A feature fusion processing method for anesthesia depth multi-modal data

Abstract

The application discloses a feature fusion processing method for anesthetic depth multi-modal data and belongs to the technical field of graphic data processing and pattern recognition, which comprises the following steps: 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 anesthetic depth index; performing deviation analysis based on the electromyographic signals and the index to obtain an electromyographic response deviation index; performing graphic feature mapping on real-time electroencephalogram signals to generate a real-time time-frequency spectrum; when the deviation index exceeds a safety threshold, performing spectrum pattern matching with a pattern template library to calculate a similarity score; and classifying and outputting a current anesthetic state pattern. The application adopts a graphic feature fusion technology of multi-modal signals, combines a dynamic spectrum generation and pattern matching mechanism, can solve a complex pattern recognition problem of physiological signal graphic data, and improves the accuracy and timeliness of anesthetic state classification.

Description

Technical Field

[0001] This invention relates to the field of graphic data processing and pattern recognition technology, and in particular to a feature fusion processing method for multimodal data on anesthesia depth. Background Technology

[0002] In the field of graphic data processing and pattern recognition technology, this invention relates to a collaborative analysis method for multi-source physiological time-series signals. This invention falls under the category of feature fusion processing of multimodal heterogeneous data, specifically targeting a real-time parsing and state classification system for physiological signal graphic data. This type of system achieves dynamic identification of target states by performing pattern matching on time-frequency spectra generated from electromyography (EMG) signals, electroencephalography (EEG) signals, and other physiological parameters.

[0003] Existing technologies primarily achieve state recognition through static template matching, employing predefined EEG atlas databases for fixed pattern comparison. To address temporal differences in multi-channel signals, some systems introduce blood oxygen fluctuation trends to establish a time delay compensation model, utilizing linear interpolation to complete signal alignment. The feature fusion stage commonly employs channel-weighted averaging or principal component analysis to integrate multi-source features, while also setting independent threshold modules to monitor electromyographic activity intensity.

[0004] Existing solutions have significant limitations: First, the signal synchronization mechanism ignores individual differences in drug metabolism kinetics, leading to spatiotemporal misalignment of multi-channel data; second, the feature fusion process fails to establish cross-modal correlation topology, making it unable to suppress the impact of heterogeneous signal interference on evaluation indicators; finally, when abnormal electromyographic activity is detected, traditional atlas matching methods cannot dynamically adjust the evaluation strategy, still relying on static template library output results. These shortcomings stem from the fact that the inherent dynamic correlation of multi-source data has not been effectively modeled. Summary of the Invention

[0005] To address the aforementioned issues, this invention provides a feature fusion processing method for multimodal data on anesthesia depth. By employing graphical feature fusion technology for multimodal signals, combined with dynamic atlas generation and pattern matching mechanisms, this method can solve the complex pattern recognition problem of physiological signal graphical data, thereby improving the accuracy and timeliness of anesthesia state classification.

[0006] The above objectives can be achieved through the following approach:

[0007] A feature fusion processing method for multimodal data on anesthesia depth includes acquiring the patient's multimodal physiological signals and electromyographic (EMG) signals, performing time axis calibration on the physiological signals to generate alignment signals, extracting multimodal feature vectors and calculating an anesthesia depth index, performing deviation analysis based on the EMG signals and the index to obtain an EMG response deviation index, performing graphical feature mapping on real-time EEG signals to generate a real-time time-frequency atlas, and when the deviation index exceeds a safety threshold, performing atlas pattern matching with a pattern template library to calculate a similarity score, and classifying and outputting the current anesthesia state pattern.

[0008] Optionally, the step of time-axis calibration of the multimodal physiological signal to generate an aligned multimodal signal includes: acquiring the blood oxygen signal of the multimodal physiological signal; extracting the slope of the blood oxygen signal change and establishing a drug metabolism rate prediction model; calculating the signal delay based on the drug metabolism rate prediction model and generating time alignment parameters; and using the time alignment parameters to perform waveform matching on the multimodal physiological signal to generate an aligned multimodal signal.

[0009] Optionally, the calculation of the anesthesia depth index based on the multimodal feature vector includes:

[0010] Based on the multimodal feature vectors, a time-domain pharmacodynamic correlation tensor is obtained; the time-domain pharmacodynamic correlation tensor is dynamically aggregated to calculate the anesthetic depth index.

[0011] Optionally, the step of extracting the slope of the blood oxygen signal change and establishing a drug metabolism rate prediction model includes: calculating the slope of the blood oxygen signal change to generate a drug concentration gradient feature matrix; and performing pharmacodynamic projection on the drug concentration gradient feature matrix to establish a drug metabolism rate prediction model.

[0012] Optionally, the step of performing graphical feature mapping on the real-time EEG signal to generate a real-time time-frequency atlas includes: performing adaptive time-frequency conversion on the real-time EEG signal to obtain a complex time-frequency component three-dimensional matrix; and performing dynamic atlas rendering on the complex time-frequency component three-dimensional matrix to generate a real-time time-frequency atlas.

[0013] Optionally, the adaptive time-frequency conversion of the real-time EEG signal to obtain a complex time-frequency component three-dimensional matrix includes: performing multi-channel collaborative filtering and time-varying window adaptive decomposition on the real-time EEG 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 a complex time-frequency component three-dimensional matrix.

[0014] Optionally, the step of performing deviation analysis on the electromyographic signal and the anesthesia depth index to obtain the electromyographic response deviation index includes: obtaining a dynamic response deviation vector based on the electromyographic signal and the anesthesia depth index; performing event-based 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.

[0015] Optionally, the step of performing event-based slicing on the dynamic response deviation vector to generate a calibration trigger signal includes: obtaining a drug-sensitive mutation point sequence based on the dynamic response deviation vector; binding the drug-sensitive mutation point sequence to physiological events to generate a calibration trigger signal.

[0016] Optionally, the step of extracting the aligned multimodal signal to generate a multimodal feature vector includes: performing dynamic path topology analysis on the aligned multimodal signal to obtain a cross-modal coupling feature set; and performing entropy fusion on the cross-modal coupling feature set to generate a multimodal feature vector.

