavelet transform is performed on the denoised wavelet coefficients to reconstruct the denoised artifact suppression signal; The denoised artifact suppression signal is input into the encoder part of the sparse autoencoder network for signal pattern recognition, and the number of neurons is reduced layer by layer through the fully connected layer to generate a low-dimensional compressed feature vector; The low-dimensional compressed feature vector is input into the decoder part, the number of neurons is increased layer by layer through the fully connected layer to ensure consistency with the dimension of the denoised artifact suppression signal, and the reconstructed artifact suppression signal is output as the preprocessed electroencephalogram data.

[0036] It should be noted that the preset correlation threshold is set based on the statistical experience criterion for distinguishing noise and effective components in wavelet decomposition of electroencephalogram signals, and an example value is 0.7.

[0037] S2, a plurality of frequency band power features are extracted from the preprocessed electroencephalogram data as nodes, and a functional connection graph is constructed by taking the time-varying intensity of the skin conductance response signal and the corrugator electromyogram signal as dynamic weights.

[0038] S2.1: Extracting the first, second, third and fourth band power features from the preprocessed electroencephalogram data as nodes; Specifically, the preprocessed electroencephalogram data is input into the first, second, third and fourth band pass filters, which are physical electronic circuits (composed of resistance, capacitance and inductance elements); The physical electronic circuit automatically allows the current of the target frequency band to pass through according to the capacitive reactance and inductive reactance characteristics of the electronic elements, while physically attenuating the current outside the target frequency band range. The current passing through the band pass filter is converted into a voltage waveform at the output terminal, generating the first, second, third and fourth band signals; The first, second, third and fourth band signals are input into the electroencephalograph, and the signal energy intensity is generated by the squarer in the built-in circuit. The squared value of the voltage amplitude of each band signal is related to the signal energy intensity by the integrator in the built-in circuit. In a fixed time window, the integral circuit automatically accumulates the squared value of the voltage amplitude of each band signal to generate a time window total energy representation value. The time window total energy representation value is converted into a scalar by the linear calibration circuit, and the first, second, third and fourth band power features are output; The first, second, third and fourth band power features are directly used as nodes.

[0039] Example: The first, second, third and fourth band pass filters cover the delta wave, theta wave, alpha wave and beta wave bands respectively.

[0040] S2.2: Hilbert-Huang transform is performed on the skin conductance response signal to extract the energy entropy value, Teager energy analysis is performed on the corrugator muscle electrical signal to extract the transient energy peak value, and the time sequence energy fluctuation of the energy entropy value and the transient energy peak value is matched to generate a dynamic weight value; Specifically, the skin conductance response signal is input into the biological signal amplifier channel one, and is standardized into a unified dimension signal by the gain adjusting potentiometer to generate a standardized skin conductance response signal; The corrugator muscle electrical signal is input into the biological signal amplifier channel two, and is standardized into a unified dimension signal by the independent gain adjusting potentiometer to generate a standardized corrugator muscle electrical signal; The standardized skin conductance response signal is connected to the time constant circuit, and the exponential decay voltage is realized through the inherent physical properties of the resistor-capacitor elements, and the decay skin conductance response voltage signal is output; The standardized corrugator muscle electrical signal is connected to the complementary time constant circuit, and the exponential enhancement voltage is realized through the physical properties of the reverse resistor-capacitor elements, and the enhanced corrugator muscle electrical voltage signal is output; The attenuated skin conductance response voltage signal and the enhanced frowning muscle electrophysiological voltage signal are input in parallel into the analog adder circuit, and are physically superimposed according to Kirchhoff's current law to generate a dynamic weight value.

[0041] S2.3: Calculate the time-delay correlation strength between nodes in the power characteristic nodes of each frequency band, and use dynamic weight values ​​to modulate the weight of the connection edges between nodes to form directed connection edges between nodes; Specifically, the time delay correlation strength and dynamic weight value between each node are input in parallel into a four-quadrant analog multiplier circuit. The product voltage is automatically output as the weight voltage signal of the connection edge through the physical characteristics of the circuit. At the same time, based on the time offset direction determined by the time delay correlation analysis (such as the frequency band power characteristic node signal being earlier than the frequency band power characteristic node signal), the trigger direction control relay is used to conduct the corresponding current path, forming a directed connection edge between nodes.

