Feature fusion processing method for anesthesia depth multi-modal data

Abstract

The invention discloses a feature fusion processing method for anesthesia depth multi-modal data, and belongs to the technical field of graphic data processing and pattern recognition, and the method comprises the steps: obtaining a multi-modal physiological signal and an electromyographic signal of a patient; performing time axis calibration on the physiological signal to generate an alignment signal; extracting a multi-modal feature vector and calculating an anesthesia depth index; performing deviation analysis on the basis of the electromyographic signal and the index to obtain an electromyographic response deviation index; performing graphical feature mapping on the real-time electroencephalogram signal to generate a real-time time-frequency map; when the deviation index exceeds a safety threshold value, performing graph pattern matching with a pattern template library to calculate a similarity score; and outputting the current anesthesia state mode in a classified manner. According to the method, a multi-modal signal graphical feature fusion technology is adopted, and a dynamic map generation and pattern matching mechanism is combined, so that the problem of complex pattern recognition of physiological signal graphic data can be solved, and the accuracy and timeliness of anesthesia state classification are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of pattern data processing and pattern recognition, in particular to a feature fusion processing method for multi-modal data of anesthesia depth. BACKGROUND

[0002] In the field of pattern data processing and pattern recognition, a collaborative analysis method for multi-source physiological time series signals is involved. The present application belongs to the category of feature fusion processing of multi-modal heterogeneous data, and specifically to a real-time analysis and state classification system for physiological signal pattern data. This type of system realizes dynamic recognition of target states by pattern matching of time-frequency graphs generated by electromyographic signals, electroencephalographic signals and other physiological parameters.

[0003] The prior art mainly realizes state recognition through static template matching, and uses a pre-defined electroencephalogram database for fixed pattern comparison. In view of the time series differences of multi-channel signals, some systems introduce blood oxygen fluctuation trends to establish a time delay compensation model, and use linear interpolation to complete signal alignment. In the feature fusion stage, channel weighted average or principal component analysis is generally used to integrate multi-source features, and an independent threshold module is set to monitor electromyographic activity intensity.

[0004] The existing scheme has significant application limitations: first, the signal synchronization mechanism ignores the differences in individual drug metabolism dynamics, resulting in spatial and temporal misplacement of multi-channel data; second, the feature fusion process does not establish a cross-modal correlation topology, which cannot suppress the influence of heterogeneous signal interference on evaluation indicators; finally, when abnormal electromyographic activity is detected, the traditional graph matching method cannot dynamically adjust the evaluation strategy and still relies on static template library output results. The root cause of these defects lies in the fact that the intrinsic dynamic correlation of multi-source data has not been effectively modeled. SUMMARY

[0005] To solve the above problems, the present application provides a feature fusion processing method for multi-modal data of anesthesia depth, which uses a graphical feature fusion technology of multi-modal signals, combines dynamic graph generation and pattern matching mechanism, and can solve the complex pattern recognition problem of physiological signal pattern data, and improve the accuracy and timeliness of anesthesia state classification.

[0006] The above objective can be achieved by the following scheme: A feature fusion processing method for multi-modal data of anesthesia depth, comprising: acquiring multi-modal physiological signals and electromyographic signals of a patient, performing time axis calibration on the physiological signals to generate aligned signals, extracting multi-modal feature vectors and calculating an anesthesia depth index, performing deviation analysis based on the electromyographic signals and the index to obtain an electromyographic response deviation index, performing graphical feature mapping on real-time electroencephalographic signals to generate real-time time-frequency graphs, when the deviation index exceeds a safety threshold, performing graph pattern matching with a pattern template library to calculate a similarity score, and classifying and outputting a current anesthesia state pattern.

[0007] Optionally, the time axis calibration of the multi-modal physiological signals to generate the aligned multi-modal signals comprises: obtaining a blood oxygen signal of the multi-modal physiological signals; extracting a change slope of the blood oxygen signal to establish a drug metabolism rate prediction model; calculating a signal delay amount based on the drug metabolism rate prediction model to generate a time alignment parameter; and performing waveform matching on the multi-modal physiological signals by using the time alignment parameter to generate the aligned multi-modal signals.

[0008] Optionally, the calculation of the anesthetic depth index based on the multi-modal feature vector comprises: obtaining a time-domain pharmacodynamic correlation tensor based on the multi-modal feature vector; and performing dynamic aggregation on the time-domain pharmacodynamic correlation tensor to calculate the anesthetic depth index.

[0009] Optionally, the extraction of the change slope of the blood oxygen signal to establish the drug metabolism rate prediction model comprises: calculating the change slope of the blood oxygen signal to generate a drug concentration gradient feature matrix; and performing pharmacokinetic projection on the drug concentration gradient feature matrix to establish the drug metabolism rate prediction model.

[0010] Optionally, the graphical feature mapping of the real-time electroencephalogram signal to generate a real-time time-frequency spectrum comprises: performing adaptive time-frequency conversion on the real-time electroencephalogram signal to obtain a complex time-frequency component three-dimensional matrix; and performing dynamic spectrum rendering on the complex time-frequency component three-dimensional matrix to generate the real-time time-frequency spectrum.

[0011] Optionally, the adaptive time-frequency conversion of the real-time electroencephalogram signal to obtain a complex time-frequency component three-dimensional matrix comprises: performing multi-channel collaborative filtering and time-varying window adaptive decomposition on the real-time electroencephalogram signal to obtain a three-dimensional complex coefficient tensor; and performing phase synchronization and energy normalization reorganization on the three-dimensional complex coefficient tensor to obtain the complex time-frequency component three-dimensional matrix.

[0012] Optionally, the deviation analysis of the electromyographic signal and the anesthetic depth index to obtain an electromyographic response deviation index comprises: obtaining a dynamic response deviation vector based on the electromyographic signal and the anesthetic depth index; performing event slicing processing on the dynamic response deviation vector to generate a calibration trigger signal; and performing time window integration fusion on the calibration trigger signal to obtain the electromyographic response deviation index.

[0013] Optionally, the event slicing processing of the dynamic response deviation vector to generate a calibration trigger signal comprises: obtaining a pharmacodynamic sensitive mutation point sequence based on the dynamic response deviation vector; and performing physiological event binding on the pharmacodynamic sensitive mutation point sequence to generate the calibration trigger signal.

[0014] Optionally, the extracting the aligned multi-modal signals to generate a multi-modal feature vector comprises: performing dynamic path topology analysis on the aligned multi-modal signals to obtain a cross-modal coupling feature set; and performing entropy value fusion on the cross-modal coupling feature set to generate a multi-modal feature vector.