[0017] Based on the same inventive concept, this invention also provides a feature fusion processing system for multimodal data of anesthesia depth. The system includes: a signal acquisition module for acquiring the patient's electromyography (EMG) signals and multimodal physiological signals, including real-time electroencephalography (EEG) signals; a signal synchronization module for time-axis calibration of the multimodal physiological signals to generate aligned multimodal signals; a feature extraction module for extracting features from the aligned multimodal signals to generate multimodal feature vectors; an anesthesia index calculation module for calculating an anesthesia depth index based on the multimodal feature vectors; an EMG response analysis module for performing deviation analysis on the EMG signals and the anesthesia depth index to obtain an EMG response deviation index; a real-time time-frequency atlas generation module for graphical feature mapping of the real-time EEG signals to generate a real-time time-frequency atlas; a pattern matching calculation module for performing pattern matching between the real-time time-frequency atlas and a preset pattern template library when the EMG response deviation index exceeds a preset safety threshold, calculating a similarity score; and an anesthesia state classification output module for matching the anesthesia state patterns corresponding to the pattern template library based on the similarity score, and classifying and outputting the current patient's anesthesia state.

[0018] Compared with the prior art, the present invention has the following advantages:

[0019] This invention establishes a safety calibration mechanism that combines primary monitoring with bypass verification by constructing a dual assessment pathway of anesthesia depth index and electromyographic response deviation index. When the conventional anesthesia depth index may not fully reflect the patient's stress state due to drug characteristics or individual patient differences, independent electromyographic response analysis can promptly capture this mismatch and trigger deeper EEG pattern analysis. This effectively avoids the risk of insufficient anesthesia or intraoperative failure due to the limitations of a single indicator, significantly improving the overall safety and reliability of anesthesia monitoring.

[0020] This invention proposes a timeline calibration method for predicting drug metabolism rates based on dynamic changes in blood oxygenation signals. This method dynamically compensates for individualized time delays in different physiological signals caused by drug metabolism and conduction, based on the patient's real-time physiological feedback, ensuring a high degree of consistency in the time reference of multimodal data before fusion analysis. Compared to methods using fixed delay parameters, this significantly improves the accuracy of subsequent feature extraction and fusion, enabling the final anesthesia depth assessment to more accurately reflect the patient's overall physiological state.

[0021] This invention transcends the analysis of isolated features of single physiological signals by employing advanced feature engineering techniques such as dynamic path topology analysis and time-domain pharmacodynamic correlation tensors. This method focuses on uncovering the dynamic coupling relationships and temporal dependencies of different physiological systems under the influence of anesthesia, extracting cross-modal system-level features that better reveal the essence of the anesthetic state. This deep feature extraction allows the assessment model of anesthetic depth to be built on a more robust physiological mechanism, enhancing the robustness and anti-interference ability of the final indicator.

[0022] This invention, upon detecting abnormal stress signals, transforms one-dimensional real-time EEG signals into an information-rich, pattern-recognition-friendly two-dimensional real-time time-frequency atlas through adaptive time-frequency conversion and dynamic atlas rendering. Combined with a pre-defined pattern template library for high-precision matching, it enables refined classification of anesthesia states. This intelligent upgrade from quantitative assessment to qualitative classification provides clinicians with more intuitive and diagnostically significant decision-making support than single numerical values, enhancing their ability to handle complex and critical situations.

[0023] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures pointed out in the description, claims and drawings. Attached Figure Description

[0024] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0025] Figure 1 This is a flowchart illustrating a feature fusion processing method for multimodal data on anesthesia depth according to an embodiment of the present invention.

[0026] Figure 2 This is a graph of electromyographic response deviation index according to an embodiment of the present invention.

[0027] Figure 3 This is a diagram illustrating the calculation process of the anesthesia depth index according to an embodiment of the present invention.

[0028] Figure 4 This is a time-domain pharmacodynamic correlation tensor diagram according to an embodiment of the present invention.

[0029] Figure 5 This is a real-time time-frequency spectrum diagram according to an embodiment of the present invention.

[0030] Figure 6This is a schematic diagram of the feature fusion processing system for multimodal data of anesthesia depth according to an embodiment of the present invention. Detailed Implementation

[0031] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0032] Reference Figure 1 One embodiment of the present invention proposes a feature fusion processing method for multimodal data of anesthesia depth. It adopts graphical feature fusion technology of multimodal signals, combined with dynamic map generation and pattern matching mechanism, which can solve the complex pattern recognition problem of physiological signal graphical data and improve the accuracy and timeliness of anesthesia state classification.

[0033] The method described in this embodiment specifically includes:

[0034] Acquire patients' electromyographic signals and multimodal physiological signals, including real-time electroencephalogram (EEG) signals;

[0035] The multimodal physiological signals are time-axis calibrated to generate aligned multimodal signals;

[0036] The aligned multimodal signals are extracted to generate multimodal feature vectors;

[0037] Based on the multimodal feature vector, the anesthesia depth index is calculated;

[0038] A deviation analysis was performed on the electromyographic signal and the anesthesia depth index to obtain the electromyographic response deviation index;

[0039] The real-time EEG signals are graphically mapped to generate a real-time time-frequency spectrum.

[0040] When the electromyographic response deviation index exceeds a preset safety threshold, the real-time time-frequency spectrum is matched with a preset pattern template library to calculate a similarity score.

[0041] Based on the similarity score, the anesthesia state pattern corresponding to the pattern template library is matched, and the current anesthesia state of the patient is classified and output.