[0042] The time-delay correlation strength between nodes in the power characteristic nodes of each frequency band is calculated using the following expression: ; In the formula, Represents the power characteristic nodes of the target frequency band With reference band power characteristic node The time-delay correlation strength, Indicates the index of the target frequency band power feature node. Indicates the power characteristic node of the reference frequency band. Indicates the time delay offset. Indicates the maximum negative time delay. Indicates the maximum positive time delay. Indicates within the time delay range Seeking the element that maximizes the time-delay correlation strength value, Indicates the total length of the time series. Indicates the effective calculation length. Indicates a point-in-time index. Indicates index for time points From 1 to Summation operation, Represents the power characteristic nodes of the target frequency band At the point of time The frequency band power value, Represents the power characteristic nodes of the target frequency band The time window average of the frequency band power value. Represents the power characteristic nodes of the target frequency band At the point of time The frequency band power deviation value, Indicates the power characteristic node of the reference frequency band. At the point of time The frequency band power value, Indicates the power characteristic node of the reference frequency band. The time window average of the frequency band power value. Indicates the power characteristic node of the reference frequency band. At the point of time The frequency band power deviation value, Represents the power characteristic nodes of the target frequency band The standard deviation of the frequency band power values, Indicates the power characteristic node of the reference frequency band. The standard deviation of the frequency band power values, This represents the normalized denominator.

[0043] It should be noted that, , and The dimension is μV 2 , , and All measurements are dimensionless, with the number of sampling points as the basic unit (e.g., at a sampling rate of 256Hz, 1 second = 256 points). The dimension is (μV) 2 ) 2 , The dimension is (μV) 2 ) 2 Final output It is dimensionless, but we maintain dimensional consistency.

[0044] S2.4: Generate a functional connection graph by using the frequency band power characteristic nodes and the directed connection edges between nodes as connection relationships.

[0045] Specifically, the power characteristics of the first, second, third, and fourth frequency bands are used as voltage source nodes connected to the circuit backplane terminals. At the same time, the directed connection edges between the nodes are used as weighted directional current paths. The terminals of the frequency band power characteristic nodes are connected to the relay contact terminals of the directed connection edges between the nodes through physical wiring on the backplane, forming a complete circuit network with four sets of voltage source nodes and multiple sets of weighted redirection paths. The topology of the complete circuit network is directly implemented as a functional connection diagram.

[0046] More preferably, by fusing the multi-modal biological signal dynamic weight mechanism, the limitations of traditional functional connection graph construction methods (such as fixed weight models based on static correlation coefficients or phase lock values) are broken through, the autonomic nervous awakening state represented by the skin conductance response signal and the pain motor response represented by the corrugator muscle electromyography signal are fused to generate a dynamic weight value through time decay, the connection strength between nodes in the electroencephalogram frequency band is modulated in real time, the quantitative coupling relationship between the pain stress response and the dynamic evolution of the brain functional network is established, the distortion problem of the sedation evaluation caused by ignoring the pain interference in the traditional method is overcome, and the time-varying weight generation mechanism realized by the hardware circuit makes the functional connection graph have millisecond-level response capability, and improves the detection sensitivity of the pain interference and the accuracy of the depth evaluation of the sedation.

[0047] S3, analyze the drug metabolism sites in the genotype data, query the constructed drug metabolism knowledge graph to generate the pharmacokinetic parameter vector, and convert it into a virtual pharmacokinetic node embedding functional connection graph to generate a dynamic functional connection graph.

[0048] S3.1: The drug metabolism knowledge graph is constructed by mapping the corresponding relationship of genotype-metabolic phenotype-drug response parameters to structured rules by calling pre-stored pharmacogenomic data and clinical pharmacokinetic data.