[0015] Based on the same inventive concept, the present application also provides a feature fusion processing system for anesthesia depth multi-modal data, which comprises: a signal acquisition module for acquiring electromyographic signals of a patient and multi-modal physiological signals including real-time electroencephalographic signals; a signal synchronization module for performing time axis calibration on the multi-modal physiological signals to generate aligned multi-modal signals; a feature extraction module for extracting the aligned multi-modal signals to generate a multi-modal feature vector; an anesthesia index calculation module for calculating an anesthesia depth index based on the multi-modal feature vector; an electromyographic response analysis module for performing deviation analysis on the electromyographic signals and the anesthesia depth index to obtain an electromyographic response deviation index; a real-time time-frequency spectrum generation module for performing graphical feature mapping on the real-time electroencephalographic signals to generate a real-time time-frequency spectrum; a pattern matching calculation module for, when the electromyographic response deviation index exceeds a preset safety threshold, performing pattern matching between the real-time time-frequency spectrum and a preset pattern template library to calculate a similarity score; and an anesthesia state classification output module for matching an anesthesia state pattern corresponding to the pattern template library based on the similarity score to classify and output an anesthesia state of the current patient.

[0016] Compared with the prior art, the present application has the following advantages: The present application establishes a safety calibration mechanism combining main monitoring and bypass verification by constructing a dual evaluation path of anesthesia depth index and electromyographic response deviation index. When the conventional anesthesia depth index may not fully reflect the stress state of the patient due to the characteristics of the drug or individual differences of the patient, the independent electromyographic response analysis can timely capture this mismatch and trigger deeper electroencephalographic pattern analysis, thereby effectively avoiding the risk of anesthesia too shallow or intraoperative failure due to the limitations of a single indicator, and significantly improving the overall safety and reliability of anesthesia monitoring.

[0017] The present application proposes a time axis calibration method based on dynamic changes of blood oxygen signals to predict drug metabolism rate. This method can dynamically compensate for the individualized time delay of different physiological signals caused by drug metabolism and conduction according to real-time physiological feedback of the patient, ensuring that the time base of multi-modal data is highly consistent before fusion analysis. Compared with the method using fixed delay parameters, this greatly improves the accuracy of subsequent feature extraction and fusion, so that the final anesthesia depth evaluation result can more truly reflect the comprehensive physiological state of the patient.

[0018] The application overcomes the analysis of isolated features of a single physiological signal by adopting dynamic path topology analysis and time domain drug efficacy correlation tensor and other advanced feature engineering technologies.The method focuses on mining the dynamic coupling relationship and time sequence dependency of different physiological systems under the influence of anesthesia, and extracts cross-modal system-level features that can better reveal the essence of the anesthesia state. This deep feature extraction enables the anesthesia depth evaluation model to be based on a more stable physiological mechanism, enhancing the robustness and anti-interference ability of the final index.

[0019] When an abnormal stress signal is detected, the application can convert one-dimensional real-time electroencephalogram signals into two-dimensional real-time time-frequency maps through adaptive time-frequency conversion and dynamic atlas rendering, which are rich in information and suitable for pattern recognition. Combined with the pre-set pattern template library for high-precision matching, the application can achieve fine classification of the anesthesia state. This intelligent upgrade from quantitative evaluation to qualitative classification provides clinicians with more intuitive and more diagnostic decision-making basis beyond single numerical values, improving the handling ability of complex and critical situations.

[0020] Other features and advantages of the present application will be set forth in the following description, and in part will be apparent from the description, or can be learned by practice of the present application. The objects and other advantages of the present application will be realized and attained by the structure particularly pointed out in the description and claims hereof as well as the appended drawings. BRIEF DESCRIPTION OF DRAWINGS

[0021] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiment or prior art description. Obviously, the drawings in the following description are 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.

[0022] Figure 1 is a flowchart of a feature fusion processing method of multi-modal data of anesthesia depth according to an embodiment of the present application.

[0023] Figure 2 is an electromyographic response deviation index chart according to an embodiment of the present application.

[0024] Figure 3 is a process chart of anesthesia depth index calculation according to an embodiment of the present application.

[0025] Figure 4 is a time domain drug efficacy correlation tensor chart according to an embodiment of the present application.

[0026] Figure 5 is a real-time time-frequency atlas chart according to an embodiment of the present application.

[0027] Figure 6is a structural schematic diagram of a feature fusion processing system of anesthesia depth multi-modal data according to an embodiment of the present application. DETAILED DESCRIPTION

[0028] To make the objectives, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are some but not all of the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative work fall within the protection scope of the present application.

[0029] Referring to Figure 1 An embodiment of the present application provides a feature fusion processing method of anesthesia depth multi-modal data, adopts a graphical feature fusion technology of multi-modal signals, combines a dynamic atlas generation and a pattern matching mechanism, can solve a complex pattern recognition problem of physiological signal graphical data, and improves the accuracy and timeliness of anesthesia state classification.

[0030] The method according to the embodiment specifically includes: obtaining an electromyographic signal of a patient and multi-modal physiological signals including real-time electroencephalographic signals; performing time axis calibration on the multi-modal physiological signals to generate aligned multi-modal signals; extracting the aligned multi-modal signals to generate multi-modal feature vectors; calculating an anesthesia depth index based on the multi-modal feature vectors; performing deviation analysis on the electromyographic signal and the anesthesia depth index to obtain an electromyographic response deviation index; performing graphical feature mapping on the real-time electroencephalographic signals to generate a real-time time-frequency atlas; when the electromyographic response deviation index exceeds a preset safety threshold, performing pattern matching on the real-time time-frequency atlas and a preset pattern template library to calculate a similarity score; based on the similarity score, matching a corresponding anesthesia state pattern of the pattern template library, and classifying and outputting an anesthesia state of the current patient.

