A high-precision anesthesia state monitoring method, system and device

Abstract

The present invention discloses a high-precision anesthesia state monitoring method, system and device, which belongs to the field of electroencephalogram signal processing and medical monitoring technology, in order to solve the influence of noise interference in electroencephalogram signals on anesthesia state monitoring. The monitoring method includes: step 1: collecting electroencephalogram signals of patients in anesthesia state; step 2: preliminarily filtering and amplitude normalizing the electroencephalogram signals to obtain new electroencephalogram signals; step 3: using the hierarchical frequency decomposition method to decompose the new electroencephalogram signals to obtain several modal functions; step 4: extracting multidimensional features from each modal function; step 5: fusing and screening the extracted multidimensional features to form a feature set; step 6: using the feature set to train the machine learning model, and using the trained machine learning model to predict the patient's anesthesia state in real time, and output the anesthesia stage classification result. The present invention can provide decision support for doctors based on real-time monitoring data and reduce the anesthesia risk during surgery.

Description

Technical Field

[0001] The present invention belongs to the technical field of electroencephalogram signal processing and medical monitoring, and in particular relates to a high-precision anesthesia state monitoring method, system and device. Background Art

[0002] In surgical operations, accurate monitoring of the depth of anesthesia is one of the key factors to ensure the success of the operation and the safety of the patient. If the anesthesia is too deep, it may lead to delayed postoperative awakening, postoperative cognitive dysfunction, and even increase the risk of intraoperative and postoperative complications; while anesthesia that is too shallow may cause the patient to regain consciousness during the operation, causing painful intraoperative awareness and severe psychological trauma. Therefore, it is of great clinical significance to accurately monitor the patient's anesthesia state and adjust the dosage of anesthetic drugs in time. Electroencephalogram (EEG), as the main tool for monitoring the depth of anesthesia, can directly reflect the effect of anesthetic drugs on the central nervous system. However, there are still certain problems in the detection of anesthesia depth through EEG.

[0003] On the one hand, EEG signals are often affected by a variety of noises and interferences, such as baseline drift, electromyographic noise, power frequency interference, and equipment noise. The baseline drift phenomenon caused by common problems such as breathing, changes in body position, and poor electrode contact is particularly prominent. These noises and interferences not only mask the characteristics of EEG activity related to the anesthetic state, but may also lead to misjudgment of the anesthetic state. Therefore, how to effectively remove noise and interference in EEG signals, especially the baseline drift problem, has become a major technical problem in current anesthetic state monitoring. Existing EEG signal processing methods such as bandpass filtering and empirical mode decomposition can remove some noise and interference to a certain extent, but they show great limitations in dealing with complex low-frequency baseline drift and various physiological noises. These traditional methods may lose some valuable anesthetic state-related signals during the denoising process, thereby affecting the accuracy of anesthesia depth assessment.

[0004] On the other hand, traditional feature extraction methods based on EEG signals, such as calculating the mean, standard deviation, power spectral density, etc., cannot fully tap the complexity and diversity of information. It is still a challenge to extract representative features from the decomposed multi-level modalities and use these features to build an effective machine learning model for real-time prediction of anesthesia status. In addition, in the process of feature selection and fusion, how to balance the contribution of each modality feature and optimize the generalization ability of the model is also a major problem in current research. Summary of the invention

[0005] The purpose of the present invention is to provide a high-precision anesthesia state monitoring method, system and device, aiming to solve the problems of insufficient accuracy and low automation level of anesthesia state monitoring in the prior art. The EEG signal is processed by a hierarchical frequency decomposition method, combined with advanced feature extraction and fusion algorithms, to construct a complete real-time monitoring solution for anesthesia state. By decomposing the EEG signal layer by layer using the hierarchical frequency decomposition method, the baseline drift and other noise interference in the EEG signal can be effectively removed, and multi-dimensional features can be extracted from the decomposed modal function. Subsequently, innovative feature fusion and selection methods, such as nonlinear feature mapping, sparse coding, adaptive weighted feature fusion, etc., are used to optimize and screen the extracted features, thereby ensuring the system's high-precision prediction of the anesthesia state.

[0006] The high-precision anesthesia status monitoring method, system and device proposed in the present invention not only fill the deficiencies of the existing technology in EEG signal noise reduction and feature extraction, but also realize real-time automatic monitoring of the anesthesia status by introducing a machine learning model. It can provide anesthesiologists with reliable decision support based on real-time monitoring data, reduce the anesthesia risk during surgery, and improve the safety and effectiveness of surgery.

[0007] In a first aspect, the present invention provides a high-precision anesthesia state monitoring method, the monitoring method comprising the following steps: Step 1: Collect EEG signals of the patient under anesthesia; Step 2: Use high-pass filter and low-pass filter to preliminarily filter the EEG signal, and normalize the amplitude of the EEG signal to obtain a new EEG signal; Step 3: Use the hierarchical frequency decomposition method to perform variational modal decomposition on the new EEG signal to obtain several modal functions; Step 4: Extract multidimensional features from each modal function; Step 5: Fuse and filter the extracted multidimensional features to form a feature set; Step 6: Use the feature set to train the machine learning model, and use the trained machine learning model to predict the patient's anesthesia status in real time and output the anesthesia stage classification results.

[0008] Furthermore, the specific operation of preliminary filtering of the EEG signal in step 2 is as follows: Step 201: using a 0.5 Hz high-pass filter to remove the baseline drift signal in the EEG signal; Step 202: Use a 50 Hz low-pass filter to remove high-frequency interference in the EEG signal.