[0049] It should be noted that the pre-stored pharmacogenomic data refers to the corresponding relationship between the nucleotide sequence of the patient's drug metabolism related gene site and the metabolic phenotype (such as fast metabolism / slow metabolism); The pre-stored clinical pharmacokinetic data refers to the quantitative values of the clearance rate, distribution volume and other parameters of the sedative drug in different metabolic phenotype groups.

[0050] Specifically, the genotype-metabolic phenotype mapping table in the pre-stored pharmacogenomic data and the metabolic phenotype-drug response parameter mapping table in the clinical pharmacokinetic data are called; The corresponding relationship of the genotype node, the metabolic phenotype node and the drug response parameter node is converted into a structured rule (the rule format is: if the genotype node meets a specific nucleotide sequence, the metabolic phenotype node is associated and further associated with the clearance rate parameter node and the distribution volume parameter node), the directed edge pointing to the metabolic phenotype node is constructed from the genotype node, the directed edge pointing to the drug response parameter node is constructed from the metabolic phenotype node, and all nodes and directed edges are integrated to form a directed graph structure with a three-level cascade structure, which is used as a drug metabolism knowledge graph.

[0051] S3.2: Analyze the drug metabolism related sites in the genotype data, generate a site embedding vector, and convert the genotype node in the drug metabolism knowledge graph into a node embedding vector; Specifically, the nucleotide sequence of a drug metabolism related site in the genotype data is extracted, each base in the nucleotide sequence is encoded into a fixed-length numerical coding sequence by a classification coding method to generate a site embedding vector; meanwhile, the genotype node in the drug metabolism knowledge graph is encoded into a numerical coding sequence of the same dimension by the classification coding method to generate a node embedding vector.

[0052] S3.3: Calculate the cosine similarity of the site embedding vector and the node embedding vector, and when the cosine similarity exceeds a preset similarity threshold, activate the genotype node and the associated metabolic phenotype node and drug response parameter node to form an activated subgraph. It should be noted that the preset similarity threshold is set based on the clinical determination standard of genotype matching in pharmacogenomics, and an example value is 0.9.

[0053] Specifically, the genotype node with a cosine similarity exceeding the preset similarity threshold is marked as an effective genotype node, and according to the directed edge connection relationship of the drug metabolism knowledge graph, the metabolic phenotype node directly connected with the effective genotype node and the drug response parameter node connected through the metabolic phenotype node are searched through signal pattern recognition, and all effective genotype nodes, searched metabolic phenotype nodes and drug response parameter nodes, and original directed edge connection relationships in the drug metabolism knowledge graph are aggregated to form an activated subgraph.

[0054] It should be noted that the genotype node with a cosine similarity not exceeding the preset similarity threshold is marked as an invalid genotype node and is directly excluded.

[0055] The cosine similarity of the site embedding vector and the node embedding vector is calculated and represented as: ; In the formula, denotes the cosine similarity of the site embedding vector and the node embedding vector, denotes the site embedding vector, denotes the node embedding vector, denotes the dimension of the embedding vector, denotes the dimension index, denotes the sum of all dimensions from to , denotes the eigenvalue of the site embedding vector in the first dimension, denotes the eigenvalue of the node embedding vector in the first dimension, denotes the eigenvalue of the site embedding vector in the first square of eigenvalue of dimension, representing node embedding vector In the first square of eigenvalue of dimension.

[0056] It should be noted that, , , , and are dimensionless, and the final output is dimensionless, keeping the dimension consistent.

[0057] S3.4: Aggregating the drug response parameter nodes in the activated subgraph to generate a pharmacokinetic parameter vector.

[0058] Specifically, two types of drug response parameter nodes, clearance rate parameter nodes and distribution volume parameter nodes, are separated from the activated subgraph; The numerical values of all clearance rate parameter nodes are averaged to generate a mean clearance rate, and the numerical values of all distribution volume parameter nodes are averaged to generate a mean distribution volume; The mean clearance rate and the mean distribution volume are combined in the order of clearance rate priority to form a two-dimensional numerical vector, generating a pharmacokinetic parameter vector.