[0031] Specifically, first in the core evaluation layer, through the collection of the patient's multi-modal physiological signals such as electroencephalogram, electrocardiogram, blood oxygen and the like, and innovatively performing time axis calibration, it is ensured that the signals are synchronized in the pharmacological sense. Then, from these aligned signals, features reflecting the coupling relationship between systems are extracted, fused into a multi-modal feature vector, and based on this, a comprehensive anesthesia depth index is calculated as the anesthesia depth evaluation benchmark under the conventional state. Secondly, in the alarm and calibration layer, the electromyogram signal is introduced as an independent reference system, and through deviation analysis, it is compared with the anesthesia depth index derived from the electroencephalogram to generate an electromyogram response deviation index, which is Figure 2The deviation index is shown. When the deviation index exceeds the safety threshold, it indicates that the core evaluation may be out of touch with the actual somatic response of the patient, and the system initiates an independent verification mechanism. The safety threshold is the core trigger condition for starting the atlas matching path, which determines the balance between clinical safety and sensitivity, and adopts a dynamic mechanism of "basic benchmark value + individual correction", which includes: basic benchmark value determination: based on large sample clinical data statistics, the correlation between electromyographic response deviation index and adverse events of anesthesia (such as intraoperative awareness, excessive anesthesia) is analyzed, and the ROC curve analysis method is used to determine the universal benchmark value; individual adjustment rule: according to the individual characteristics of the patient (age, weight, underlying disease, type of surgery), the threshold is adjusted, for example: for elderly patients (≥65 years old), the threshold is lowered by 10%-15% (such as 0.6-0.63) due to decreased nerve and muscle sensitivity; for obese patients (BMI≥30), the threshold is raised by 5%-8% (such as 0.73-0.76) due to higher baseline electromyographic signal; for neurosurgery, the threshold is lowered by 5% (such as 0.66) for higher sensitivity, while for local anesthesia assisted sedation, the threshold is raised by 10% (such as 0.77); clinical calibration process: a threshold prediction model is constructed through multi-center clinical data, and the initial threshold is automatically generated after inputting the patient's characteristics, and the anesthesiologist can fine-tune it within a range of ±0.05 according to clinical experience, and verify the baseline signal before surgery for 3-5 minutes (if the baseline deviation index is continuously higher than the initial threshold, secondary correction is automatically triggered). The mechanism generates real-time time-frequency atlas by processing real-time electroencephalogram signals, and performs pattern matching with a pattern template library containing various known anesthesia state patterns. The pattern template library is constructed through clinical data training, including standard time-frequency atlas templates and feature vectors corresponding to various typical anesthesia states, and the specific establishment method includes: clinical data collection: collect electroencephalogram signals of patients of different ages, weights and surgery types in various anesthesia states (such as stable anesthesia, light anesthesia, intraoperative awareness, electromyographic interference, etc.), and record the dosage of anesthetic drugs, vital signs and the anesthesia state label judged by the doctor; standardization preprocessing: the collected electroencephalogram signals are processed in the same way as the "real-time time-frequency atlas generation" (adaptive time-frequency conversion, complex time-frequency component three-dimensional matrix construction, dynamic atlas rendering), to generate standardized time-frequency atlas; template clustering and labeling: the standardized time-frequency atlas is classified by clustering algorithm (such as K-means), and the anesthesia state label marked by the doctor is combined to determine the anesthesia state pattern corresponding to each type of atlas (such as wave energy is dominant, low activity in wave band, "light anesthesia" corresponds Waveband enhancement and increase of high frequency components); feature vector extraction: extracting key time-frequency features (such as energy proportion of each frequency band, peak frequency, phase synchronicity, etc.) of each type of standard atlas to form a feature vector as the template core data, and storing in the pattern template library; dynamic optimization: regularly incorporating new clinical data to update the template library through incremental learning, and correcting the template features to adapt to individual differences and changes in clinical scenarios. The structure of the pattern template library covers 6 types of core anesthesia state patterns, and the specific features are as follows:

[0032] The template library adopts a multi-dimensional index structure, and the feature vectors of real-time time-frequency atlas and template feature vectors are quickly matched through cosine similarity algorithm. A similarity score ≥ 80% is determined as a successful match, and a similarity score < 50% triggers manual review. Finally, according to the similarity score of the match, the anesthesia state of the current patient is accurately classified and output, realizing the identification and confirmation of abnormal states.

[0033] Optionally, the time axis calibration of the multi-modal physiological signals to generate the aligned multi-modal signals comprises: Obtaining a blood oxygen signal of the multi-modal physiological signals; Extracting the change slope of the blood oxygen signal, and establishing a drug metabolism rate prediction model; Based on the drug metabolism rate prediction model, calculating the signal delay amount to generate a time alignment parameter; Using the time alignment parameter to perform waveform matching on the multi-modal physiological signals to generate the aligned multi-modal signals.

[0034] Specifically, first, from the obtained multi-modal physiological signals, such as electroencephalogram signals, electrocardiogram signals, and blood pressure signals, the blood oxygen signal is specifically extracted. The blood oxygen signal, especially the change of its saturation, can indirectly reflect the influence of anesthetic drugs on the patient's circulatory system and tissue perfusion, and its change rate is related to the diffusion and action rate of drugs in the body. Then, the extracted blood oxygen signal is processed to extract its change slope. The change slope is a quantification of the local change rate of the blood oxygen signal in the time sequence, which can be regarded as a dynamic indicator reflecting the drug effect kinetics. Calculating the change slope generates a drug metabolism related dynamic feature vector. To establish a drug metabolism rate prediction model, the core of the model is to use the above drug metabolism related dynamic feature vector to predict the time delay between the central nervous system signals and the peripheral physiological signals caused by the drug. Specifically, the change slope of the blood oxygen signal is used as a key input variable to dynamically predict the distribution and metabolism rate of the drug in the body, and then the time delay amount between different physiological signals is calculated. This delay amount is the time alignment parameter, and the establishment of the model can be represented as: , wherein, represents the calculated signal delay, i.e. the time alignment parameter. represents the slope of the real-time calculated blood oxygen signal, which reflects the immediate dynamics of drug effect. represents a set of baseline physiological parameters of patients, such as age, weight, etc., which are used as static inputs of the model to improve the accuracy of personalized prediction. represents the functional form of the drug metabolism rate prediction model, which can be a mathematical equation based on physiological principles. The baseline linear model is as follows: , wherein is the signal delay, is the blood oxygen saturation signal, is the absolute value of the slope of the blood oxygen signal change, is the baseline physiological parameter, denotes the model coefficients (determined by fitting clinical data). Finally, the calculated dynamic time alignment parameter is used to perform waveform matching on the multi-modal physiological signals. The search window of dynamic time warping is constrained by the time alignment parameter, and the minimum distortion path of each signal and the reference electroencephalogram signal is calculated to achieve waveform matching. The principle of dynamic time warping is to calculate the optimal alignment path of two time series by dynamic programming, minimizing the cumulative distance (such as Euclidean distance), as follows: , wherein is the cumulative distance, is the distance measure between points. denotes the selection of the path with the minimum cumulative distance from the three possible predecessor positions (i.e. the position of the last step). The waveform matching process takes one signal as the reference, usually the electroencephalogram signal with the most direct response, and then adjusts the time domain of other physiological signals, such as heart rate and blood pressure, according to their respective time alignment parameters. By this way, the waveforms of various signals that originally deviated in time axis are accurately aligned, ensuring that the multi-modal data at the same time point can truly reflect the comprehensive physiological state of the patient at that time, and finally generating the aligned multi-modal signals, providing a high-quality synchronous data basis for subsequent feature extraction and fusion analysis.