[0042] Specifically, firstly, at the core assessment layer, multimodal physiological signals from patients, such as EEG, ECG, and blood oxygenation, are collected and innovatively time-axis calibrated to ensure pharmacological synchronization of all signals. Then, features reflecting inter-system coupling are extracted from these aligned signals and fused into a multimodal feature vector. Based on this vector, a comprehensive anesthesia depth index is calculated as a benchmark for assessing anesthesia depth under normal conditions. Secondly, at the alarm and calibration layer, electromyography (EMG) signals are introduced as an independent reference system. Bias analysis compares this EMG signal with the anesthesia depth index derived from EEG, generating an EMG response bias index, as shown below. Figure 2As shown. When this deviation index exceeds the safety threshold, it indicates that the core assessment may be disconnected from the patient's actual physical response, and the system immediately initiates an independent verification mechanism. The safety threshold is the core trigger condition for initiating the atlas matching path. Its determination requires a balance between clinical safety and sensitivity, employing a dynamic mechanism of "basic benchmark value + individualized correction." Specifically, this includes: Determining the basic benchmark value: Based on large-sample clinical data statistics, a general benchmark value is determined by analyzing the correlation between the electromyographic response deviation index and adverse anesthesia events (such as intraoperative awareness, excessive anesthesia) using ROC curve analysis; Individualized adjustment rules: Corrections are made according to the patient's individual characteristics (age, weight, underlying diseases, type of surgery). For example, for elderly patients (≥65 years old), due to decreased neuromuscular sensitivity, the threshold is lowered by 10%-15% (e.g., 0). (6-0.63); For obese patients (BMI≥30), due to the higher baseline EMG signal, the threshold is increased by 5%-8% (e.g., 0.73-0.76); for neurosurgical procedures requiring higher sensitivity, the threshold is decreased by 5% (e.g., 0.66), while for local anesthesia with assisted sedation, the threshold is increased by 10% (e.g., 0.77); Clinical calibration process: A threshold prediction model is constructed using multi-center clinical data. After inputting patient characteristics, an initial threshold is automatically generated, and then fine-tuned by the anesthesiologist within ±0.05 based on clinical experience. Baseline signal verification is performed for 3-5 minutes before surgery (if the baseline deviation index remains higher than the initial threshold, a secondary correction is automatically triggered). This mechanism generates a real-time time-frequency spectrum by processing real-time EEG signals and performs pattern matching with a pattern template library containing various known anesthesia state patterns. The pattern template library is built through training on clinical data and contains standard time-frequency atlas templates and feature vectors corresponding to various typical anesthesia states. The specific construction method includes: Clinical data acquisition: collecting EEG signals from patients of different ages, weights, and surgical types under various anesthesia states (such as stable anesthesia, superficial anesthesia, intraoperative awareness, electromyography interference, etc.), simultaneously recording anesthetic drug dosage, vital signs, and physician-judged anesthesia state labels; Standardized preprocessing: applying the same processing flow (adaptive time-frequency conversion, construction of a three-dimensional matrix of complex time-frequency components, dynamic atlas rendering) to the collected EEG signals to generate standardized time-frequency atlases; Template clustering and labeling: classifying the standardized time-frequency atlases using clustering algorithms (such as K-means), and combining the physician-labeled anesthesia state labels to determine the anesthesia state pattern corresponding to each atlas class (e.g., "stable anesthesia" corresponds to...). Wave energy is dominant. Low activity in the band indicates "insufficient anesthesia". (Band enhancement and increased high-frequency components); Feature vector extraction: Key time-frequency features (such as energy proportion of each frequency band, peak frequency, phase synchronization, etc.) are extracted for each type of standard spectrum to form feature vectors as core template data, which are stored in the pattern template library; Dynamic optimization: New clinical data are periodically incorporated, and the template library is updated through incremental learning to correct template features to adapt to individual differences and changes in clinical scenarios. The structure of the pattern template library covers 6 core anesthesia state patterns, and the specific features are shown in the table below:

[0043]

[0044] The template library employs a multi-dimensional index structure, rapidly matching the feature vectors of real-time time-frequency maps with the template feature vectors using a cosine similarity algorithm. A similarity score ≥80% is considered a successful match, while <50% triggers manual review. Ultimately, based on the matching similarity score, the current patient's anesthesia status is accurately classified and output, enabling the identification and confirmation of abnormal states.

[0045] Optionally, the step of performing time-axis calibration on the multimodal physiological signals to generate aligned multimodal signals includes:

[0046] Obtain the blood oxygenation signal of the multimodal physiological signals;

[0047] Extract the slope of the blood oxygenation signal change and establish a drug metabolism rate prediction model;

[0048] Based on the drug metabolism rate prediction model, the signal delay is calculated, and time alignment parameters are generated.

[0049] The waveform of the multimodal physiological signal is matched using the time alignment parameter to generate an aligned multimodal signal.

[0050] Specifically, firstly, from the acquired multimodal physiological signals, such as electroencephalogram (EEG) and electrocardiogram (ECG) signals, blood oxygen saturation signals are specifically extracted from the blood pressure signals. Changes in blood oxygen saturation, in particular, can indirectly reflect the effects of anesthetic drugs on the patient's circulatory system and tissue perfusion; its rate of change is correlated with the rate of drug diffusion and action in the body. Next, the extracted blood oxygen signals are processed to extract their slope. This slope quantifies the local rate of change of the blood oxygen signal over time and can be considered a dynamic indicator reflecting pharmacokinetics. This slope is calculated to generate a dynamic feature vector related to drug metabolism. To establish a drug metabolism rate prediction model, the core of this model lies in using the aforementioned dynamic feature vector related to drug metabolism to predict the time delay between central nervous system signals and peripheral physiological signals caused by the drug. Specifically, the model uses the slope of the blood oxygen signal as a key input variable to dynamically predict the distribution and metabolic rate of the drug in the body, thereby calculating the expected time delay between different physiological signals. This delay is the time alignment parameter, and the model can be expressed as follows:

[0051] ,

[0052] in, This represents the calculated signal delay, i.e., the time alignment parameter. The slope of the change in the blood oxygen signal, calculated in real time, reflects the immediate dynamics of the drug effect. Baseline physiological parameters, such as age and weight, represent a group of patients and serve as static inputs to the model to improve the accuracy of personalized predictions. The functional form representing a drug metabolism rate prediction model can be a mathematical equation based on physiological principles. The basic linear model is as follows:

[0053] ,

[0054] in This is the signal delay. This is a signal of blood oxygen saturation. This represents the absolute value of the slope of the blood oxygen signal change. For the first Baseline physiological parameters, The model coefficients are represented (determined through fitting clinical data). Finally, waveform matching of the multimodal physiological signals is performed using the calculated dynamically changing time alignment parameters. Dynamic time warping is constrained using the aforementioned time alignment parameters. The search window calculates the minimum distortion path between each signal and the reference EEG signal to achieve waveform matching. The principle of dynamic time warping is to calculate the optimal alignment path between two time series through dynamic programming, minimizing the cumulative distance (such as Euclidean distance). The formula is as follows:

[0055] ,

[0056] in To accumulate distance, It is a metric for the distance between points. This indicates selecting the path with the shortest cumulative distance from three possible preceding positions (i.e., the previous position). The waveform matching process uses one signal as a reference, usually the most direct response to the electroencephalogram (EEG) signal, and then adjusts other physiological signals, such as heart rate and blood pressure, in the time domain by shifting or scaling them according to their respective time alignment parameters. In this way, the waveforms of various signals that were originally biased on the time axis are accurately aligned, ensuring that the multimodal data at the same time point can truly reflect the patient's comprehensive physiological state at that moment. Finally, aligned multimodal signals are generated, providing a high-quality synchronous data foundation for subsequent feature extraction and fusion analysis.

[0057] Optionally, the calculation of the anesthesia depth index based on the multimodal feature vector includes:

[0058] Based on the multimodal feature vectors, the temporal pharmacodynamic correlation tensor is obtained;

[0059] The anesthesia depth index is calculated by dynamically aggregating the time-domain pharmacodynamic correlation tensor.