[0059] S3.5: Convert the pharmacokinetic parameter vector into a pharmacokinetic parameter sequence, and convert each frequency band power feature node into an electroencephalogram frequency band power sequence. By calculating the correlation coefficient between the pharmacokinetic parameter sequence and the electroencephalogram frequency band power sequence, the interaction strength between the brain region and drug metabolism is generated. Specifically, the pharmacokinetic parameter vector is divided into clearance rate parameter sequences and distribution volume parameter sequences for consecutive time periods according to a fixed time window length, generating a pharmacokinetic parameter sequence. At the same time, the power values of the first frequency band power feature node, the second frequency band power feature node, the third frequency band power feature node, and the fourth frequency band power feature node in the same time period are obtained, generating an electroencephalogram frequency band power sequence. According to the ratio of the clearance rate parameter sequence in the pharmacokinetic parameter sequence to the electroencephalogram frequency band power sequence, a Pearson correlation coefficient is generated. According to the ratio of the distribution volume parameter sequence in the pharmacokinetic parameter sequence to the electroencephalogram frequency band power sequence, a Pearson correlation coefficient is generated. The absolute value of each Pearson correlation coefficient is taken as the interaction strength between the brain region and drug metabolism.

[0060] S3.6: According to the interaction strength between the brain region and drug metabolism, by aggregating all interaction strengths, a virtual pharmacokinetic node is generated and embedded in the functional connectivity graph, generating a dynamic functional connectivity graph.

[0061] Specifically, the interaction strength between the brain region and the drug metabolism includes eight groups of interaction strength values of the clearance rate parameter and the first frequency band power feature, the clearance rate parameter and the second frequency band power feature, the clearance rate parameter and the third frequency band power feature, the clearance rate parameter and the fourth frequency band power feature, the distribution volume parameter and the first frequency band power feature, the distribution volume parameter and the second frequency band power feature, the distribution volume parameter and the third frequency band power feature, and the distribution volume parameter and the fourth frequency band power feature. By aggregating the eight groups of interaction strength values, the average value of the eight groups of interaction strength values is calculated to generate a single numerical value as the weight attribute value of the virtual pharmacokinetic node. The weight attribute value of the virtual pharmacokinetic node is directly assigned to the virtual pharmacokinetic node to form the virtual pharmacokinetic node. The virtual pharmacokinetic node is connected to all frequency band power feature nodes in the functional connection graph to form a dynamic functional connection graph.

[0062] Preferably, the pharmacogenomic data and the brain function network are dynamically integrated to break through the limitation of the traditional sedation monitoring that ignores individual metabolic differences in the population pharmacokinetic model. By constructing a three-layer knowledge graph of genotype-metabolic phenotype-drug response parameter, individualized pharmacokinetic parameter analysis is realized to generate a virtual pharmacokinetic node embedded in a functional connection graph. A real-time interaction channel between the genetic characteristics of drug metabolism and brain network activity is established to overcome the prediction bias problem of traditional methods caused by not considering genotype differences. At the same time, by quantifying the dynamic relationship between brain regions and pharmacokinetic parameters, the drug-brain function interaction effect that cannot be captured by existing independent brain networks or independent pharmacokinetic models is solved to provide a cross-scale pharmacodynamic real-time prediction basis for individualized precision sedation.

[0063] S4, input the dynamic functional connection graph into the graph convolutional neural network, and combine the hypergraph adaptive operator with the adjacent convolution of the virtual pharmacokinetic node to generate a node-level feature tensor.

[0064] S4.1: Take each frequency band power feature node as a point set in the base space and take the virtual pharmacokinetic node as a point set in the fiber space to construct a fiber bundle geometry. Specifically, each base space point is associated with a fiber space point to form a fiber bundle geometry through point-to-point mapping.

