Anesthesia early warning method based on multi-source signal fusion

Abstract

The invention discloses an anesthesia early warning method based on multi-source signal fusion. The anesthesia early warning method comprises the following steps that physiological signals in the anesthesia process are collected in real time and preprocessed; a multi-modal Transform attention network is adopted to extract correlation features in the time sequence physiological signals, and multi-modal fusion features are obtained; predicting the fusion feature at the next moment by using a long-short-term memory neural network to obtain a prediction error; updating a noise and observation covariance matrix of the Kalman filter according to the prediction error; the dynamically adjusted Kalman filter calibrates the multi-modal fusion features in real time; according to the calibration characteristics, anesthesia depth and consciousness state prediction values are calculated in real time, and early warning is output. The real-time performance, the stability and the accuracy of anesthesia state monitoring are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of anesthesia monitoring, and in particular to an anesthesia early warning method based on multi-source signal fusion. Background Art

[0002] With the development of medical monitoring technology, anesthesia status monitoring has gradually become an important part of ensuring the safety of patients during surgery and the quality of their postoperative recovery. At present, clinical anesthesia monitoring technology uses a single modality indicator based on EEG signals as a criterion for determining the state of consciousness. By analyzing the EEG signals collected from the patient's scalp and extracting their spectral characteristics, the patient's anesthesia depth can be assessed. However, EEG signals are extremely susceptible to factors such as poor electrode contact, electromyographic artifacts, and individual patient differences. Early warning systems that rely solely on EEG signal analysis are prone to delays, misjudgments, and instability.

[0003] In recent years, researchers have attempted to use multi-source physiological signals for fusion analysis to improve the accuracy of anesthesia state assessment. These methods collect multi-source signals such as the patient's EEG, ECG, blood pressure, and blood oxygen saturation signals separately, and use traditional data fusion technology to perform simple combination or weighted averaging to obtain more reliable anesthesia depth judgment. However, most of these existing technical solutions use static fusion methods with fixed fusion weights, fail to achieve adaptive adjustment for different patients and clinical environments, make it difficult to dynamically respond to real-time changes in anesthesia status, and cannot effectively suppress the adverse effects of clinical interference factors on the fusion results.

[0004] Furthermore, while some existing methods attempt to use neural networks or Kalman filtering techniques to process multi-source physiological signals, they often fail to incorporate the temporal correlations between modal signals and lack in-depth modeling and dynamic analysis of the intrinsic relationships between modalities. Consequently, existing anesthesia monitoring methods are insufficiently sensitive to abnormal fluctuations, and early warning signals experience significant latency, making them unable to meet the stringent real-time and accuracy requirements of clinical anesthesia safety.

[0005] Therefore, how to provide an anesthesia early warning method based on multi-source signal fusion is an urgent problem that those skilled in the art need to solve. Summary of the Invention

[0006] One purpose of the present invention is to propose an anesthesia warning method based on multi-source signal fusion. In response to the problems in the existing technology that the multi-source physiological signal fusion method is static, cannot adapt to dynamic interference, and has delayed warning response, a real-time anesthesia monitoring method using a multimodal Transformer attention network combined with a long short-term memory neural network and a Kalman filter dynamic fusion is proposed. The present invention has the advantages of high real-time warning, strong anti-interference ability and significantly improved accuracy.

[0007] An anesthesia early warning method based on multi-source signal fusion according to an embodiment of the present invention includes:

[0008] Real-time collection of physiological signals during anesthesia of the subjects, and processing to obtain time-series physiological signal data;

[0009] The multimodal Transformer attention network is used to extract the time-correlation features and cross-correlation features within the temporal physiological signal data to obtain multimodal fusion features.

[0010] A long short-term memory neural network is used to predict the fusion features of the next moment based on the multimodal fusion features of the current moment, and the predicted fusion feature value is obtained. The predicted value is then compared with the actual multimodal fusion features of the next moment to obtain the prediction error.

[0011] Based on the prediction error, update the noise covariance matrix and observation covariance matrix of the Kalman filter;

[0012] The multimodal fusion features output by the multimodal Transformer attention network are calibrated using the adjusted Kalman filter to generate calibrated multimodal fusion features;

[0013] The predicted values ​​of the subject's anesthesia depth and consciousness state are calculated based on the calibrated multimodal fusion features. When the predicted values ​​exceed the pre-set anesthesia depth safety threshold or consciousness recovery risk threshold, an anesthesia warning signal is output.

[0014] Optionally, the real-time collection of physiological signals during anesthesia of the subject and the processing to obtain time-series physiological signal data are specifically:

[0015] The EEG signals were collected from multiple preset electrode positions on the subject's scalp, and the myoelectric artifacts were removed by bandpass filtering. The adaptive mean filtering method was used to remove the slowly varying baseline drift to obtain time-series EEG signal data.

[0016] The ECG signal of the subject's chest is collected at the standard lead position, high-frequency environmental noise is eliminated through a bandpass filter, and nonlinear baseline drift is removed using the wavelet transform method to obtain time-series ECG signal data;

[0017] Invasive blood pressure signals are collected from the radial or brachial artery of the subject. Spike interference is eliminated using a median filter, and baseline drift is removed using a low-pass filter to obtain time-series blood pressure signal data.