[0009] Furthermore, the specific process of performing variational modal decomposition of the new EEG signal using the hierarchical frequency decomposition method in step 3 is as follows: Step 301: Perform the first-layer decomposition on the new EEG signal to obtain k First-order mode functions; Step 302: k The first-order mode function is decomposed into the second level The second-order mode function, l It means that each first-order mode function is decomposed into Second-order mode functions, update The frequency distribution of the second-order mode function and calculate The frequency and amplitude characteristics of the second-order mode function; Step 303: Use the frequency of the first-order mode function as the reference frequency for baseline drift signal screening. k- Select the first-order mode function containing the reference frequency p The first-order low-frequency mode function is selected p The first-order low-frequency mode function The third-order modal function is decomposed into three layers for each second-order modal function, the frequency distribution of the third-order modal function is updated, and the frequency characteristics and amplitude characteristics of the third-order modal function are calculated; Step 304: Introduce the baseline drift signal standard frequency, from A second-order low-frequency modal function containing the standard frequency of the baseline drift signal is selected from the second-order modal functions, and the third-order modal function under the selected second-order low-frequency modal function is decomposed into a fourth-order modal function, the frequency distribution of the fourth-order modal function is updated, and the frequency characteristics and amplitude characteristics of the fourth-order modal function are calculated; Step 305: Continue to select the third-order low-frequency modal function containing the standard frequency of the baseline drift signal from the third-order modal function, and perform the next-level decomposition on the fourth-order modal function under the selected third-order low-frequency modal function, and repeat step 304 until n The first low-frequency mode function n+ 1 layer decomposition n+ If the first-order mode function does not contain the standard frequency of the baseline drift signal, stop the decomposition and update n+ The frequency distribution of the 1st order mode function and calculate n+ Frequency and amplitude characteristics of the first-order mode function; Step 306: Select the first-order modal function, the first-order low-frequency modal function, the second-order low-frequency modal function, the third-order low-frequency modal function, and n The first-order low-frequency mode function is used as a candidate mode for baseline drift signal reconstruction; Step 307: Reconstruct the baseline drift signal using the candidate mode, using the following formula: ; in, represents the signal time, represents the reconstructed baseline drift signal, is the low-frequency mode function of each order after hierarchical decomposition, express The total number of low-frequency mode functions; Step 308: Subtract the reconstructed baseline drift signal from the new EEG signal to obtain a purified EEG signal, as shown in the following formula: ; in, Represents a new EEG signal. In the purified signal, baseline drift and other low-frequency interference have been removed, retaining the patient's actual EEG activity.

[0010] Furthermore, in step 301, the new EEG signal is decomposed into k independent first-order mode functions, each of which is compactly distributed near its center frequency, and the optimal and ; (1); in, Represents signal time, which is used to describe the changes of the signal in the time domain. Indicates The first-order mode function is in time The signal value on For the A first-order mode function, For the The center frequency of the first-order mode function, is the impulse function, used as the convolution kernel, is an ordinal unit used to represent the complex form of frequency components. is the convolution operator, used for convolution of the signal and the impulse function. is a derivative operation on time, which represents the bandwidth minimization constraint; and the center frequency of each first-order mode function obtained by the first-level decomposition is equally spaced.

[0011] Furthermore, in steps 302-305, according to formula (1), The order mode functions are divided into multiple independent order mode function, and The frequency distribution of the order mode function is updated using the following formula: (2); in, represents the signal frequency, is the number of iterations, For the The modal function is in The center frequency at the iteration, Corresponding to The Lagrange multiplier of the iteration is used as the optimization constraint term. It is The modal function is in The spectrum update value of the iteration, is the frequency domain representation of the new EEG signal, is the penalty coefficient, represents the modal variable, and Respectively represent The modal function is in The spectrum update value of the iteration and the The modal function is in The spectrum update value of the iteration; but The center frequency of the first mode function is calculated using the following formula: (3); The frequency characteristic calculated in steps 302-305 is the center frequency, which is obtained by formula (3); Each iterative decomposition will update the Lagrange multiplier, and the update formula of the Lagrange multiplier is: (4); in, is the updated value of the Lagrange multiplier in the frequency domain update process, is the time step parameter in the dual ascent process, is the number of iterations.

[0012] Furthermore, the modal functions in step 4 include first-order modal functions, second-order modal functions, third-order modal functions, fourth-order modal functions, and Order mode function; Multidimensional features include: The center frequency and bandwidth of each modal function are used to reflect the main frequency characteristics of the new EEG signal; The frequency domain energy distribution and time-frequency energy matrix of each modal function are used to quantify the time-frequency characteristics of the new EEG signal; The higher-order statistics of each modal function include kurtosis and skewness, which are used to assess the nonlinearity and complexity of new EEG signals; The coherence and frequency overlap ratio between modal functions are used to describe the information correlation and independence between modal functions.

[0013] Furthermore, the frequency of the first-order mode function of the first digit is 1 Hz, and the standard frequency of the baseline drift signal is 0.5 Hz.

[0014] Furthermore, the specific process of fusing and screening the extracted multi-dimensional features in step 5 is as follows: Step 501: Mapping the extracted multi-dimensional features to a high-dimensional feature space using a non-linear feature mapping technique to capture the complex non-linear relationship between the multi-dimensional features; Step 502: sparsely represent the multi-dimensional features in the high-dimensional feature space by using a sparse coding algorithm; Step 503: using an adaptive feature weighting method, dynamically learning the contribution of each multidimensional feature in the feature set under different anesthesia states, and adjusting the weight of each multidimensional feature; Step 504: introducing a multi-view feature learning method, constructing feature subspaces for the multi-dimensional features extracted from each modal function, and jointly training these subspaces through a co-training algorithm to mine implicit information between different modal functions; Step 505: After processing steps 501-504, a fused feature set is obtained, and an integrated feature selection algorithm is used to screen the fused feature set.

[0015] In a second aspect, the present invention further provides a high-precision anesthesia state monitoring system, which is implemented using the above-mentioned monitoring method and includes: The EEG signal acquisition module collects the patient's EEG signals during surgery through non-invasive electrodes; The signal processing module first performs preliminary filtering on the EEG signal to obtain a new EEG signal, and then uses the hierarchical frequency decomposition method to perform variational mode decomposition on the new EEG signal to obtain several mode functions, and variational mode decomposition of the baseline drift signal; The feature extraction and fusion module extracts the multidimensional features of each modal function, and fuses and filters the multidimensional features to obtain a feature set; The machine learning prediction module uses the feature set to train the machine learning model, and uses the trained machine learning model to predict the patient's anesthesia status in real time and output the anesthesia stage classification result; The visualization interface module displays the classification results of anesthesia stages.