[0065] S4.2: Perform an adjacent convolution operation on the base space and the fiber space according to the Riemann curvature of the fiber bundle geometry to generate fiber bundle connection parameters. Specifically, based on the Riemann curvature properties of the fiber bundle geometry, the Riemann curvature properties are input into a voltage-to-resistance conversion circuit. The resistance values ​​are generated through an operational amplifier and a logarithmic circuit. The resistance values ​​are arranged in spatial coordinates to form a resistance matrix. The resistance matrix is ​​physically equivalent to a continuous convolution kernel. The base space is input into a voltage follower circuit, which outputs a scalar field quantity. The fiber space is input into a multi-channel digital-to-analog converter, which outputs a vector field quantity. Tensor product integration is performed using continuous convolution kernels, scalar field quantities, and vector field quantities to output fiber bundle connection parameters.

[0066] S4.3: Optimize fiber bundle connection parameters using Chern-Wey canonical theory and perform Hodge decomposition to generate node-level feature tensors.

[0067] Specifically, the fiber bundle connection parameters are input into the gauge transformation operation, and the spatial distribution of the fiber bundle connection parameters is adjusted according to the principle of minimum action to generate the gauge-invariant optimized fiber bundle connection parameters. The optimized fiber bundle connection parameters are decomposed into appropriate form components, co-appropriate form components, and harmonic form components by applying the Hodge decomposition theorem. The harmonic form components are then extracted as node-level feature tensors.

[0068] S5. Perform attention pooling on the node-level feature tensor, calculate the sedation depth index, and generate a pain interference factor based on the connection edge weights of the dynamic functional connection graph.

[0069] S5.1: Perform cross-modal attention pooling on the node-level feature tensor to generate a fused feature vector; Specifically, the mutual information intensity of different modal feature components in the node-level feature tensor is calculated, and attention weight coefficients are assigned based on the mutual information intensity. The attention weight coefficients are combined with the corresponding feature components of the node-level feature tensor, and then summed and aggregated along the feature dimension to generate a fused feature vector.

[0070] S5.2: Specifically, the fused feature vectors are input into the hierarchical regression network to calculate the sedation depth index, expressed as: ; In the formula, Indicates the depth of sedation index. Indicates the scaling factor. Represents the hyperbolic tangent function. This represents the dimension index of the fused feature vector. This represents the total dimension of the fused feature vector. Indicates to arrive Perform a traversal and summation. Indicates the first The feature weight values ​​corresponding to the dimension. The fused feature vector represents the first... Dimensional feature values Indicates a constant offset. This represents the translation coefficient.

[0071] It should be noted that, , , , , and All are dimensionless, and the final output is... Since it is dimensionless, we maintain dimensional consistency. The scaling factor is used to linearly map the hyperbolic tangent function output range [-1,1] to the clinically required sedation index range [-50,50], with the example value being a fixed constant of 50; the translation factor is used to translate the mapped value [-50,50] to the standard sedation index range [0,100], with the example value being a fixed constant of 50. Example: If the sedation depth index is [0,30], the patient is conscious; if the sedation depth index is [31,60], the patient is mildly sedated; if the sedation depth index is [61,85], the patient is moderately sedated; if the sedation depth index is [86,100], the patient is deeply sedated.

[0072] S5.3: Specifically, based on the edge weights of the dynamic functional connection graph, the entropy value of the dynamic graph is calculated, and the expression is: ; In the formula, Indicates a point in time The entropy value of the dynamic graph at that time. This represents the total number of directed edges in the dynamic function connection graph. Indicates the index of the directed connection edge. This indicates that for directed connection edges from arrive Perform a traversal and summation. Indicates the first A directed connection edge at time point The weighting percentage of time, Indicates to Perform logarithmic operations to base 2.

[0073] It should be noted that, Dimensionless It is dimensionless, and the final output is... It is dimensionless, but we maintain dimensional consistency.

[0074] S5.4: Perform multi-scale wavelet analysis on the entropy value of the dynamic graph to generate a pain interference factor.

[0075] Specifically, a dynamic graph entropy value sequence is formed according to dynamic graph entropy values at multiple different time points, a discrete wavelet transform is used to perform multi-scale decomposition on the dynamic graph entropy value sequence to generate an approximate coefficient sequence set and a detail coefficient sequence set, a high-frequency detail coefficient subsequence corresponding to different scales is separated from the detail coefficient sequence set through signal pattern recognition, and a root mean square value of the high-frequency detail coefficient subsequence is calculated as a pain interference factor.