[0060] Specifically, starting with the multimodal feature vectors obtained in the previous step, these vectors aggregate key information from different physiological signals at each time point. Based on these time-series-based multimodal feature vectors, the system constructs a time-domain pharmacodynamic correlation tensor, as shown in the figure below. Figure 4 As shown. The temporal pharmacodynamic correlation tensor is constructed as follows: a three-dimensional tensor is constructed with a sliding time window as the time dimension and multimodal features as the other two dimensions. The tensor elements are filled by calculating the mutual information value of any two features within the window, quantifying the dynamic correlation between features. Here, the temporal pharmacodynamic correlation tensor is a three-dimensional or higher-dimensional mathematical object used to characterize the relationship between different physiological features over time. Specifically, within a sliding short time window, the correlation strength between any two features in the multimodal feature vector is calculated. For example, indicators such as cross-correlation, mutual information, or phase synchronization index can be used for quantification. These pairwise correlation strength values ​​are combined into a two-dimensional matrix, which completely describes the coupling network between all features at a specific time point. The correlation strength calculation method is based on phase amplitude coupling:

[0061] ,

[0062] in Indicates the change in blood oxygen signal Phase angle in the frequency band (4-8Hz), Indicates brain electrical signals Phase angle in the frequency band (30-50Hz), Indicates the number of sampling points within the time window. It is the imaginary unit. Indicates the first time within a fixed time window The system generates a series of correlation matrices at discrete sampling time points. As the time window moves, these matrices are stacked along the time axis to form a time-domain pharmacodynamic correlation tensor. One dimension of this tensor is time, and the other two dimensions represent different pairs of physiological features. The element values ​​reflect the coupling strength of specific feature pairs at specific times, intuitively presenting the dynamic evolution pattern of the physiological system network under the action of anesthetic drugs. Subsequently, in order to extract a single, intuitive anesthetic depth index from this complex high-dimensional tensor, dynamic aggregation of the time-domain pharmacodynamic correlation tensor is required. The dynamic aggregation of the time-domain pharmacodynamic correlation tensor to calculate the anesthetic depth index includes the following steps: a three-layer LSTM network is used, the tensor is input in time slices, and the number of hidden units in each layer is 64, 32, and 16 respectively. The aggregated value is calculated recursively over time and then mapped to the anesthetic depth index range of 0-100 using the Sigmoid function. Dynamic aggregation is not a simple summation or averaging, but a complex mapping process that projects high-dimensional information onto a one-dimensional scalar. This process aims to assign different weights to different features based on their indicative importance to the depth of anesthesia, and then comprehensively calculate the final depth of anesthesia index. The calculation process of the depth of anesthesia index is shown in the figure below. Figure 3 As shown. The aggregation process can be executed by a pre-trained model function, whose input is the correlation matrix at a specific time point, i.e., a slice of the time-domain pharmacodynamic correlation tensor. This process can be represented as:

[0063] ,

[0064] in, Represents a point in time The obtained anesthesia depth index. It is the time-domain pharmacodynamic correlation tensor in time. The two-dimensional slice, i.e. the feature correlation matrix at that moment. This represents a dynamic aggregation function, which can be a weighted summation based on expert knowledge. The weighted summation formula is as follows:

[0065] ,

[0066] in Indicates time At that time, the first The blood oxygenation characteristic and the first Phase-amplitude coupling strength of individual EEG features Represents the feature combination weight matrix. This represents the dimension of blood oxygenation characteristics. Indicates the time dimension. Represents the dimensions of EEG features. This indicates the rate of increase in drug efficacy. The anesthesia depth exponential function can identify and amplify the coupling patterns most relevant to changes in consciousness level, while suppressing irrelevant or noisy interference, ultimately outputting a continuous value that can accurately quantify the patient's anesthesia depth.

[0067] Optionally, the step of extracting the slope of the blood oxygen signal change and establishing a drug metabolism rate prediction model includes:

[0068] The slope of the change in the blood oxygen signal is calculated to generate a drug concentration gradient feature matrix;

[0069] A pharmacodynamic projection is performed on the drug concentration gradient feature matrix to establish a drug metabolism rate prediction model.

[0070] Specifically, the process begins with calculating the slope of the change in the blood oxygen signal. Blood oxygen saturation is a sensitive indicator reflecting the combined effects of anesthetic drugs on the respiratory and circulatory systems, and its rate of change can indirectly characterize the intensity of the drug effect. By performing temporal difference operations on continuously acquired blood oxygen signals, its instantaneous slope can be obtained. These slope values ​​calculated at consecutive time points. By constructing a time series and selecting data within a specific time window, a drug concentration gradient feature matrix can be generated. Each element of this matrix quantifies the rate of physiological response elicited by the drug at a given moment. Therefore, the matrix as a whole depicts the dynamic spectrum of drug effects over time and can be considered an indirect representation of the drug concentration gradient in the effect room. After generating the drug concentration gradient feature matrix, this method performs pharmacodynamic projection on it to establish a drug metabolism rate prediction model. This pharmacodynamic projection is a model transformation-based analytical method, the core of which is to transform the observed, high-dimensional physiological data, i.e., the drug concentration gradient feature matrix... This is mapped to a low-dimensional pharmacokinetic model space that can describe drug absorption, distribution, and metabolism. This process can be expressed as:

[0071] ,

[0072] in, The predicted drug metabolism rate is the model's output. It is the input drug concentration gradient feature matrix. This is a nonlinear function representing the pharmacodynamic projection. This function itself constitutes the model structure to be established, and its form can be constructed based on classical pharmacological models such as the Emax model or the compartmental model. The pharmacodynamic projection is based on a modified Emax-compartmental hybrid model, and its core function expression is:

[0073] ,

[0074] In the formula These are the time-series values ​​of drug concentration derived from the drug concentration gradient feature matrix; Maximum metabolic rate represents the highest metabolic level that a drug can reach in the body; The half-maximum metabolic concentration is the drug concentration at which the metabolic rate reaches half of the maximum metabolic rate. To eliminate the rate constant, dynamic corrections will be made based on physiological parameters such as the patient's age, weight, and underlying diseases. The time variable is used. The model establishment process is as follows: First, the slope of blood oxygen signal changes, the corresponding drug concentration time series data, and the patients' physiological parameter information are collected from multi-center clinical patients; then, the nonlinear least squares method is used to analyze the above function. The model is fitted with parameters to minimize the deviation between the predicted results and the actual observed data. Finally, the model is optimized through 10-fold cross-validation to ensure that the prediction error of the model in different patient groups is controlled within 8%, thereby obtaining a stable and reliable drug metabolism rate prediction model. These are a set of undetermined parameters for the model, such as the maximum effect rate and the half-maximal effect concentration. These parameters need to be determined by analyzing the actual observed matrices. The estimation is performed by fitting the model's predicted behavior. Parameters are then adjusted using an optimization algorithm. This allows the model output to best reproduce the physiological changes represented by the input matrix; this process completes the projection. The final function obtained... It is a predictive model that takes the rate of change in blood oxygen signal as input and can dynamically output the individualized drug metabolism rate.