[0016] In a third aspect, the present invention also provides a high-precision anesthesia status monitoring device, which includes a memory storing a computer program and a processor for executing the computer program. When the computer program is executed by the processor, the steps of the above-mentioned monitoring method are implemented.

[0017] Compared with the prior art, the present invention has the following advantages: 1. The core innovation of the present invention lies in the noise reduction processing of EEG signals. First, the collected EEG signals are subjected to multi-level hierarchical decomposition through the hierarchical frequency decomposition method, and the complex EEG signals are decomposed into modal functions (IMFs) of different frequency components, which can adaptively separate low-frequency baseline drift and other noise components while retaining high-frequency effective EEG activity signals, thereby providing a clearer and more accurate signal basis for subsequent anesthesia status monitoring. Compared with traditional noise reduction methods, hierarchical variational modal decomposition shows significant advantages in removing low-frequency baseline drift and electromyographic noise, ensuring that key information in the signal will not be lost.

[0018] 2. The present invention also proposes a complete set of feature extraction and fusion methods, aiming to extract a feature set with diagnostic value from the net EEG signal after noise reduction. In the process of hierarchical variational modal decomposition, the extracted features include the center frequency, bandwidth, frequency domain energy distribution, time-frequency energy matrix, coherence between modes, frequency overlap ratio and high-order statistics (such as kurtosis and skewness) of each mode. These features can accurately reflect the dynamic changes of the patient's EEG activity under anesthesia, especially the feature changes related to the depth of anesthesia. The extracted features not only cover the frequency domain characteristics, but also capture the volatility and complexity of the signal in different stages of anesthesia through time domain analysis.

[0019] 3. The present invention adopts a combination of multiple methods in the feature fusion and screening part, which greatly improves the monitoring accuracy and adaptability. First, the original feature set is mapped to a high-dimensional space by nonlinear feature mapping technology to ensure that the implicit relationship between the features is captured. Then, the sparse coding algorithm is used to decompose and sparsely represent the features, effectively remove redundant features, and improve the computational efficiency of the system. In addition, the adaptive weighted feature fusion technology is adopted to dynamically adjust the feature weights through an algorithm based on a gradient boosting decision tree to ensure that each feature plays the best role in different anesthesia stages. In addition, the present invention adopts a multi-perspective feature learning method, and regards the multi-frequency modality obtained by decomposition as multiple different perspectives for feature learning. Through the co-training algorithm, the implicit information between different modalities can be further mined, and the independence and consistency of the features can be ensured. The introduction of multi-perspective learning not only improves the effect of feature fusion, but also enhances the adaptability of the system in different surgical scenarios. Finally, the feature set is screened by an integrated feature selection algorithm, combined with methods such as recursive feature elimination and Lasso regression, and finally retains the features with the most diagnostic value, reduces redundancy, and improves the generalization ability of the model.

[0020] 4. The present invention selects the best model suitable for the classification of anesthesia status through a variety of algorithm comparison experiments, including support vector machine, random forest and neural network. The support vector machine ensures high-precision classification of different anesthesia stages by maximizing the construction of classification boundaries; the random forest enhances the stability and noise resistance of the model by integrating multiple weak classifiers; the neural network captures the complex nonlinear changes of EEG signals during anesthesia through the deep learning structure of multi-layer perceptron. Combined with cross-validation technology, the optimal model was finally selected to achieve real-time prediction of four different stages of awake, light anesthesia, moderate anesthesia and deep anesthesia.

[0021] 5. The present invention significantly improves the accuracy and automation level of anesthesia status monitoring. During surgery, it can effectively reduce the risk of too shallow or too deep anesthesia, ensuring the safety of the surgery and the quality of postoperative recovery of the patient. Through the present invention, anesthesiologists can make more accurate anesthesia management decisions based on real-time monitoring data, improving the overall surgical effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 The figure is a flow chart of a high-precision anesthesia state monitoring method.

[0023] Figure 2 This is a flow chart for performing variational mode decomposition of new EEG signals using the hierarchical frequency decomposition method. DETAILED DESCRIPTION

[0024] The preferred embodiments of the present invention are described in detail below in conjunction with the accompanying drawings so that the advantages and features of the present invention can be more easily understood by those skilled in the art, thereby making a clearer and more definite definition of the protection scope of the present invention.

[0025] Embodiment: See also Figure 1 This embodiment specifically provides a high-precision anesthesia state monitoring method, which includes the following steps: Step 1: Collect EEG signals of the patient under anesthesia: A non-invasive method is used to collect the patient's EEG signals under anesthesia, that is, electrodes are placed on the patient's scalp to record the signals. This embodiment uses a multi-electrode system to collect EEG signals. The electrodes are fixed by standard electrode sheets, and the sampling frequency of collecting EEG data is 256Hz. The system consists of four electrode sheets, and the electrodes are placed on the patient's forehead area. The collected EEG signals cover the patient's complete EEG activity data during the entire operation.

[0026] At this point, the collection of EEG signals of the patient under anesthesia during the operation has been completed.

[0027] Step 2: Use high-pass filter and low-pass filter to preliminarily filter the EEG signal, and normalize the amplitude of the EEG signal to obtain a new EEG signal: Baseline drift is a very low-frequency interference. The amplitude of the EEG signal affected by baseline drift will increase. Baseline drift is mainly caused by the patient's breathing, movement, and movement of the electrode. The patient's breathing will cause the patient's body to fluctuate to a certain extent, and the baseline drift caused by this situation is unavoidable. In addition, due to surgical needs, the doctor may need to move the patient during the operation, which will also cause a large amount of baseline drift. Secondly, due to the long operation time and the complex operating room environment, the patient's sweating, the reduction of the viscosity of the electrode paste, and the accidental contact with the wire may cause the electrode for collecting EEG signals to change position to a certain extent during the operation, thereby causing baseline drift.