[0076] S6. When the pain interference factor exceeds a preset pain threshold and the sedation depth index meets a preset condition, a sedation optimization instruction is generated, and an infusion rate adjustment instruction is generated when the virtual pharmacokinetics node displays abnormal metabolism.

[0077] S6.1: When the pain interference factor exceeds the preset pain threshold and the sedation depth index is lower than a first sedation threshold, a sedation depth improvement instruction is generated. It should be noted that the preset pain threshold is set based on the pain intensity in the pre-stored clinical pain assessment scale that requires drug intervention, and an example value is 0.7; the first sedation threshold is set according to the insufficient sedation warning line of the sedation depth monitoring index (such as AAI index), and an example value is 50 (corresponding to AAI index ≤ 50, which requires to improve sedation).

[0078] Specifically, the size relationship between the pain interference factor value and the preset pain threshold is compared, the size relationship between the sedation depth index value and the first sedation threshold is compared, if the pain interference factor value exceeds the preset pain threshold and the sedation depth index value is lower than the first sedation threshold, a sedation depth improvement instruction is generated, if the pain interference factor value does not exceed the preset pain threshold and the sedation depth index value is higher than the first sedation threshold, it indicates that the current pain response does not reach the severity that requires intervention and the sedation level meets the minimum requirement, and no instruction is generated.

[0079] S6.2: When the pain interference factor exceeds the preset pain threshold and the sedation depth index is between the first and second sedation thresholds, an analgesic drug addition instruction is generated. It should be noted that the second sedation threshold is set according to the excessive sedation warning line of the sedation depth monitoring index, and an example value is 30 (corresponding to AAI index ≤ 30, which requires to stop deepening sedation).

[0080] Specifically, it is confirmed whether the pain interference factor exceeds the preset pain threshold and whether the sedation depth index is between the first and second sedation thresholds, when both conditions are met, the analgesic drug addition instruction generation program is activated, and the analgesic drug addition instruction is output, if both conditions cannot be met, the current sedation and analgesia remain unchanged.

[0081] S6.3: The clearance rate parameter value is analyzed from the virtual pharmacokinetics node, and when the clearance rate parameter value exceeds the tolerance error, it is determined that the metabolism is abnormal. Specifically, the clearance parameter value is extracted from the virtual pharmacokinetic node, and a pre-stored standard clearance reference value is obtained; An absolute difference value between the clearance parameter value and the standard clearance reference value is calculated, and when the absolute difference value exceeds a tolerance error, it is determined that the virtual pharmacokinetic node has a metabolic abnormal state.

[0082] It should be noted that the tolerance error is set based on the range of individualized differences allowed in clinical pharmacokinetic parameter monitoring, and an example value is 10% of the standard clearance reference value (for example, when the standard clearance is 30 L / h, the tolerance error threshold = 3 L / h).

[0083] S6.4: Adjust the infusion rate amount according to the severity of the metabolic abnormality, and generate an infusion rate adjustment instruction.

[0084] It should be noted that the infusion rate amount refers to the change value of the unit time administration amount of intravenous sedative drugs adjusted according to the severity of the metabolic abnormality, expressed in percentage change (for example, -5% means a 5% reduction in hourly administration amount).

[0085] Specifically, the absolute difference value is matched with a pre-set severity level classification (for example, mild abnormality corresponds to an absolute difference value less than 10%, moderate abnormality corresponds to an absolute difference value greater than or equal to 10% and less than 20%, and severe abnormality corresponds to an absolute difference value greater than or equal to 20%); According to the severity level, a pre-set infusion rate adjustment amount mapping table (for example, mild abnormality corresponds to a rate reduction of 5%, moderate abnormality corresponds to a rate reduction of 10%, and severe abnormality corresponds to a rate reduction of 20%) is looked up to generate an infusion rate adjustment instruction.