[0035] Optionally, the calculation of the anesthesia depth index based on the multi-modal feature vector comprises: obtaining a time-domain pharmacodynamic correlation tensor based on the multi-modal feature vector; performing dynamic aggregation on the time-domain pharmacodynamic correlation tensor to calculate the anesthesia depth index.

[0036] Specifically, first, the multi-modal feature vectors obtained from the previous step are taken as the starting point. These vectors collect key information from different physiological signals at each time point. Based on these time-sequenced multi-modal feature vectors, the system constructs a time-domain pharmacodynamic association tensor, as shown in Figure 4 The time-domain pharmacodynamic association tensor is constructed as follows: a three-dimensional tensor is constructed with a sliding time window as the time dimension and multi-modal features as the other two dimensions. The tensor elements are filled by calculating the mutual information value of any two features within the window to quantify the dynamic association between features. The time-domain pharmacodynamic association tensor here is a three-dimensional or higher-dimensional mathematical object that represents the mutual relationship between different physiological features over time. The specific construction method is to calculate the association strength between any two features in the multi-modal feature vector within a sliding short time window, for example, using mutual correlation, mutual information, or phase synchronization index as an indicator to quantify. These pairwise association strength values are combined into a two-dimensional matrix that fully describes the coupling network between all features at a specific time point. The calculation method of association strength is based on phase amplitude coupling: , where represents the phase angle of the blood oxygen signal after transformation in the 4-8 Hz frequency band, represents the phase angle of the electroencephalogram signal in the 30-50 Hz frequency band, represents the number of sampling points within the time window, is the imaginary unit, represents the th feature in the fixed time window, th feature in the fixed time window, a series of correlation matrices, and stacking these matrices along the time axis forms a time-domain pharmacodynamic correlation tensor. One dimension of the tensor is time, and the other two dimensions represent different physiological feature pairs. The element values of the tensor reflect the coupling strength of a specific feature pair at a specific time, and intuitively present the pattern of the dynamic evolution of the physiological system network under the action of anesthetic drugs. Subsequently, in order to extract a single and intuitive anesthetic depth index from this complex high-dimensional tensor, dynamic aggregation of the time-domain pharmacodynamic correlation tensor is required. The calculation of the anesthetic depth index obtained by dynamic aggregation of the time-domain pharmacodynamic correlation tensor includes the following steps: using a three-layer LSTM network, inputting the tensor according to time slicing, the number of hidden units of each layer being 64, 32 and 16 respectively, calculating the output aggregation value through time series recursion, and then mapping it to the anesthetic depth index range of 0-100 through the Sigmoid function. Dynamic aggregation is not a simple summation or averaging, but a complex mapping process of projecting high-dimensional information to a one-dimensional scalar. This process aims to give different weights to different feature correlation pairs according to their importance in indicating anesthetic depth, and finally calculate the anesthetic depth index. The process of calculating the anesthetic depth index is shown in FIG. 1. Figure 3 The anesthetic depth index function can identify and amplify the coupling pattern most relevant to the change in consciousness level, while suppressing the interference of irrelevant or noise correlation, and finally output a continuous value that can accurately quantify the anesthetic depth of the patient.​​​​​​​​​​​​​​​​​​

[0037] Optionally, the extracting the change slope of the blood oxygen signal, and establishing a drug metabolism rate prediction model comprises: calculating the change slope of the blood oxygen signal to generate a drug concentration gradient feature matrix; performing pharmacodynamic projection on the drug concentration gradient feature matrix to establish a drug metabolism rate prediction model.

[0038] Specifically, the process starts with calculating the change slope of the blood oxygen signal. Blood oxygen saturation is a sensitive indicator reflecting the comprehensive influence of anesthetic drugs on the respiratory and circulatory systems, and its change rate can indirectly represent the intensity change of drug effect. By performing time difference operation on the continuously collected blood oxygen signal, the instantaneous change slope can be obtained. These slope values calculated at consecutive time points are constructed into a time series, and data within a certain time window is selected to generate a drug concentration gradient feature matrix . Each element of the matrix quantifies the physiological response speed caused by the drug acting on the body at a certain time, so the matrix as a whole depicts the dynamic map of the change of drug effect over time, which can be regarded as an indirect representation of the drug concentration gradient in the effect compartment. After generating the drug concentration gradient feature matrix, the method performs pharmacodynamic projection on it to establish a drug metabolism rate prediction model. The pharmacodynamic projection here is an analysis method based on model conversion, and its core is to map the observed physiological representation data, i.e. the drug concentration gradient feature matrix to a low-dimensional pharmacodynamic model space that can describe drug absorption, distribution, and metabolism. This process can be expressed as: , where represents the predicted drug metabolism rate, which is the output of the model. is the input drug concentration gradient feature matrix. is a nonlinear function representing pharmacodynamic projection, which itself is the model structure to be established. Its form can be constructed based on classical pharmacological models such as Emax model or compartment model. The pharmacodynamic projection is based on a modified Emax-compartment hybrid model, and the core function expression is: , where is the time series value of drug concentration derived from the drug concentration gradient feature matrix; is the maximum metabolism rate, representing the highest metabolism level that the drug can reach in the body; is the half-metabolism concentration, i.e. the drug concentration at which the metabolism rate reaches half of the maximum metabolism rate. To eliminate the rate constant, dynamic correction will be made according to the patient's age, weight, underlying disease and other physiological parameters; is a time variable. The model building process is as follows: first, collect the multi-center clinical patient's blood oxygen signal change slope, the corresponding drug concentration time series data and the patient's physiological parameter information; then, use the nonlinear least squares method to fit the parameters in the above function , so that the deviation between the model prediction results and the actual observation data is minimized; finally, through 10-fold cross-validation, the model is optimized to ensure that the prediction error of the model in different patient groups is controlled within 8%, so as to obtain a stable and reliable drug metabolism rate prediction model. is a set of undetermined parameters of the model, such as the maximum effect rate, half effect concentration, etc. These parameters need to be estimated by fitting the actually observed matrix with the model prediction behavior. By adjusting the parameters through the optimization algorithm, the model output can best reproduce the physiological change process represented by the input matrix, which is the projection process. The final function is a prediction model that takes the blood oxygen signal change rate as input and dynamically outputs the individual drug metabolism rate.