[0028] In the process of EEG signal acquisition, in addition to baseline drift, it also includes equipment noise and other high-frequency signal interference. Therefore, before analyzing the signal, preliminary noise reduction processing is required first. Since baseline drift is a common low-frequency noise, its frequency is usually less than 0.5 Hz. Therefore, a 0.5Hz high-pass filter is used in this embodiment to remove the low-frequency component. This step can significantly reduce the low-frequency baseline drift interference introduced by the patient's breathing, movement and electrode movement. In addition, equipment noise and high-frequency electrical signals may also affect the validity of EEG signals. A 50Hz low-pass filter can remove high-frequency interference, including power line interference (50Hz power frequency noise) and equipment vibration signals. The combination of high-pass and low-pass filtering ensures that the filtered signal is mainly retained in the frequency band of 0.5Hz to 50Hz, which is also the main frequency band range of EEG activity.

[0029] In order to further reduce the amplitude changes caused by electrode contact differences or other environmental factors during signal acquisition, the signal is normalized. The amplitude normalization process adjusts the signal amplitudes of different channels to the same range, ensuring the comparability of signals between different electrodes. The normalized signal amplitude is standardized to the 0-1 interval, ensuring the accuracy of the subsequent decomposition process.

[0030] Step 3: Use the hierarchical frequency decomposition method to perform variational modal decomposition on the new EEG signal to obtain several modal functions: Baseline drift is a low-frequency interference that can cause a significant increase in the amplitude of the EEG signal. This interference mainly comes from the patient's breathing, body movement, and displacement of the electrode. Since the patient's breathing process causes the body to rise and fall slightly, an inevitable baseline drift will occur. At the same time, during the operation, if the doctor needs to operate or move the patient, a larger drift may be caused. In addition, the long duration of the operation and the complex environment will also aggravate this problem. For example, the patient may sweat during the operation, resulting in a decrease in the adhesion of the electrode paste, or accidentally touch the wire. These factors will cause slight changes in the position of the electrode, and ultimately cause the baseline drift in the EEG signal to increase. This embodiment extracts the baseline drift signal through multi-layer VMD decomposition (hierarchical frequency decomposition method).

[0031] Based on the new EEG signal obtained by the preliminary denoising and normalization in the previous step, VMD decomposition is used to obtain Mode The decomposition criteria are obtained by The signal is fully reconstructed by summing the modes while minimizing the sum of the bandwidths of all modes. Each mode is compactly distributed around its center frequency. Nearby, it can be represented by frequency and amplitude modulation properties. The optimal and : (1); in, Represents signal time and is used to describe the changes of the signal in the time domain. Indicates The first-order mode function is in time The signal value on For the A first-order mode function represents the component within a certain frequency range of the signal. for Abbreviation of . For the The center frequency of the first-order mode function, is an impulse function, used as a convolution kernel. It is an ordinal unit used to represent the complex form of frequency components. It is a convolution operator used for convolution of a signal and an impulse function. is the derivative operation with respect to time, representing the bandwidth minimization constraint.

[0032] The above optimization problem is solved by Lagrange multipliers, and the saddle point is found through multiple iterations of the alternating direction multiplier method. The process mainly includes: 1) Modal update: The update process of each mode in the frequency domain is implemented by the following formula: (2); in, represents the signal frequency, is the number of iterations, For the The modal function is in The center frequency at the iteration, Corresponding to The Lagrange multiplier of the iteration is used as the optimization constraint term. It is The modal function is in The spectrum update value of the iteration, is the frequency domain representation of the new EEG signal, is the penalty coefficient, Represents the modal variable, which is the commonly used representative index meaning, and Respectively represent The modal function is in The spectrum update value of the iteration and the The modal function is in The spectrum update value of the iteration; 2) Center frequency update: In order to ensure that the mode can be accurately decomposed into the correct frequency range, the center frequency of the mode is updated during each iteration. Need to be updated continuously. The update formula of center frequency is as follows: (3); 3) Lagrange multiplier update: In order to ensure the accuracy of signal reconstruction, VMD uses the Lagrange multiplier method to constrain the modal decomposition. The update formula of the Lagrange multiplier is as follows: (4); in, is the updated value of the Lagrange multiplier in the frequency domain update process, is the time step parameter in the dual ascent process, is the number of iterations; each iteration will update the Lagrange multiplier and gradually approach the optimal decomposition of the signal.

[0033] 4) The amplitude characteristics are expressed by the maximum amplitude and the root mean square amplitude, which are calculated using the following formulas: ; ; in, N represents the length of the time series, Indicates The modal function in time The signal value on .

[0034] A key step in VMD decomposition is to determine the number of modes. In the embodiment of the invention, based on the frequency characteristics of the EEG signal and multiple experimental results, five modes are finally selected, corresponding to EEG signal components in different frequency ranges. It determines the bandwidth of each mode and affects the accuracy of modal decomposition. The value is selected as 2000, which can ensure the reasonable distribution of frequency bandwidth between modalities, especially in the low frequency band, which has a better effect on the extraction of baseline drift. Another important parameter of VMD is the center frequency of the modality. After many experiments, the center frequency of the modality is set to 0.5Hz, 2Hz, 4Hz, 8Hz and 12Hz, corresponding to different EEG frequency ranges.

[0035] Through VMD decomposition, the different frequency components in the signal are separated into several modal functions, each of which represents a certain frequency band in the signal. Usually, the low-frequency mode contains a lot of baseline drift and other interference signals, while the high-frequency mode retains the effective EEG activity related to the patient's anesthesia state.

[0036] The low-frequency modes obtained by the initial variational mode decomposition are further decomposed into finer sub-modes; the modal function is hierarchically decomposed to further decompose each mode into sub-modes.

[0037] In the obtained modal function, the low-frequency mode also contains a lot of baseline drift and noise interference. Therefore, this embodiment further refines and decomposes these low-frequency modes, and decomposes the low-frequency modes into finer sub-modes through a hierarchical optimization process.