[0075] Optionally, the step of graphically mapping the real-time EEG signal to generate a real-time time-frequency atlas includes:

[0076] Adaptive time-frequency conversion is performed on the real-time EEG signal to obtain a complex three-dimensional matrix of time-frequency components;

[0077] Dynamic spectrum rendering is performed on the three-dimensional matrix of the complex time-frequency components to generate a real-time time-frequency spectrum.

[0078] Specifically, an adaptive time-frequency transformation is first performed on the real-time EEG signal. This transformation is an advanced signal processing technique whose core advantage lies in its ability to dynamically adjust the analysis resolution based on the characteristics of the signal itself. For typical non-stationary signals like EEG signals, which contain both slowly changing low-frequency rhythms and sudden high-frequency oscillations, the adaptive time-frequency transformation can analyze different frequency components using different time and frequency windows, thus achieving optimal resolution in both time and frequency. This transformation process decomposes the input real-time EEG signal, and its output is a complex three-dimensional matrix of time-frequency components. The three dimensions of this matrix represent time, frequency, and the EEG signal channel, respectively. Each element in the matrix is ​​a complex number containing the amplitude and phase information of the EEG signal component at a specific time point, frequency, and channel. Next, the obtained complex three-dimensional matrix of time-frequency components is dynamically rendered to generate a visualized real-time time-frequency map. The rendering process mainly focuses on the distribution of signal energy or power in the time-frequency plane. Typically, the power of the signal at a specific time and frequency point is proportional to the square of the amplitude of the corresponding complex component. This process can be represented as:

[0079] ,

[0080] in, Indicates at a point in time and frequency The power spectral density value on. It is extracted from the complex time-frequency component three-dimensional matrix, corresponding to the complex values ​​of a specific time and frequency after the fusion of one or more channels. The rendering step is to process the calculated series of... The values ​​are converted into corresponding colors using a preset color mapping table. For example, high power values ​​correspond to warm colors like red, and low power values ​​correspond to cool colors like blue. Since EEG signals are continuously input, the above conversion and rendering process is also continuously and dynamically performed, thus forming a color image on the display interface that updates in real time, i.e., a real-time time-frequency spectrum. The real-time time-frequency spectrum is as follows: Figure 5 As shown in the figure. The horizontal axis of this spectrum represents time, the vertical axis represents frequency, and the color intensity represents the energy intensity of that frequency component at a specific time.

[0081] Optionally, the adaptive time-frequency conversion of the real-time EEG signal to obtain a complex three-dimensional matrix of time-frequency components includes:

[0082] The real-time EEG signal is subjected to multi-channel collaborative filtering and time-varying window adaptive decomposition to obtain a three-dimensional complex coefficient tensor;

[0083] The three-dimensional complex coefficient tensor is phase-synchronized and energy-normalized for recombination to obtain a three-dimensional matrix of complex time-frequency components.

[0084] Specifically, the real-time EEG signal is first subjected to multi-channel collaborative filtering and time-varying window adaptive decomposition. The collaborative filtering here does not filter each channel independently, but rather utilizes the spatial correlation between multi-channel signals to identify and suppress common noise interference, such as power line interference or motion artifacts, thereby maximizing the signal-to-noise ratio while preserving the true brain neural activity signal. Next, 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 dynamically adjusts the length of the analysis time window based on the characteristics of local frequency components of the signal. For long-duration low-frequency rhythms in the EEG signal, a longer time window is used to obtain high frequency resolution; for transient high-frequency activities, a shorter time window is used to ensure accurate time localization. The output of this decomposition process is a three-dimensional complex coefficient tensor, whose dimensions represent the EEG channel, frequency, and time, respectively. Each element in the tensor is a complex number containing the signal amplitude and phase information of a specific channel at a specific time and frequency point. Then, the generated three-dimensional complex coefficient tensor is reconstructed by phase synchronization and energy normalization. Phase synchronization aims to correct and align phase deviations between different channels caused by signal propagation delays or measurement errors. By calculating the phase consistency across channels within a specific frequency band, the phase information of each channel is adjusted to a common reference standard, which enhances the effective characterization of coordinating neural oscillations in the whole brain or specific brain regions. Energy normalization is then performed to eliminate inherent energy differences between different channels, frequency bands, and individuals, ensuring that subsequent analysis and comparison are not affected by the absolute amplitude of the signal. Normalization maps the energy value of each channel or frequency band to a uniform scale range, for example, through Z-score normalization. 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 time-frequency information but has also undergone cross-channel coordinating processing and normalization, resulting in higher consistency and comparability.

[0085] Optionally, the deviation analysis of the electromyographic signal and the anesthesia depth index to obtain the electromyographic response deviation index includes:

[0086] Based on the electromyographic signal and the anesthesia depth index, a dynamic response deviation vector is obtained;

[0087] The dynamic response deviation vector is processed into event-based slices to generate a calibration trigger signal;

[0088] The calibration trigger signal is integrated and fused over a time window to obtain the electromyographic response deviation index.

[0089] Specifically, the system first generates a dynamic response deviation vector based on real-time acquired electromyography (EMG) signals and a 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 obtained through complex calculations, the two cannot be directly compared. Therefore, the system first preprocesses the EMG signals, for example, by calculating their energy or root mean square value within a sliding time window, to obtain a time-series index that continuously reflects the intensity of muscle activity. Subsequently, the system correlated this electromyographic activity intensity index with the anesthesia depth index. Comparisons are made to generate a dynamic response bias vector. Ideally, deep anesthesia (low...) ) should respond to low electromyographic activity. When low activity occurs But Gao This situation is known as deviation. The deviation vector can be represented by the following functional relationship:

[0090] ,

[0091] in, It is a non-linear function, when Significantly higher than by At the expected baseline level, The value will increase significantly. Next, the system performs event-based slicing on the dynamic response deviation vector to generate a calibration trigger signal. The purpose of this step is to transform continuous deviation values ​​into discrete, meaningful "deviation events." The system sets a deviation threshold; when the dynamic response deviation vector... When the value exceeds this threshold, a significant electromyographic-central mismatch event is considered to have occurred. At this time, the system generates a pulsed calibration trigger signal. The signal value is 1 at the moment the event occurs and 0 at other times. This approach ignores minor, clinically insignificant fluctuations, focusing only on strong bias events that may indicate a patient response. Finally, the system performs time-window integration and fusion on the generated calibration trigger signals to calculate the final electromyographic response bias index. Since a single bias event may be caused by random interference, a stable and reliable index needs to consider the cumulative effect of bias events over a period of time. Therefore, the system integrates or sums the calibration trigger signals within a retrospective time window, as follows:

[0092] ,

[0093] In this formula, The electromyographic response deviation index at the current moment. It is the length of the time window. This refers to calibration trigger signals that occurred within that time window. The index reflects the density or duration of significant deviation events in the recent past.