[0039] Optionally, the mapping of the real-time electroencephalogram signal to a graphical feature to generate a real-time time-frequency spectrum includes: Adaptive time-frequency conversion is performed on the real-time electroencephalogram signal to obtain a complex time-frequency component three-dimensional matrix; Dynamic atlas rendering is performed on the complex time-frequency component three-dimensional matrix to generate a real-time time-frequency spectrum.

[0040] Specifically, first, the real-time electroencephalogram signal is subjected to adaptive time-frequency conversion. This conversion is an advanced signal processing technique, the core advantage of which lies in the ability to dynamically adjust the resolution of the analysis according to the characteristics of the signal itself. For a typical non-stationary signal such as an electroencephalogram signal, which contains both slow-changing low-frequency rhythms and possible burst high-frequency oscillations, adaptive time-frequency conversion can analyze different frequency components using different time windows and frequency windows, thereby achieving optimal resolution in both time and frequency. The conversion process decomposes the input real-time electroencephalogram signal, and its output is a complex time-frequency component three-dimensional matrix. The three dimensions of the matrix represent time, frequency, and electroencephalogram channel, respectively. Each element in the matrix is a complex number, containing the amplitude and phase information of the electroencephalogram signal component at a specific time point, specific frequency, and specific channel. Next, the obtained complex time-frequency component three-dimensional matrix is subjected to dynamic atlas rendering to generate a visualized real-time time-frequency atlas. The rendering process mainly focuses on the distribution of the signal's energy or power in the time-frequency plane. Generally, the power of the signal at a certain time-frequency point is proportional to the square of the amplitude of the corresponding complex component. This process can be represented as: , wherein, represents the power spectral density value at time point and frequency . is a complex value extracted from the complex time-frequency component three-dimensional matrix, corresponding to a specific time and frequency after fusion of a single or multiple channels. The rendering step is to convert the calculated series of values into corresponding colors through a preset color mapping table. For example, high power values correspond to warm colors such as red, and low power values correspond to cool colors such as blue. Since the electroencephalogram signal is continuously input, the above conversion and rendering process is also continuously and dynamically performed, thereby forming a color image that is updated in real time with time on the display interface, i.e., a real-time time-frequency atlas, as shown in Figure 5 . The horizontal axis of the atlas represents time, the vertical axis represents frequency, and the color depth represents the energy intensity of the frequency component at a specific time.

[0041] Optionally, the adaptive time-frequency conversion of the real-time electroencephalogram signal to obtain a complex time-frequency component three-dimensional matrix comprises: subjecting the real-time electroencephalogram signal to multi-channel collaborative filtering and time-varying window adaptive decomposition to obtain a three-dimensional complex coefficient tensor; subjecting the three-dimensional complex coefficient tensor to phase synchronization and energy normalization reorganization to obtain a complex time-frequency component three-dimensional matrix.

[0042] Specifically, first, the real-time electroencephalogram signal is subjected to multi-channel cooperative filtering and time-varying window adaptive decomposition. The cooperative filtering here is not independent filtering of each channel, but utilizes the spatial correlation between multi-channel signals to identify and suppress common noise interference such as power frequency interference or motion artifacts, thereby maximizing the signal-to-noise ratio while retaining the true brain neural activity signal. Then, time-varying window adaptive decomposition is performed on the filtered multi-channel signal. This is an advanced signal decomposition technique, such as a variant of adaptive wavelet transform or empirical mode decomposition, which can dynamically adjust the length of the analysis time window according to the characteristics of the local frequency components of the signal. For low-frequency rhythms with long duration in the electroencephalogram signal, a longer time window is used to obtain high frequency resolution; for transient high-frequency activity, a shorter time window is switched to ensure accurate time positioning. The output of this decomposition process is a three-dimensional complex coefficient tensor, whose dimensions represent the electroencephalogram channel, frequency and time, respectively. Each element in the tensor is a complex number containing the signal amplitude and phase information at a specific frequency point at a specific time for a specific channel. Then, the three-dimensional complex coefficient tensor generated above is subjected to phase synchronization and energy normalization reorganization. The purpose of phase synchronization processing is to correct and align the phase deviations caused by signal transmission delay or measurement error between different channels. By calculating the phase consistency across channels in a specific frequency band, the phase information of each channel is adjusted to a common reference, which enhances the effective representation of the coordinated neural oscillation activity of the whole brain or specific brain regions. Subsequently, energy normalization is performed to eliminate the inherent energy differences between different channels, different frequency bands and different individuals, so that subsequent analysis and comparison are not disturbed by the absolute amplitude of the signal. Normalization can map the energy values of each channel or frequency band to a unified scale range, such as through Z-score standardization. After phase synchronization and energy normalization, the original three-dimensional complex coefficient tensor is reorganized to form the final complex time-frequency component three-dimensional matrix. The data in this matrix not only reflects the time-frequency information, but also undergoes cross-channel coordination and standardization, with higher consistency and comparability.

[0043] Optionally, the bias analysis on the electromyographic signal and the depth of anesthesia index includes: Based on the electromyographic signal and the depth of anesthesia index, a dynamic response bias vector is obtained; The dynamic response bias vector is subjected to event slicing processing to generate a calibration trigger signal; The calibration trigger signal is subjected to time window integration fusion to obtain an electromyographic response bias index.