[0086] It should be noted that the pre-set infusion rate adjustment amount mapping table establishes a quantitative correspondence between the severity level classification of metabolic abnormality and the infusion rate adjustment amount through multi-center clinical research data.

[0087] In summary, the present application realizes precise quantification of the pain-sedation coupling effect by using the time-varying intensity of the skin conductance response signal and the corrugator muscle electrical signal as the dynamic weight in the dynamic functional connectivity graph construction link. The sympathetic nerve activation intensity and the body movement response intensity are fused into a dynamic edge weight, the connection strength between the nodes in the electroencephalogram frequency band is modulated, and the frequency domain disturbance characteristics of the pain interference on the brain function network are captured. The specificity of the pain-related brain network disorder can be improved, the false positive rate can be reduced, the dynamic weight can change nonlinearly with the pain stimulation intensity, and the pain contribution in the frequency band power attenuation can be effectively separated, providing high signal-to-noise ratio for sedation optimization decision.

[0088] It should be noted that the above examples are only used to illustrate the technical solutions of the present application but not limit the present application. Although the present application is described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present application can be modified or equivalently replaced, without departing from the spirit and scope of the technical solutions of the present application, which should be covered in the scope of the claims of the present application.

Claims

1. A sedation depth monitoring method based on graph convolutional neural network, characterized in that: The application relates to a method for monitoring and optimizing the sedation and pain management of a patient, comprising the following steps: Collecting multi-source biomarkers of a patient, filtering and removing artifacts from electroencephalogram signals to generate preprocessed electroencephalogram data; Extracting multiple frequency band power features from the preprocessed electroencephalogram data as nodes, constructing connecting edges between the nodes by taking the time-varying intensity of a skin conductance response signal and a corrugator electromyogram signal as dynamic weights, and generating a functional connection graph; Analyzing drug metabolism sites in genotype data, querying a constructed drug metabolism knowledge graph to generate a pharmacokinetic parameter vector, and converting the pharmacokinetic parameter vector into a virtual pharmacokinetic node embedded in the functional connection graph to generate a dynamic functional connection graph; Inputting the dynamic functional connection graph into a graph convolutional neural network, combining an adjacency convolution of the virtual pharmacokinetic node with a hypergraph adaptive operator to generate a node-level feature tensor; Performing attention pooling processing on the node-level feature tensor, calculating a sedation depth index, and generating a pain interference factor according to the connecting edge weights of the dynamic functional connection graph; When the pain interference factor exceeds a preset pain threshold and the sedation depth index meets a preset condition, generating a sedation optimization instruction, and when the virtual pharmacokinetic node shows abnormal metabolism, generating an infusion rate adjustment instruction.

2. The sedation deepness monitoring method based on graph convolution neural network according to claim 1, wherein: The multi-source biomarkers include electroencephalogram signals, skin conductance response signals, corrugator electromyogram signals and genotype data.

3. The sedation deepness monitoring method based on graph convolutional neural network according to claim 1, wherein: The preprocessed electroencephalogram data is generated in the following specific steps, Signal amplification is performed on the electroencephalogram signals to generate amplified electroencephalogram signals, first band-pass filtering is performed on the amplified electroencephalogram signals to generate wideband filtered signals, and second narrowband filtering is performed synchronously to generate sedation characteristic frequency band signals; The wideband filtered signals are subjected to empirical mode decomposition to separate high-frequency artifact components; Coherence analysis is performed on the sedation characteristic frequency band signals and the high-frequency artifact components to remove artifact components irrelevant to sedation characteristics, and artifact suppression signals are generated; Multi-scale wavelet decomposition denoising is performed on the artifact suppression signals, and the artifact suppression signals are reconstructed into denoised artifact suppression signals, which are input into a sparse autoencoder network to generate preprocessed electroencephalogram data.