[0094] Optionally, the step of performing event-based slicing on the dynamic response deviation vector to generate a calibration trigger signal includes:

[0095] Based on the dynamic response deviation vector, the sequence of drug-sensitive mutation points is obtained;

[0096] Physiological events are bound to the drug-sensitive mutation point sequence to generate a calibration trigger signal.

[0097] Specifically, based on the dynamic response bias vector generated in the previous steps, a drug-sensitive mutation point sequence is first identified and extracted. Dynamic response bias vector This is a continuous time series reflecting the real-time difference between electromyographic activity and the anesthesia depth index. The system continuously monitors the first derivative or rate of change of this vector. When it experiences a sudden, abnormally large positive increase far exceeding normal fluctuations within a very short period, and the amplitude of the vector itself also exceeds a preset sensitivity threshold, that time point is marked as a pharmacodynamic sensitivity mutation point. This process aims to capture sudden increases in the deviation signal; such mutations are often more indicative of the patient's instantaneous response to external stimuli (such as surgical incision or traction) or of a rapid drop in anesthetic drug concentration below a critical level than slowly increasing deviations. Through monitoring the entire process... Through analysis, the system obtains a discrete sequence consisting of the timestamps of these mutation points, i.e., the drug-sensitive mutation point sequence. After obtaining this sequence, the system binds it to physiological events to generate the final calibration trigger signal. The core of this step is to correlate abstract mathematical mutation points with real, probable clinical events. The system checks whether there are synchronous changes in other physiological signals, such as a sudden increase in heart rate or blood pressure, within a very short time window before and after each drug-sensitive mutation point. If such a multi-system synchronized stress response pattern exists, the reliability of the mutation point is enhanced, and it is formally confirmed as an event requiring in-depth analysis. The system then converts this confirmed mutation point timestamp into a pulse signal, forming the calibration trigger signal sequence. This binding process acts as a verification and filtering mechanism, using cross-validation of multiple physiological parameters to improve the accuracy of triggering decisions.

[0098] Optionally, the step of extracting the aligned multimodal signal to generate a multimodal feature vector includes:

[0099] Dynamic path topology analysis is performed on the aligned multimodal signals to obtain a cross-modal coupling feature set;

[0100] Entropy fusion is performed on the cross-modal coupled feature set to generate a multimodal feature vector.

[0101] Specifically, dynamic path topology analysis is first performed on these time-synchronized multimodal signals. The dynamic path topology analysis employs a Granger causality analysis algorithm based on a state-space model, as follows: The state-space model state equations are constructed as follows:

[0102] ,

[0103] For state transition moments, For the input matrix, This is the process noise vector. It is an external input vector. State vector ， z Indicates a point in time The system state vector. The observation equation is:

[0104] ,

[0105] These are observations of multimodal physiological signals. For the observation matrix, To observe the noise vector, dynamic path topology analysis goes beyond the isolated examination of a single signal. Instead, it treats physiological signals from different sources, such as EEG, ECG, and blood pressure, as nodes in a dynamic network. This analysis uses computational techniques to quantify the direction, intensity, and time delay of information flow between these nodes within continuous sliding time windows. In this way, the system can construct a functional connectivity network topology that evolves over time. The output of this analysis is a cross-modal coupling feature set, which contains a series of quantitative indicators describing the network topology, such as those ranging from heart rate variability to EEG. Features such as the information transmission intensity of band energy or the phase-locking value between blood pressure fluctuations and brain electrical activity collectively characterize the complex interaction patterns between the central nervous system and the autonomic nervous system under anesthesia. Then, entropy fusion is performed on the obtained cross-modal coupled feature set to generate the final multimodal feature vector. This fusion process aims to intelligently filter and integrate numerous coupled features to highlight the most informative features under the current anesthesia state. Specifically, the system calculates the information entropy of each coupled feature over a period of time. Information entropy is used here as an indicator of the amount of information or determinism contained in a feature; generally, a feature that exhibits a stable and distinguishable pattern under different depths of anesthesia has a lower information entropy and higher information value. Based on this, the system assigns a weight to each coupled feature, which is inversely proportional to the information entropy of that feature. This process can be expressed as:

[0106] ,

[0107] in, It is the generated multimodal feature vector. It is the first after normalization. A cross-modal coupling eigenvalue. It is the entropy weight corresponding to this feature, and its calculation is the same as that of the feature. Information entropy Correlation, the lower the information entropy, the higher the weight. The higher the value, the greater its contribution to the final vector. By multiplying each normalized feature value by its dynamically calculated entropy weight, the system ultimately generates a multimodal feature vector that contains multidimensional coupled information and has been optimized for information value.

[0108] Based on the same inventive concept, such as Figure 6 As shown, the present invention also provides a feature fusion processing system for multimodal data of anesthesia depth, the system comprising:

[0109] The signal acquisition module is used to acquire the patient's electromyographic signals and multimodal physiological signals, including real-time electroencephalogram (EEG) signals.

[0110] The signal synchronization module is used to perform time axis calibration on the multimodal physiological signals to generate aligned multimodal signals;

[0111] The feature extraction module is used to extract the aligned multimodal signal and generate a multimodal feature vector;

[0112] The anesthesia index calculation module is used to calculate the anesthesia depth index based on the multimodal feature vector.

[0113] The electromyography response analysis module is used to perform deviation analysis on the electromyography signal and the anesthesia depth index to obtain the electromyography response deviation index.

[0114] The real-time time-frequency spectrum generation module is used to perform graphical feature mapping on the real-time EEG signal to generate a real-time time-frequency spectrum.

[0115] The pattern matching calculation module is used to perform pattern matching between the real-time time-frequency spectrum and the preset pattern template library when the electromyographic response deviation index exceeds the preset safety threshold, and calculate the similarity score.