[0044] Specifically, first, a dynamic response deviation vector is generated based on the real-time collected EMG signals and the calculated depth of anesthesia index. Since EMG signals are high-frequency raw physiological waveforms, while the depth of anesthesia index is a low-frequency, normalized index calculated through complex computation, the two cannot be directly compared. Therefore, the system first pre-processes the EMG signals, for example, calculates the energy or root mean square value in a sliding time window, to obtain a time series index that can continuously reflect the intensity of muscle activity . Subsequently, the system compares this muscle activity intensity index with the depth of anesthesia index to generate a dynamic response deviation vector . Ideally, deep anesthesia (low ) should correspond to low muscle activity. When there is low but high , it is a deviation. This deviation vector can be represented by the following functional relationship: , where is a nonlinear function, and when is significantly higher than the baseline level expected by , the value of will significantly increase. Then, the system eventizes the slice processing of the dynamic response deviation vector to generate a calibration trigger signal. The purpose of this step is to convert continuous deviation values into discrete, meaningful "deviation events". The system sets a deviation threshold, and when the value of the dynamic response deviation vector exceeds the threshold, it is considered that a significant EMG-central mismatch event has occurred. At this time, the system generates a pulse-like calibration trigger signal , which is 1 at the moment of the event and 0 at other times. This processing method can ignore small, clinically insignificant fluctuations and focus only on strong deviation events that may indicate patient responsiveness. Finally, the system integrates the generated calibration trigger signal in a time window to calculate the final EMG response deviation index. Since a single deviation event may be caused by accidental interference, a stable and reliable index needs to consider the cumulative effect of deviation events in a period of time. Therefore, the system integrates or sums the calibration trigger signal in a retrospective time window, which is calculated as follows: , In this formula, is the EMG response deviation index at the current time, is the length of the time window, is the calibration trigger signal in the time window. This index reflects the density or duration of significant deviation events in the recent period.

[0045] Optionally, the eventizing and slicing the dynamic response deviation vector to generate a calibration trigger signal comprises: Based on the dynamic response deviation vector, a pharmacodynamic sensitive mutation point sequence is obtained; Physiological event binding is performed on the pharmacodynamic sensitive mutation point sequence to generate a calibration trigger signal.

[0046] Specifically, first, based on the dynamic response deviation vector generated in the previous step, a pharmacodynamic sensitive mutation point sequence is identified and extracted. The dynamic response deviation vector is a continuous time series reflecting the real-time difference between electromyographic activity and depth of anesthesia index. The system continuously monitors the first derivative or rate of change of the vector, and when it appears a positive sharp increase far beyond normal fluctuations in a very short time, and the amplitude of the vector itself also exceeds a preset sensitivity threshold, the time point is marked as a pharmacodynamic sensitive mutation point. This process aims to capture the sudden increase of the deviation signal. Such mutations often indicate the patient's instantaneous response to external stimuli (such as surgical incision or traction), or indicate that the concentration of anesthetic drugs has rapidly decreased below the critical level. By analyzing the entire monitoring process , the system can obtain a discrete sequence composed of the timestamps of these mutation points, i.e. the pharmacodynamic sensitive mutation point sequence. After obtaining the sequence, the system performs physiological event binding to generate the final calibration trigger signal. The core of this step is to associate abstract mathematical mutation points with real and possible clinical events. The system checks whether there are synchronous changes in other physiological signals, such as sudden increases in heart rate or blood pressure, within a very short time window before and after each pharmacodynamic sensitive mutation point. If such a multi-system synchronous stress response pattern exists, the reliability of the mutation point is enhanced and it is formally confirmed as an event that needs to start in-depth analysis. The system then converts the timestamp of this confirmed mutation point into a pulse signal to form a calibration trigger signal sequence . This binding process is equivalent to a verification and filtering mechanism that uses cross-verification of multiple physiological parameters to improve the accuracy of the triggering decision.

[0047] Optionally, the extracting the aligned multi-modal signals to generate a multi-modal feature vector comprises: Performing dynamic path topology analysis on the aligned multi-modal signals to obtain a cross-modal coupling feature set; Performing entropy value fusion on the cross-modal coupling feature set to generate a multi-modal feature vector.

[0048] Specifically, first, dynamic path topology analysis is performed on these time-synchronized multi-modal signals. Dynamic path topology analysis uses a Granger causality analysis algorithm based on a state space model, which is as follows: a state space model is constructed, and the state equation is: , is the state transition matrix, is the input matrix, is the process noise vector, is the external input vector, is the state vector ， z denotes the system state vector at time point . The observation equation is: , is the multi-modal physiological signal observation, is the observation matrix, is the observation noise vector. The dynamic path topology analysis goes beyond the isolated investigation of a single signal, and instead treats physiological signals from different sources, such as electroencephalogram, electrocardiogram, blood pressure, etc., as nodes in a dynamic network. This analysis quantifies the direction, strength and time lag of information flow between these nodes within a continuous sliding time window through computational techniques. In this way, the system is able to construct a functional connectivity network topology graph that evolves over time. The output of this analysis is a cross-modal coupling feature set, which contains a series of quantitative indices that describe the network topology structure, such as the information transfer strength from heart rate variability to electroencephalogram band energy, or the phase locking value between blood pressure fluctuation and electroencephalogram activity, etc. These features collectively characterize the complex interaction patterns between the central nervous system and the autonomic nervous system under anesthesia. Then, the obtained cross-modal coupling feature set is subjected to entropy fusion to generate the final multi-modal feature vector. This fusion process aims to intelligently select and integrate numerous coupling features to highlight the most informative features under the current anesthesia state. In implementation, the system calculates the information entropy of each coupling feature over time. Information entropy is used here as an indicator of the information content or certainty of a feature, and generally a feature that exhibits stable and distinguishable patterns under different anesthesia depths has a lower information entropy and higher information value. Based on this, the system assigns a weight to each coupling feature, which is inversely proportional to the information entropy of the feature. This process can be represented as: , where, is the generated multi-modal feature vector. is the normalized th cross-modal coupling feature value. is the entropy weight corresponding to the feature, whose information entropy is The lower the information entropy, the higher the weight to amplify the contribution of the feature in the final vector. By multiplying each normalized feature value with its dynamically calculated entropy weight, the system finally generates a multi-modal feature vector that contains both multi-dimensional coupling information and optimized information value.

[0049] Based on the same inventive concept, as Figure 6 indicated, the present application also provides a feature fusion processing system for multi-modal anesthesia depth data, which comprises: a signal acquisition module for acquiring electromyographic signals of a patient and multi-modal physiological signals including real-time electroencephalographic signals; a signal synchronization module for time axis calibration of the multi-modal physiological signals to generate aligned multi-modal signals; a feature extraction module for extracting the aligned multi-modal signals to generate a multi-modal feature vector; an anesthesia index calculation module for calculating an anesthesia depth index based on the multi-modal feature vector; an electromyographic response analysis module for deviation analysis of the electromyographic signals and the anesthesia depth index to obtain an electromyographic response deviation index; a real-time time-frequency spectrum generation module for graphical feature mapping of the real-time electroencephalographic signals to generate a real-time time-frequency spectrum; a pattern matching calculation module for pattern matching of the real-time time-frequency spectrum with a preset pattern template library to calculate a similarity score when the electromyographic response deviation index exceeds a preset safety threshold; an anesthesia state classification output module for matching the anesthesia state pattern corresponding to the pattern template library based on the similarity score to classify and output the anesthesia state of the current patient.