[0038] Hierarchical decomposition is a further decomposition of low-frequency modes to ensure that every detail of the low-frequency signal can be accurately extracted. Specifically, the low-frequency mode obtained by VMD is first decomposed twice, and the goal of each decomposition is to further refine the frequency components of the signal. The objective function of hierarchical decomposition is similar to that of the initial VMD decomposition, see formula (1). During each decomposition process, the center frequency and bandwidth parameters of the mode need to be readjusted. For the first step of the hierarchical decomposition, the center frequency of the mode is further refined from the original 0.5Hz to 0.2Hz and 1Hz, so that the changes in baseline drift can be accurately captured.

[0039] During the hierarchical decomposition process, the penalty coefficient needs to be further optimized to ensure that the frequency range of the signal can be further refined. During the hierarchical decomposition process, the bandwidth penalty coefficient is set to 500, which allows each sub-mode to cover a finer frequency range. Through the optimization of this parameter, the hierarchical decomposition process can peel off the baseline drift signal layer by layer, ensuring that the effective signal in the high-frequency mode is not affected.

[0040] Select the modes obtained from the first-level decomposition and the sub-modes with gradually decreasing frequencies after hierarchical decomposition; for these low-frequency modes, calculate their frequency distribution and amplitude characteristics, and select the mode with the lowest frequency and large amplitude as the main source of baseline drift; superimpose the selected low-frequency modes and reconstruct the baseline drift signal. The step flow chart is as follows Figure 2 shown.

[0041] The specific steps include: Step 301: Perform the first-layer decomposition on the new EEG signal to obtain k First-order mode functions; Step 302: k The first-order mode function is decomposed into the second level Second-order mode functions, update The frequency distribution of the second-order mode function and calculate The frequency and amplitude characteristics of the second-order mode function; Step 303: Use the frequency of the first-order mode function as the reference frequency for baseline drift signal screening. k- Select the modal function containing the reference frequency from 1 modal function p The low-frequency mode functions are selected p The low frequency mode function The third-order modal function is decomposed into three layers for each second-order modal function, the frequency distribution of the third-order modal function is updated, and the frequency characteristics and amplitude characteristics of the third-order modal function are calculated; Step 304: Introduce the baseline drift signal standard frequency, from A second-order low-frequency modal function containing the standard frequency of the baseline drift signal is selected from the second-order modal functions, and the third-order modal function under the selected second-order low-frequency modal function is decomposed into a fourth-order modal function, the frequency distribution of the fourth-order modal function is updated, and the frequency characteristics and amplitude characteristics of the fourth-order modal function are calculated; Step 305: Continue to select the third-order low-frequency mode function containing the standard frequency of the baseline drift signal from the third-order mode function, and perform the next-level decomposition on the fourth-order mode function under the selected third-order mode function, and repeat step 304 until n The first low-frequency mode function n+ 1 layer decomposition n+ If the first-order mode function does not contain the standard frequency of the baseline drift signal, stop the decomposition and update n+ The frequency distribution of the 1st order mode function and calculate n+ Frequency and amplitude characteristics of the first-order mode function; Step 306: Select the first-order modal function, the second-order low-frequency modal function, the third-order low-frequency modal function, the second-third-order low-frequency modal function, and the n The first-order low-frequency mode function is used as a candidate mode for baseline drift signal reconstruction; Step 307: Reconstruct the baseline drift signal using the candidate mode, using the following formula: ; in, represents the reconstructed baseline drift signal, is the low-frequency mode function of each order after hierarchical decomposition, M stands for The total number of low-frequency mode functions; by superimposing these low-frequency mode functions, a complete baseline drift signal can be finally obtained.

[0042] Step 308: Subtract the reconstructed baseline drift signal from the new EEG signal to obtain a purified EEG signal, as shown in the following formula: ; in, Represents a new EEG signal. In the purified signal, baseline drift and other low-frequency interference have been removed.

[0043] Based on the above steps, each low-frequency modal function will be further decomposed into several multi-order modal functions during the hierarchical decomposition process. As the hierarchical decomposition proceeds, the frequency components of each multi-order modal function will gradually decrease, and the amplitude characteristics will become clearer. This enables us to accurately identify which multi-order modal functions are the main components of baseline drift and which multi-order modal functions are other interference signals. After the hierarchical decomposition is completed, the frequency distribution and amplitude characteristics of each multi-order modal function need to be analyzed. Generally speaking, during long-term surgery, the amplitude of baseline drift tends to increase significantly due to the patient's breathing, body position changes, and electrode movement. Therefore, among the sub-modal modes after hierarchical decomposition, we will select those modes with larger amplitudes as candidate modes for baseline drift. At the same time, baseline drift is a very low-frequency signal, and its frequency is usually less than 1 Hz, and may even be less than 0.1 Hz. The purpose of hierarchical decomposition is to further strip out these low-frequency components so that we can accurately identify baseline drift.

[0044] To achieve this goal, the frequency and amplitude analysis method is used in this embodiment. Specifically, for each multi-order modal function after hierarchical decomposition, its frequency distribution and amplitude characteristics are calculated, and those modes with lower frequencies and larger amplitudes are selected as the main sources of baseline drift. The frequencies of these multi-order modal functions are usually lower than 1Hz, and the amplitudes account for more than 30% of the total signal amplitude. This optimization process ensures that the most representative baseline drift components can be extracted from the sub-modes after hierarchical decomposition, and can minimize the impact of other interfering signals on baseline drift extraction.

[0045] The reconstructed baseline drift signal represents the comprehensive reflection of all low-frequency components, mainly including low-frequency interference caused by patient breathing, body position changes, poor electrode contact, etc. These interference signals will continue to change during the operation, so the reconstructed baseline drift signal can dynamically reflect these changes.

[0046] The accuracy of baseline drift signal extraction is greatly improved through hierarchical decomposition and modal optimization. The multi-level processing of hierarchical decomposition ensures that low-frequency interference signals are gradually stripped away, while the optimized modal selection ensures that only those low-frequency modes that are highly correlated with baseline drift are retained.

[0047] The final reconstructed baseline drift signal can not only accurately reflect the baseline drift changes of the patient during the operation, but also effectively avoid the mixing of other interference signals. The successful implementation of this process makes it possible to extract effective EEG activities from EEG signals in the future.