[0116] The anesthesia status classification and output module is used to match the anesthesia status pattern corresponding to the pattern template library based on the similarity score, and classify and output the current patient's anesthesia status.

[0117] To verify the feasibility of this invention in practice, it was applied to a general anesthesia surgical monitoring scenario in a tertiary hospital's surgical center. This surgical center aims to improve the accuracy and safety of anesthesia depth monitoring, particularly in addressing individual patient differences and the risk of latent awakening caused by intraoperative stimulation. The method of this invention was integrated into an anesthesia monitoring device for real-time monitoring of a patient undergoing laparoscopic cholecystectomy. The patient was a 55-year-old male, ASA (American Society of Anesthesiologists) class II. Traditional monitors rely solely on single indicators such as the bispectral index of electroencephalography (EEG), which may lead to misjudgments due to electromyographic interference or individual pharmacokinetics differences. This embodiment aims to verify that this invention, through multimodal data fusion and a dual verification mechanism, can more accurately and reliably assess the patient's true anesthesia status.

[0118] In this embodiment, the system continuously acquires the patient's multimodal physiological signals through a signal acquisition module, including 4-channel real-time electroencephalogram (EEG), frontalis muscle electromyography (EMG), electrocardiogram (ECG), blood oxygen saturation (SpO2), and non-invasive blood pressure (NIBP). To address the temporal inconsistency caused by delays in physiological conduction and drug metabolism, the signal synchronization module first extracts the slope of the blood oxygen signal change, establishes a drug metabolism rate prediction model through pharmacodynamic projection, calculates the dynamic time delay of each signal channel, and performs time axis calibration on all multimodal signals to generate aligned multimodal signals. During the anesthesia maintenance phase, the feature extraction module performs dynamic path topology analysis on the aligned signals, extracts a cross-modal coupling feature set reflecting the coupling relationship between the brain-heart-muscle system, and generates a multimodal feature vector through entropy fusion. The anesthesia index calculation module constructs a time-domain pharmacodynamic correlation tensor based on this vector and obtains a continuous and stable depth of anesthesia index (ADI) through dynamic aggregation calculation. This index is maintained between 40 and 50 during the stable anesthesia period, indicating an appropriate depth of anesthesia. To verify the core advantages of this invention, this embodiment focuses on recording the scene at the 52-minute mark of the surgery, where increased pneumoperitoneum pressure triggered intense traction. At this time, the EEG index of a traditional monitor only fluctuated slightly from 45 to 48, without issuing any alarms. However, the system of this invention processes the data in the following steps: First, the electromyography (EMG) response analysis module detects a sudden surge in the patient's EMG signal energy, but the ADI value does not change significantly. The system then generates a dynamic response deviation vector based on both and identifies a drug-sensitive mutation point. By binding this to a physiological event of a slight increase in heart rate, the system generates a highly reliable calibration trigger signal. After time-window integration and fusion of this signal, the EMG response deviation index rapidly climbs from 0.2 to 0.92 within 3 seconds, exceeding the preset safety threshold of 0.8.

[0119] Exceeding this threshold immediately triggered the system's deep analysis process. The real-time time-frequency atlas generation module performed multi-channel collaborative filtering and time-varying window adaptive decomposition on the real-time EEG signal, and after phase synchronization and energy normalization reconstruction, generated a high-resolution complex three-dimensional matrix of time-frequency components. Subsequently, a real-time time-frequency atlas was generated through dynamic atlas rendering. The pattern matching calculation module matched this real-time time-frequency atlas with a preset pattern template library. The matching results showed that the similarity score between the current atlas and the "light anesthesia with noxious stimulus response" pattern template was as high as 0.95. Based on this score, the anesthesia status classification output module finally output the classification result as "anesthesia status: light anesthesia - high risk of body movement" and issued a high-level warning to the anesthesiologist. Based on this clear classification warning, the anesthesiologist promptly added propofol. Five minutes later, the patient's electromyographic activity returned to stability, the electromyographic response deviation index dropped back to 0.3, the ADI stabilized at 42, and the real-time time-frequency atlas also returned to the slow-wave dominant pattern of deep anesthesia. The system successfully avoided a potential intraoperative awareness event.

[0120] Table 1. Monitoring data of the stable state during anesthesia (30-35 minutes after the start of surgery)

[0121] time Heart rate (bpm) Blood pressure (mmHg) Blood 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

[0122] Table 2. System response analysis data during key stimulus events (52 minutes after the start of surgery).

[0123] Timestamp (seconds) Events / Indicators numerical values System Actions / Outputs T+52:00 traction stimulation occur - T+52:01 Electromyographic signal energy Instantaneous increase of 85% Calculate the dynamic response deviation vector T+52:01 Anesthesia Depth Index (ADI) 46 Deviation analysis initiated T+52:03 Electromyographic response deviation index 0.92 Exceeding the safety threshold (0.8) T+52:04 System Mode Triggering Deep Analysis Generate real-time time-frequency spectrum T+52:05 Pattern matching similarity score 0.95 Matching the "light anesthesia-noxious stimulus" template T+52:05 Classification status output Anesthesia too weak Issue a high-level warning

[0124] Table 3. Performance Comparison of the Invention System and Traditional Patient Monitors

[0125] Performance indicators This invention system Traditional single-parameter monitor Performance indicators Latent body movement response detection time <5 seconds >30 seconds or missed report Latent body movement response detection time Status classification accuracy (complex working conditions) Approximately 96% Approximately 75% Status classification accuracy (complex working conditions) False alarm rate <2% Approximately 10% (electromyographic interference) False alarm rate Clarity of early warning information Qualitative classification (e.g., light anesthesia) Single value Clarity of early warning information

[0126] Tables 1 to 3 above record the application data of this invention in real surgical monitoring scenarios, detailing the system's performance in stable-phase monitoring, critical event response, and overall performance. Table 1 shows that during the stable anesthesia phase, the system can output a smooth and reliable anesthesia depth index, providing a stable state reference for clinicians. Table 2 clearly demonstrates the core advantages of this invention: when the traditional indicator (ADI) does not change significantly, the system successfully captures latent stress responses that are difficult for clinicians to detect visually through electromyographic response deviation analysis, completing the entire process from deviation detection and depth analysis to accurate classification and early warning within just 5 seconds. This rapid response mechanism buys valuable time for clinical intervention. The comparative data in Table 3 further highlights the technical superiority of this invention. Compared to traditional monitors, the system of this invention is faster and more accurate in detecting latent stimulus responses, and due to the introduction of pattern-matching-based secondary validation, its early warnings are highly specific and have a low false alarm rate. By providing qualitative state classification beyond a single numerical value, this invention provides anesthesiologists with richer and more decision-making-valuable information, greatly improving the safety and accuracy of anesthesia management.