[0050] To verify the feasibility of the present application in implementation, the present application is applied to the general anesthesia operation monitoring scene of the operation center of a certain third-grade class-A hospital. The operation center aims to improve the accuracy and safety of anesthesia depth monitoring, especially in dealing with individual differences of patients and the risk of implicit awakening caused by intraoperative stimulation. The method of the present application is integrated into an anesthesia monitoring device to monitor a patient in a laparoscopic cholecystectomy. The surgical patient is a 55-year-old male with an ASA (American Society of Anesthesiologists) classification of II. The traditional monitor only relies on single indicators such as bispectral index, which may lead to misjudgment due to electromyographic interference or individual pharmacokinetic differences. This example aims to verify that the present application can more accurately and reliably evaluate the true anesthesia state of the patient through multi-modal data fusion and a double verification mechanism.

[0051] In this embodiment, the system continuously acquires the patient's multi-modal physiological signals through the signal acquisition module, including 4-channel real-time electroencephalogram (EEG), frontalis electromyogram (EMG), electrocardiogram (ECG), oxygen saturation signal (SpO2) and non-invasive blood pressure (NIBP). To solve the problem of time inconsistency caused by physiological conduction and drug metabolism delay, the signal synchronization module first extracts the change slope of the oxygen signal, establishes a drug metabolism rate prediction model through pharmacokinetic projection, and calculates the dynamic time delay of each signal channel. Time axis calibration is performed on all multi-modal signals to generate aligned multi-modal signals. During the anesthesia maintenance stage, the feature extraction module performs dynamic path topology analysis on the aligned signals, extracts the cross-modal coupling feature set reflecting the coupling relationship between brain-heart-muscle system, and generates a multi-modal feature vector through entropy fusion. The anesthesia index calculation module constructs a time-domain pharmacodynamic correlation tensor based on this vector, and calculates a continuous and stable anesthesia depth index (ADI) through dynamic aggregation. The index maintains between 40-50 during the stable anesthesia period, indicating an appropriate anesthesia depth. To verify the core advantages of the invention, this embodiment focuses on recording the scene when the surgery is performed to the 52nd minute, when the high abdominal pressure causes strong traction stimulation. At this time, the electroencephalogram index of the traditional monitor only fluctuates slightly from 45 to 48, and does not issue any alarm. However, the system of the invention processes as follows: First, the electromyogram response analysis module detects a transient sharp increase in the patient's electromyogram energy, but the ADI value does not change significantly. The system then generates a dynamic response deviation vector based on the two and identifies a pharmacodynamic sensitive mutation point. By binding with the physiological event of slightly accelerated heart rate, the system generates a high-credibility calibration trigger signal. After time window integration and fusion of the signal, the electromyogram response deviation index rapidly rises from 0.2 to 0.92 within 3 seconds, exceeding the preset safety threshold of 0.8.

[0052] The exceeding of the threshold value immediately triggers the deep analysis process of the system. The real-time time-frequency map generation module performs multi-channel collaborative filtering and time-varying window adaptive decomposition on the real-time electroencephalogram signal, and after phase synchronization and energy normalization reorganization, a high-resolution complex time-frequency component three-dimensional matrix is generated. Subsequently, through dynamic map rendering, a real-time time-frequency map is generated. The pattern matching calculation module matches the real-time time-frequency map with the preset pattern template library. The matching result shows that the similarity score of the current map with the "shallow anesthesia with nociceptive stimulation response" pattern template is as high as 0.95. Based on this score, the anesthesia state classification output module finally outputs the classification result as "anesthesia state: light anesthesia-high body movement risk", and sends a high-level warning to the anesthesiologist. According to this clear classification warning, the anesthesiologist timely adds propofol dose. 5 minutes later, the patient's myoelectric activity returns to stable, the myoelectric response deviation index falls to 0.3, the ADI stabilizes at 42, and the real-time time-frequency map also returns to the slow wave dominant pattern of deep anesthesia. The system successfully avoids a potential intraoperative awareness event.

[0053] Table 1 Anesthesia stable period state monitoring data table (30-35 minutes after the start of the operation) Time Heart rate (bpm) Blood pressure (mmHg) Oxygen saturation (%) Anesthesia depth index (ADI) Classification status T+30min 65 110 / 65 99 48 Suitable anesthesia T+31min 66 108 / 64 99 46 Suitable anesthesia T+32min 65 112 / 66 99 45 Suitable anesthesia T+33min 64 111 / 65 100 44 Suitable anesthesia T+34min 65 109 / 63 99 45 Suitable anesthesia Table 2 System response analysis data table during key stimulation events (52 minutes after the start of the operation) Timestamp (seconds) Event / indicator Value System action / output T+52:00 Tug stimulus Occurrence - T+52:01 Myoelectric signal energy Instantaneous increase of 85% Calculate dynamic response deviation vector T+52:01 Anesthesia depth index (ADI) 46 Deviation analysis start T+52:03 Myoelectric response deviation index 0.92 Exceed safe threshold (0.8) T+52:04 System mode Trigger depth analysis Generate real-time time-frequency spectrum T+52:05 Pattern matching similarity score 0.95 Match with "light anesthesia-harmful stimulus" template T+52:05 Classification status output Anesthesia too light Issue high-level warning Table 3 Performance comparison table of the system of the present application and the traditional monitor Performance indicator The system of the present invention Traditional single-parameter monitor Performance indicator Latent body movement response detection time <5 seconds > 30 seconds or false alarm Latent body movement response detection time State classification accuracy rate (complex working condition) About 96% About 75% State classification accuracy rate (complex working condition) False alarm rate <2% About 10% (myoelectric interference) False alarm rate Warning information clarity Qualitative classification (e.g., light anesthesia) Single value Warning information clarity The above Tables 1 to 3 record the application data of the present application in a real operation monitoring scene, which details the performance of the system in stable period monitoring, key event response and comprehensive performance. As can be seen from Table 1, in the stable period of anesthesia, the system of the present application can output a smooth and reliable anesthesia depth index, providing a stable state reference for the clinic. Table 2 clearly shows the core advantage of the present application: in the case where the traditional index (ADI) does not change significantly, the system successfully captures the subtle stress response that is difficult for the clinician to detect with the naked eye through myoelectric response deviation analysis, and completes the whole process from deviation detection, depth analysis to precise classification warning in just 5 seconds. This fast response mechanism wins valuable time for clinical intervention. The comparison data in Table 3 further highlights the technical superiority of the present application. Compared with the traditional monitor, the system of the present application is faster and more accurate in detecting subtle stimulation responses, and due to the introduction of secondary verification based on pattern matching, the specificity of the warning is strong and the false positive rate is low. By providing qualitative state classification beyond a single numerical value, the present application provides more abundant and more decision-making valuable information for anesthesiologists, greatly improving the safety and accuracy of anesthesia management.