[0048] Step 4: Extract multidimensional features from each modal function: In the feature extraction and calculation stage, we make full use of the intermediate calculation results and modal features in the decomposition process to construct a comprehensive and effective feature set. First, based on the modal functions obtained from each layer of decomposition, we extract the center frequency, bandwidth, and frequency domain energy distribution of each multi-order modal function. These features belong to frequency domain features. The center frequency is as mentioned above. The bandwidth and frequency domain energy distribution are both conventional methods in this field. The bandwidth will be introduced below and will not be repeated here. The python mne package contains tools for calculating frequency domain energy distribution, so its calculation will not be repeated here.

[0049] In addition to common frequency domain features, the present invention also uses the coherence between modal functions to evaluate the correlation between different modalities. The higher the coherence, the more information overlaps between modal functions, which may reflect similar physiological activities. The calculation formula for inter-modal coherence is as follows: (7); in, is modal and The mutual spectrum of and The modal and The power spectral density of the modal. Through coherence analysis, modes with high independence can be screened out to reduce feature redundancy. In addition, the frequency overlap ratio is used to quantify the degree of overlap of frequency components between different modes. The smaller the frequency overlap ratio, the better the frequency separation between the modes. The calculation formula for the frequency overlap ratio is: (8); in, and is the frequency range, and Mode and A lower overlap ratio means better frequency separation between the modal functions, which reduces the mutual influence between the modal functions.

[0050] The construction of the time-frequency energy matrix can help observe the dynamic changes of the time-frequency of EEG signals in different stages of anesthesia, and provide further basis for feature extraction. By constructing the time-frequency energy matrix, the time-frequency changes of EEG signals can be captured, which helps analyze the energy concentration trend in different stages of anesthesia. The time-frequency energy matrix is ​​a digital representation of the frequency density map, which is carried on the EEG device, so this part will not be explained in detail.

[0051] High-order statistics are also one of the important features extracted by the present invention, which are used to quantify the nonlinearity and distribution characteristics of modal signals. High-order statistics include kurtosis and skewness. Kurtosis is used to describe the sharpness of the signal distribution and can identify abnormal EEG activity patterns. The calculation formula is: (9); in, N Indicates the length of the time series data, Indicates The modal function in time The signal value on and The modal The mean and standard deviation of the signal. Skewness is used to describe the symmetry of the signal, and the calculation formula is: (10); The combination of skewness and kurtosis can help identify nonlinear changes in signals under different anesthetic states, thereby improving the diagnostic value of the features.

[0052] During the hierarchical decomposition process, the modal bandwidth will gradually decrease, which can reflect the accuracy of modal decomposition. The calculation of the bandwidth reduction degree is used to quantify the refinement effect of the mode after hierarchical decomposition. The calculation formula is: (11); in, For the hierarchical decomposition n The bandwidth of the layer, the greater the bandwidth reduction, means that the signal is effectively refined, indicating that the modal decomposition effect is good. Through hierarchical decomposition, the modal bandwidth is gradually reduced, so that the low-frequency interference signal is stripped out layer by layer, leaving the effective signal related to the anesthesia state.

[0053] Step 5: Fuse and filter the extracted multidimensional features to form a feature set: Feature fusion and screening uses a highly innovative feature engineering method, combining advanced mathematical models and feature fusion algorithms in the field of machine learning to build a more accurate and adaptable feature set. Different from traditional recursive feature elimination (RFE), this step uses a multi-level feature fusion strategy, combined with automated feature generation technology and deep feature representation methods to ensure that the extracted features can maximize the capture of dynamic changes in the anesthesia state, while improving the generalization and robustness of the model.

[0054] Through nonlinear feature mapping technology, the original feature set is mapped to a high-dimensional feature space. This process uses kernel functions for nonlinear transformation, among which commonly used kernel functions include Gaussian kernel (RBF kernel) and polynomial kernel. The mathematical expression of kernel mapping is as follows: (12); in, is the mapping function of the feature, is the adjustment parameter of the kernel function, and is the input feature vector. Through this mapping, the nonlinear relationship between features can be captured in high-dimensional space, thereby obtaining more implicit information. This nonlinear feature mapping is particularly suitable for processing multi-modal and multi-frequency interactive features in complex EEG signals, making the frequency coupling relationship in the original signal clearer.

[0055] In high-dimensional space, sparse coding algorithms are used to decompose and sparsely represent features. Sparse representation reduces feature dimensions by seeking to minimize representation errors while retaining the core structural information of the signal. The optimization goal of sparse representation is: (13); in, is the input feature, is a sparse dictionary matrix, is the sparse coefficient, is the regularization parameter. Through sparse coding, a small number of highly representative information can be extracted from a large number of features, and L1 regularization can effectively control the complexity of the model to avoid overfitting. The feature set after sparse representation can more effectively extract key information from the signal, especially in removing redundant features and noise.

[0056] Feature fusion adopts an adaptive weighted feature fusion method. The contribution of different modal features is different, so the weight of each feature needs to be adjusted dynamically. The present invention uses an adaptive feature weighting method based on gradient boosting decision tree (GBDT) to dynamically learn the contribution of each feature under different anesthesia states. GBDT constructs multiple weak classifiers and gradually learns the weighted coefficients of the features, and its loss function is: (14); in, is the predicted value of the model, is the true label, is a regularization term that controls the complexity of the model. By iteratively optimizing the loss function, GBDT can learn the weight of each feature, so that high-weight features contribute more to the prediction and low-weight features are suppressed. Adaptive weighted feature fusion can flexibly respond to individual differences of different patients, especially in the monitoring of the patient's anesthesia status, feature weight adjustment can improve the accuracy and robustness of the prediction.

[0057] In addition, in order to further improve the efficiency of feature fusion, the present invention introduces a multi-view feature learning method. This method regards the different frequency modes obtained by decomposition as multi-view information, and constructs a feature subspace for each mode. Then, these subspaces are jointly trained through a co-training algorithm. The core idea of ​​co-training is that the feature subspaces under different modes can complement each other and improve the generalization ability of the model. Multi-view feature learning can fully explore the implicit correlation between cross-modalities in EEG signals based on the fusion of different modal features.