[0127] It should be noted that the electrical connections between the various units described above do not necessarily represent direct or indirect connections. Any indirect connection method can be applied to the embodiments of the present invention as long as it achieves the purpose of the present invention. The above descriptions are merely exemplary embodiments of the present invention and should not be construed as limiting the scope of the present invention.

[0128] All equivalent changes and modifications made in accordance with the teachings of this invention are still within the scope of this invention. Those skilled in the art will readily conceive of other embodiments of this invention upon considering the specification and the disclosure of practical truth. This application is intended to cover any variations, uses, or adaptations of this invention that follow the general principles of this invention and include common knowledge or conventional techniques in the art not described herein.

Claims

1. A feature fusion processing method for multimodal data on anesthesia depth, characterized in that, The method includes: Acquire patients' electromyographic signals and multimodal physiological signals, including real-time electroencephalogram (EEG) signals; The multimodal physiological signals are time-axis calibrated to generate aligned multimodal signals; The aligned multimodal signals are extracted to generate multimodal feature vectors; Based on the multimodal feature vector, the anesthesia depth index is calculated; A deviation analysis is performed on the electromyographic signal and the anesthesia depth index to obtain an electromyographic response deviation index, including: obtaining a dynamic response deviation vector based on the electromyographic signal and the anesthesia depth index; performing event-based slicing on the dynamic response deviation vector to generate a calibration trigger signal; and performing time-window integration and fusion on the calibration trigger signal to obtain the electromyographic response deviation index, including: obtaining a drug-sensitive mutation point sequence based on the dynamic response deviation vector; and binding the drug-sensitive mutation point sequence to physiological events to generate a calibration trigger signal. The real-time EEG signal is graphically feature-mapped to generate a real-time time-frequency spectrum, including: adaptive time-frequency transformation of the real-time EEG signal to obtain a complex time-frequency component three-dimensional matrix; dynamic spectrum rendering of the complex time-frequency component three-dimensional matrix to generate a real-time time-frequency spectrum, including: multi-channel collaborative filtering and time-varying window adaptive decomposition of the real-time EEG signal to obtain a three-dimensional complex coefficient tensor; phase synchronization and energy normalization recombination of the three-dimensional complex coefficient tensor to obtain a complex time-frequency component three-dimensional matrix; When the electromyographic response deviation index exceeds a preset safety threshold, the real-time time-frequency spectrum is matched with a preset pattern template library to calculate a similarity score. Based on the similarity score, the anesthesia state pattern corresponding to the pattern template library is matched, and the current anesthesia state of the patient is classified and output.

2. The feature fusion processing method for multimodal data of anesthesia depth according to claim 1, characterized in that, The step of performing time-axis calibration on the multimodal physiological signals to generate aligned multimodal signals includes: Obtain the blood oxygenation signal of the multimodal physiological signals; Extract the slope of the blood oxygenation signal change and establish a drug metabolism rate prediction model; Based on the drug metabolism rate prediction model, the signal delay is calculated, and time alignment parameters are generated. The waveform of the multimodal physiological signal is matched using the time alignment parameter to generate an aligned multimodal signal.

3. The feature fusion processing method for multimodal data of anesthesia depth according to claim 1, characterized in that, The anesthesia depth index calculated based on the multimodal feature vector includes: Based on the multimodal feature vectors, the temporal pharmacodynamic correlation tensor is obtained; The anesthesia depth index is calculated by dynamically aggregating the time-domain pharmacodynamic correlation tensor.

4. The feature fusion processing method for multimodal data of anesthesia depth according to claim 2, characterized in that, The step of extracting the slope of the blood oxygen signal change and establishing a drug metabolism rate prediction model includes: The slope of the change in the blood oxygen signal is calculated to generate a drug concentration gradient feature matrix; A pharmacodynamic projection is performed on the drug concentration gradient feature matrix to establish a drug metabolism rate prediction model.

5. The feature fusion processing method for multimodal data of anesthesia depth according to claim 1, characterized in that, The step of extracting the aligned multimodal signal to generate a multimodal feature vector includes: Dynamic path topology analysis is performed on the aligned multimodal signals to obtain a cross-modal coupling feature set; Entropy fusion is performed on the cross-modal coupled feature set to generate a multimodal feature vector.

6. A feature fusion processing system for multimodal data of anesthesia depth, applied to a feature fusion processing method for multimodal data of anesthesia depth as described in any one of claims 1-5, characterized in that, The system includes: The signal acquisition module is used to acquire the patient's electromyographic signals and multimodal physiological signals, including real-time electroencephalogram (EEG) signals. The signal synchronization module is used to perform time axis calibration on the multimodal physiological signals to generate aligned multimodal signals; The feature extraction module is used to extract the aligned multimodal signal and generate a multimodal feature vector; The anesthesia index calculation module is used to calculate the anesthesia depth index based on the multimodal feature vector. The electromyography (EMG) response analysis module is used to perform deviation analysis on the EMG signal and the anesthesia depth index to obtain an EMG response deviation index. This includes: obtaining a dynamic response deviation vector based on the EMG signal and the anesthesia depth index; performing event-based slicing on the dynamic response deviation vector to generate a calibration trigger signal; and performing time-window integration and fusion on the calibration trigger signal to obtain the EMG response deviation index. This includes: obtaining a drug-sensitive mutation point sequence based on the dynamic response deviation vector; and binding the drug-sensitive mutation point sequence to physiological events to generate a calibration trigger signal. A real-time time-frequency atlas generation module is used to perform graphical feature mapping on the real-time EEG signal to generate a real-time time-frequency atlas. This includes: performing adaptive time-frequency transformation on the real-time EEG signal to obtain a complex three-dimensional matrix of time-frequency components; and performing dynamic atlas rendering on the complex three-dimensional matrix of time-frequency components to generate a real-time time-frequency atlas. This includes: performing multi-channel collaborative filtering and time-varying window adaptive decomposition on the real-time EEG signal to obtain a three-dimensional complex coefficient tensor; and performing phase synchronization and energy normalization recombination on the three-dimensional complex coefficient tensor to obtain a complex three-dimensional matrix of time-frequency components. The pattern matching calculation module is used to perform pattern matching between the real-time time-frequency spectrum and the preset pattern template library when the electromyographic response deviation index exceeds the preset safety threshold, and calculate the similarity score. The anesthesia status classification and output module is used to match the anesthesia status pattern corresponding to the pattern template library based on the similarity score, and classify and output the current patient's anesthesia status.

Images