[0054] It should be noted that the electrical connection between the various units described above does not necessarily indicate a direct connection, and the indirect connection mode can also be applied to the embodiments of the present application as long as the purpose of the present application is achieved. The above is only an exemplary embodiment of the present application, and cannot limit the scope of the present application.

[0055] That is, any equivalent changes and modifications made in accordance with the teachings of the present application are still within the scope of the present application. Other embodiments of the present application will be readily apparent to those skilled in the art upon considering the description and practice of the principles disclosed herein. The present application is intended to cover any variations, uses, or adaptive changes of the present application that follow the general principles of the present application and include common knowledge or conventional techniques in the art that are not described in the present application.

Claims

1. A feature fusion processing method of anesthesia depth multi-modal data, characterized in that, The method comprises: obtaining the electromyogram signal of a patient and the multi-modal physiological signal including real-time electroencephalogram signal; time axis calibration is performed on the multi-modal physiological signal to generate aligned multi-modal signal; extraction is performed on the aligned multi-modal signal to generate multi-modal feature vector; based on the multi-modal feature vector, the depth of anesthesia index is calculated; deviation analysis is performed on the electromyogram signal and the depth of anesthesia index to obtain electromyogram response deviation index; graphical feature mapping is performed on the real-time electroencephalogram signal to generate real-time time-frequency spectrum; when the electromyogram response deviation index exceeds the preset safety threshold, the real-time time-frequency spectrum is matched with the preset mode template library to calculate the similarity score; based on the similarity score, the anesthesia state mode corresponding to the mode template library is matched, and the anesthesia state of the current patient is classified and output.

2. The method of claim 1, wherein the method further comprises: The time axis calibration of the multi-modal physiological signal to generate the aligned multi-modal signal comprises: obtaining the blood oxygen signal of the multi-modal physiological signal; extracting the change slope of the blood oxygen signal to establish a drug metabolism rate prediction model; based on the drug metabolism rate prediction model, the signal delay is calculated to generate the time alignment parameter; waveform matching is performed on the multi-modal physiological signal by using the time alignment parameter to generate the aligned multi-modal signal.

3. The method of claim 1, wherein the method further comprises: The calculation of the depth of anesthesia index based on the multi-modal feature vector comprises: based on the multi-modal feature vector, a time-domain drug efficacy correlation tensor is obtained; dynamic aggregation is performed on the time-domain drug efficacy correlation tensor to calculate the depth of anesthesia index.

4. The method of claim 2, wherein the method further comprises: The extraction of the change slope of the blood oxygen signal to establish the drug metabolism rate prediction model comprises: calculating the change slope of the blood oxygen signal to generate a drug concentration gradient feature matrix; pharmacokinetic projection is performed on the drug concentration gradient feature matrix to establish the drug metabolism rate prediction model.

5. The method of claim 1, wherein the method further comprises: The graphical feature mapping of the real-time electroencephalogram signal to generate the real-time time-frequency spectrum comprises: adaptive time-frequency conversion is performed on the real-time electroencephalogram signal to obtain a complex time-frequency component three-dimensional matrix; dynamic spectrum rendering is performed on the complex time-frequency component three-dimensional matrix to generate the real-time time-frequency spectrum.

6. The method of claim 5, wherein the method further comprises: The adaptive time-frequency conversion of the real-time electroencephalogram signal to obtain the complex time-frequency component three-dimensional matrix comprises: multi-channel collaborative filtering and time-varying window adaptive decomposition are performed on the real-time electroencephalogram signal to obtain a three-dimensional complex coefficient tensor; phase synchronization and energy normalization reorganization are performed on the three-dimensional complex coefficient tensor to obtain the complex time-frequency component three-dimensional matrix.

7. The method of claim 1, wherein the method further comprises: The deviation analysis of the electromyogram signal and the depth of anesthesia index to obtain the electromyogram response deviation index comprises: based on the electromyogram signal and the depth of anesthesia index, a dynamic response deviation vector is obtained; event slicing processing is performed on the dynamic response deviation vector to generate a calibration trigger signal; time window integration fusion is performed on the calibration trigger signal to obtain the electromyogram response deviation index.

8. The method of claim 7, wherein the method further comprises: The event slicing processing of the dynamic response deviation vector to generate the calibration trigger signal comprises: based on the dynamic response deviation vector, a drug efficacy sensitive mutation point sequence is obtained; The physiological event binding is performed on the drug-sensitive mutation point sequence to generate a calibration trigger signal.

9. The method of claim 1, wherein the method further comprises: The extracting of the aligned multi-modal signal includes: The dynamic path topology analysis is performed on the aligned multi-modal signal to obtain a cross-modal coupling feature set; The entropy value fusion is performed on the cross-modal coupling feature set to generate a multi-modal feature vector.

10. A feature fusion processing system of anesthetic depth multi-modal data, applied to the feature fusion processing method of anesthetic depth multi-modal data according to any one of claims 1-9, characterized in that, The system includes: A signal acquisition module configured to acquire an electromyography signal of a patient and multi-modal physiological signals including real-time electroencephalography signals; A signal synchronization module configured to perform time axis calibration on the multi-modal physiological signals to generate aligned multi-modal signals; A feature extraction module configured to extract the aligned multi-modal signals to generate a multi-modal feature vector; An anesthesia index calculation module configured to calculate an anesthesia depth index based on the multi-modal feature vector; An electromyography response analysis module configured to perform deviation analysis on the electromyography signal and the anesthesia depth index to obtain an electromyography response deviation index; A real-time time-frequency spectrum generation module configured to perform graphical feature mapping on the real-time electroencephalography signals to generate a real-time time-frequency spectrum; A pattern matching calculation module configured to, when the electromyography response deviation index exceeds a preset safety threshold, perform pattern matching between the real-time time-frequency spectrum and a preset pattern template library to calculate a similarity score; An anesthesia state classification output module configured to, based on the similarity score, match an anesthesia state pattern corresponding to the pattern template library to classify and output an anesthesia state of the patient.

Images