4. The sedation deepness monitoring method based on graph convolutional neural network according to claim 1, wherein: The connecting edges between the nodes are constructed to generate the functional connection graph in the following specific steps, First, second, third and fourth frequency band power features are extracted from the preprocessed electroencephalogram data as nodes; The skin conductance response signals are subjected to Hilbert-Huang transform decomposition to extract energy entropy values, and the corrugator electromyogram signals are subjected to Teager energy analysis to extract transient energy peak values; dynamic weight values are generated by matching the time sequence energy fluctuations of the energy entropy values and the transient energy peak values; The time delay correlation intensity between each node in each frequency band power feature node is calculated, and the connecting edge weights between each node are modulated by the dynamic weight values to form directed connecting edges between the nodes; The frequency band power feature nodes and the directed connecting edges between the nodes are taken as connecting relationships to generate the functional connection graph.

5. The sedation deepness monitoring method based on graph convolutional neural network according to claim 1, wherein: The drug metabolism knowledge graph is constructed by calling pre-stored pharmacogenomic data and clinical pharmacokinetic data, mapping the corresponding relationship among genotype-metabolic phenotype-drug response parameters into structured rules.

6. The graph convolutional neural network-based sedation depth monitoring method of claim 1, wherein: The pharmacokinetic parameter vector is generated by querying the constructed drug metabolism knowledge graph in the following specific steps, The drug metabolism related sites in the genotype data are analyzed, a site embedding vector is generated, and a genotype node in the drug metabolism knowledge graph is converted into a node embedding vector; The cosine similarity of the site embedding vector and the node embedding vector is calculated, and when the cosine similarity exceeds a preset similarity threshold, the genotype node and the associated metabolic phenotype node and drug response parameter node are activated to form an activated subgraph; The drug response parameter nodes in the activated subgraph are aggregated to generate a pharmacokinetic parameter vector.

7. The graph convolutional neural network-based sedation depth monitoring method of claim 1, wherein: The dynamic functional connectivity graph is generated, and the specific steps are as follows, The pharmacokinetic parameter vector is converted into a pharmacokinetic parameter sequence, each frequency band power feature node is converted into an electroencephalogram frequency band power sequence, and the correlation coefficient of the pharmacokinetic parameter sequence and the electroencephalogram frequency band power sequence is calculated to generate the interaction strength between the brain region and the drug metabolism; According to the interaction strength between the brain region and the drug metabolism, all interaction strengths are aggregated to generate a virtual pharmacokinetic node and embed it in the functional connectivity graph to generate a dynamic functional connectivity graph.

8. The graph convolutional neural network-based sedation depth monitoring method of claim 1, wherein: The node-level feature tensor is generated, and the specific steps are as follows, Each frequency band power feature node is taken as a base space, and the virtual pharmacokinetic node is taken as a fiber space to construct a fiber bundle geometry; According to the Riemann curvature of the fiber bundle geometry, an adjacency convolution operation is performed on the base space and the fiber space to generate fiber bundle connection parameters; The fiber bundle connection parameters are optimized by the Chen-Weinberg gauge theory, and the Hodge decomposition is performed to generate a node-level feature tensor.

9. The graph convolutional neural network-based sedation depth monitoring method of claim 1, wherein: The pain interference factor is generated, and the specific steps are as follows, The node-level feature tensor is processed by cross-modal attention pooling to generate a fusion feature vector; The fusion feature vector is input into a hierarchical regression network to calculate a sedation depth index; According to the connection edge weight of the dynamic functional connectivity graph, a dynamic graph entropy value is calculated; The dynamic graph entropy value is subjected to multi-scale wavelet analysis to generate a pain interference factor.

10. The graph convolutional neural network-based sedation depth monitoring method of claim 1, wherein: The infusion rate adjustment instruction is generated, and the specific steps are as follows, When the pain interference factor exceeds a preset pain threshold and the sedation depth index is lower than a first sedation threshold, a sedation depth improvement instruction is generated; When the pain interference factor exceeds the preset pain threshold and the sedation depth index is between the first and second sedation thresholds, an analgesic drug addition instruction is generated; The clearance rate parameter value is analyzed from the virtual pharmacokinetic node, and when the clearance rate parameter value exceeds a tolerance error, it is determined that the metabolism is abnormal; According to the severity of the metabolic abnormality, the infusion rate is adjusted to generate an infusion rate adjustment instruction.

Images