[0058] Finally, the ensemble feature selection algorithm is used to screen the fused feature set. This algorithm combines a variety of feature selection methods, including recursive feature elimination, Lasso regression, etc., and finally determines the feature set with the most diagnostic value through weighted voting. The ultimate goal of ensemble feature selection is to improve the stability of the model and the reliability of the selection results through the combination of different selectors.

[0059] Through the above-mentioned feature fusion method, the present invention realizes the full-process optimization from high-dimensional nonlinear mapping, sparse coding, feature weighted fusion, multi-view learning to integrated feature selection, which significantly improves the monitoring accuracy of the anesthesia state and the generalization ability of the model.

[0060] Step 6: Use the feature set to train the machine learning model, and use the trained machine learning model to predict the patient's anesthesia status in real time, and output the anesthesia stage classification results: The present invention uses a variety of machine learning models to classify and predict the anesthesia state. Comparative experiments are carried out through models such as support vector machines, random forests and neural networks to find the optimal anesthesia state classification scheme. The support vector machine can accurately classify the EEG signals of different anesthesia stages by constructing a hyperplane that maximizes the classification boundary. The random forest combines the results of multiple decision trees through ensemble learning to improve the stability and anti-interference of the classification. The neural network uses a multi-layer perceptron structure to perform deep learning on the extracted complex nonlinear features, and can capture the nonlinear dynamic changes of different anesthesia stages.

[0061] Finally, the machine learning model is trained and classified according to the input feature set, and the output includes prediction results for four different stages: awake, light anesthesia, moderate anesthesia, and deep anesthesia. These prediction results are displayed to the anesthesiologist through a visual interface, helping the doctor to monitor the patient's anesthesia status in real time during the operation and adjust the anesthesia dose as needed to ensure the smooth progress of the operation and the safety of the patient.

[0062] The use process of the present invention is as follows: A patient undergoing general anesthesia was selected as the research subject, and his EEG signal data was collected during the intraoperative and postoperative stages. The device for collecting EEG signals was the Masimo SedLine brain function monitor. During the monitoring period, the patient was in a standard surgical environment and the sampling frequency was set to 256 Hz. The collected EEG signals were first preprocessed and bandpass filtered, with the filter range set to 0.5 Hz to 50 Hz. The signal after preliminary filtering was further decomposed by the H-VMD method, and reconstructed after removing the baseline drift noise. While decomposing with H-VMD, the system calculated and stored the various features of the intermediate modality, which were integrated as input features of the anesthesia state pre-training model to predict the patient's anesthesia state.

[0063] Before the start of anesthesia induction, the model output state is "awake". During the process, alpha waves and beta waves weakened, and theta waves gradually became dominant. At this time, the model output state is "light anesthesia". During the anesthesia maintenance stage, theta waves are enhanced, accompanied by a small amount of alpha wave activity, and the model output state is "moderate anesthesia". When burst suppression occurs or the spectrum energy is concentrated in the delta wave band, the model output state is "deep anesthesia". In this specific application, the classification model accurately predicted the different anesthesia states of the patient in different anesthesia intervals during the operation. This result is basically consistent with the state judged by BIS, which proves the effectiveness and reliability of the method of the present invention in predicting anesthesia states.

[0064] The above are only a few embodiments of the present application and do not constitute any form of limitation to the present application. Although the present application is disclosed as above with preferred embodiments, it is not intended to limit the present application. Any technician familiar with the profession, without departing from the scope of the technical solution of the present application, using the technical content disclosed above to make slight changes or modifications are equivalent to equivalent implementation cases and fall within the scope of the technical solution.

Claims

1. A high-precision anesthesia state monitoring method, characterized in that: The monitoring method comprises the following steps: Step 1: Collect EEG signals of the patient under anesthesia; Step 2: Use high-pass filter and low-pass filter to preliminarily filter the EEG signal, and normalize the amplitude of the EEG signal to obtain a new EEG signal; Step 3: Use the hierarchical frequency decomposition method to perform variational modal decomposition on the new EEG signal to obtain several modal functions; Step 4: Extract multidimensional features from each modal function; Step 5: Fuse and filter the extracted multidimensional features to form a feature set; Step 6: Use the feature set to train the machine learning model, and use the trained machine learning model to predict the patient's anesthesia status in real time and output the anesthesia stage classification results.

2. A high-precision anesthesia state monitoring method according to claim 1, characterized in that: The specific operation of preliminary filtering of EEG signals in step 2 is as follows: Step 201: using a 0.5 Hz high-pass filter to remove the baseline drift signal in the EEG signal; Step 202: Use a 50 Hz low-pass filter to remove high-frequency interference in the EEG signal.

3. A high-precision anesthesia state monitoring method according to claim 1, characterized in that: The specific process of using the hierarchical frequency decomposition method to perform variational modal decomposition on the new EEG signal in step 3 is as follows: Step 301: Perform the first-layer decomposition on the new EEG signal to obtain k First-order mode functions; Step 302: k The first-order mode function is decomposed into the second level The second-order mode function, l It means that each first-order mode function is decomposed into l Second-order mode functions, update The frequency distribution of the second-order mode function is calculated The frequency and amplitude characteristics of the second-order mode function; Step 303: Use the frequency of the first-order mode function as the reference frequency for baseline drift signal screening. k- Select the first-order mode function containing the reference frequency p The first-order low-frequency mode function is selected p The first-order low-frequency mode function The third-order modal function is decomposed into three layers for each second-order modal function, the frequency distribution of the third-order modal function is updated, and the frequency characteristics and amplitude characteristics of the third-order modal function are calculated; Step 304: Introduce the baseline drift signal standard frequency, from A second-order low-frequency modal function containing the standard frequency of the baseline drift signal is selected from the second-order modal functions, and the third-order modal function under the selected second-order low-frequency modal function is decomposed into a fourth-order modal function, the frequency distribution of the fourth-order modal function is updated, and the frequency characteristics and amplitude characteristics of the fourth-order modal function are calculated; Step 305: Continue to select the third-order low-frequency modal function containing the standard frequency of the baseline drift signal from the third-order modal function, and perform the next-level decomposition on the fourth-order modal function under the selected third-order low-frequency modal function, and repeat step 304 until n The first low-frequency mode function n+ 1 layer decomposition n+ If the first-order mode function does not contain the standard frequency of the baseline drift signal, stop the decomposition and update n+ The frequency distribution of the 1st order mode function and calculate n+ Frequency and amplitude characteristics of the first-order mode function; Step 306: Select the first-order modal function, the first-order low-frequency modal function, the second-order low-frequency modal function, the third-order low-frequency modal function, and n The first-order low-frequency mode function is used as a candidate mode for baseline drift signal reconstruction; Step 307: Reconstruct the baseline drift signal using the candidate mode, using the following formula: ; in, represents the signal time, represents the reconstructed baseline drift signal, is the low-frequency mode function of each order after hierarchical decomposition, express The total number of low-frequency mode functions; Step 308: Subtract the reconstructed baseline drift signal from the new EEG signal to obtain a purified EEG signal, as shown in the following formula: ; in, Represents a new EEG signal. In the purified signal, baseline drift and other low-frequency interference have been removed, retaining the patient's actual EEG activity.

4. A high-precision anesthesia state monitoring method according to claim 3, characterized in that: In step 301, the new EEG signal is decomposed into k independent first-order mode functions, each of which is compactly distributed near its center frequency, and the optimal and : （1）; in, Represents signal time, which is used to describe the changes of the signal in the time domain. Indicates The first-order mode function is in time The signal value on For the A first-order mode function, For the The center frequency of the first-order mode function, is the impulse function, used as the convolution kernel, is an ordinal unit used to represent the complex form of frequency components. is the convolution operator, used for convolution of the signal and the impulse function. is a derivative operation on time, which represents the bandwidth minimization constraint; and the center frequency of each first-order mode function obtained by the first-level decomposition is equally spaced.

5. A high-precision anesthesia state monitoring method according to claim 4, characterized in that: In steps 302-305, according to formula (1), The order mode functions are divided into several independent order mode function, and The frequency distribution of the order mode function is updated using the following formula: (2); in, represents the signal frequency, is the number of iterations, For the The modal function is in The center frequency at the iteration, Corresponding to The Lagrange multiplier of the iteration is used as the optimization constraint term. It is The modal function is in The spectrum update value of the iteration, is the frequency domain representation of the new EEG signal, is the penalty coefficient, represents the modal variable, and Respectively represent The modal function is in The spectrum update value of the iteration and the The modal function is in The spectrum update value of the iteration; but The center frequency of the first mode function is calculated using the following formula: （3）； The frequency characteristic calculated in steps 302-305 is the center frequency, which is obtained by formula (3); Each iterative decomposition will update the Lagrange multiplier, and the update formula of the Lagrange multiplier is: （4）； in, is the updated value of the Lagrange multiplier in the frequency domain update process, is the time step parameter in the dual ascent process, is the number of iterations.

6. A high-precision anesthesia state monitoring method according to claim 5, characterized in that: The modal functions in step 4 include first-order modal functions, second-order modal functions, third-order modal functions, fourth-order modal functions, and n+ 1st order mode function; Multidimensional features include: The center frequency and bandwidth of each modal function are used to reflect the main frequency characteristics of the new EEG signal; The frequency domain energy distribution and time-frequency energy matrix of each modal function are used to quantify the time-frequency characteristics of the new EEG signal; The higher-order statistics of each modal function include kurtosis and skewness, which are used to assess the nonlinearity and complexity of new EEG signals; The coherence and frequency overlap ratio between modal functions are used to describe the information correlation and independence between modal functions.

7. A high-precision anesthesia state monitoring method according to claim 5, characterized in that: The frequency of the first-order mode function is 1 Hz, and the standard frequency of the baseline drift signal is 0.5 Hz.

8. A high-precision anesthesia state monitoring method according to claim 7, characterized in that: The specific process of fusing and screening the extracted multidimensional features in step 5 is as follows: Step 501: Mapping the extracted multi-dimensional features to a high-dimensional feature space using a non-linear feature mapping technique to capture the complex non-linear relationship between the multi-dimensional features; Step 502: sparsely represent the multi-dimensional features in the high-dimensional feature space by using a sparse coding algorithm; Step 503: using an adaptive feature weighting method, dynamically learning the contribution of each multidimensional feature under different anesthesia states, and adjusting the weight of each multidimensional feature; Step 504: introducing a multi-view feature learning method, constructing feature subspaces for the multi-dimensional features extracted from each modal function, and jointly training these subspaces through a co-training algorithm to mine implicit information between different modal functions; Step 505: After processing steps 501-504, a fused feature set is obtained, and an integrated feature selection algorithm is used to screen the fused feature set.

9. A high-precision anesthesia status monitoring system, characterized in that: The monitoring system is implemented using the monitoring method according to any one of claims 1 to 8, including: The EEG signal acquisition module collects the patient's EEG signals during surgery through non-invasive electrodes; The signal processing module first performs preliminary filtering on the EEG signal to obtain a new EEG signal, then uses the hierarchical frequency decomposition method to perform variational mode decomposition on the new EEG signal to obtain several mode functions, and removes the baseline drift signal; The feature extraction and fusion module extracts the multidimensional features of each modal function, and fuses and filters the multidimensional features to obtain a feature set; The machine learning prediction module uses the feature set to train the machine learning model, and uses the trained machine learning model to predict the patient's anesthesia status in real time and output the anesthesia stage classification result; The visualization interface module displays the classification results of anesthesia stages.

10. A high-precision anesthesia status monitoring device, the monitoring device comprising a memory storing a computer program and a processor for executing the computer program, characterized in that: When the computer program is executed by a processor, the steps of the monitoring method according to any one of claims 1 to 8 are implemented.

Images