[0018] The blood oxygen saturation signal at the fingertips of the subjects is collected in real time. High-frequency environmental noise and motion artifacts are removed through a bandpass filter, and periodic baseline drift is dynamically removed using a sliding average filter to obtain time-series blood oxygen saturation signal data.

[0019] The time-series EEG signal data, time-series ECG signal data, time-series blood pressure signal data, and time-series blood oxygen saturation signal data are synchronized to a unified time series, and the physiological signal data of each channel are adjusted through an interpolation algorithm based on a cubic spline function to obtain the time-series physiological signal data.

[0020] Optionally, the multimodal fusion features are obtained through the multimodal Transformer attention network, specifically:

[0021] The time series physiological signal data are segmented using sliding windows with fixed length and time overlap ratio to obtain physiological signal segment sequences of each signal data;

[0022] Perform multi-scale self-attention calculations simultaneously on data points in any sequence of physiological signal segments of each signal data to obtain the internal attention features of each modality;

[0023] Using the EEG signal segment sequence as the dominant modality reference sequence, a cross-attention weight sequence between the EEG signal and other physiological signals is generated. Using the ECG signal, blood pressure signal, and blood oxygen saturation signal segment sequences as the dominant modality reference sequences, a cross-attention weight sequence between each modality is obtained.

[0024] According to the internal attention features of each modality and the cross-attention weight sequence between each modality, the internal features of each modality and the cross-modal interaction features are fused to obtain a fused feature sequence;

[0025] The correlation entropy value of each feature in the fusion feature sequence is calculated, and the fusion feature with higher entropy value is selected as the effective feature to obtain the multimodal fusion feature.

[0026] Optionally, the multimodal Transformer attention network includes a modality-independent attention feature encoding unit, a self-attention unit, a modality-interaction attention calculation unit, a cross-modality attention timing correction unit, and an attention weight adaptive update unit:

[0027] The modality-independent attention feature encoding unit performs feature mapping on the data points in each modality signal sequence to generate an initial feature vector sequence;

[0028] The self-attention unit uses multiple independent and parallel multi-head attention calculation mechanisms to perform self-attention calculations on the data points within the modality for the initial feature vector sequence, thereby obtaining the time correlation features within each modal signal;

[0029] The modal interaction attention calculation unit uses the time correlation features of each modal signal as the dominant modal reference sequence in a rotation manner, and performs cross-modal attention calculation with the time correlation features of other modal signals one by one to obtain cross-modal cross-attention weights;

[0030] The cross-modal attention timing correction unit performs real-time correction on the cross-modal attention weight whose change rate exceeds a preset threshold through a timing correction method based on sliding window mean calculation, and outputs the cross-modal attention weight;

[0031] The attention weight adaptive update unit calculates the attention entropy value of the internal time correlation feature of each modal signal and the attention entropy value of the cross-modal cross-attention weight, and adjusts the contribution weight of each modal signal to the overall fusion feature in real time according to the attention entropy value.

[0032] Optionally, the prediction error is obtained by using a long short-term memory neural network, specifically:

[0033] The multimodal fusion features at the current moment are input into a long short-term memory neural network with a three-layer stacked structure to predict the fusion feature prediction value at the next moment;

[0034] Calculate the difference between the fusion feature prediction value and the actual multimodal fusion feature one by one to generate a feature prediction residual sequence that corresponds one to one with the actual multimodal fusion feature;

[0035] A residual statistical analysis method based on a dynamic adaptive sliding window is used. The sliding window length is fixed as a benchmark and the sliding window length is dynamically adjusted according to the residual change rate calculated in real time.

[0036] The absolute difference between the current fusion feature residual value and the residual mean value in the corresponding dynamic sliding window is compared in real time. When the absolute difference is greater than the abnormal threshold, the fusion feature residual value is identified as a mutation or abnormal interference in the fusion feature sequence.

[0037] For mutations or abnormal interferences in the identified fusion features, the information corresponding to the abnormal features is recorded in real time to obtain the prediction error.

[0038] Optionally, the updating of the Kalman filter is specifically as follows:

[0039] The absolute value of the prediction error corresponding to each fusion feature in the prediction error is used as the basic data of the noise covariance matrix and the observation covariance matrix of the Kalman filter;

[0040] According to the real-time prediction error absolute value corresponding to each fusion feature in the prediction error sequence, the prediction error absolute value mean and prediction error absolute value variance of each fusion feature prediction error absolute value in a sliding time window of fixed length are calculated and mapped respectively;

[0041] According to the noise covariance matrix and observation covariance matrix updated at the current moment, the Kalman filter's state prediction equation and state update equation are used to calculate the Kalman gain matrix between the fusion feature estimate and the actual fusion feature measurement value;

[0042] Calculate the value change rate of the diagonal elements of the Kalman gain matrix between two consecutive moments, and determine whether the value change rate exceeds the preset Kalman gain change rate threshold, and identify the obtained abnormally sensitive fusion features;

[0043] The Kalman filter is dynamically adjusted based on the abnormally sensitive fusion features obtained through identification.

[0044] Optionally, the multimodal fusion features output by the multimodal Transformer attention network are calibrated using the adjusted Kalman filter to generate calibrated multimodal fusion features, specifically:

[0045] Input the multimodal fusion features into the dynamically adjusted Kalman filter for state calibration;

[0046] Based on the real-time updated noise covariance matrix and observation covariance matrix, the state prediction equation of the Kalman filter is used to generate the current moment's fusion feature prediction state value by using the matrix product between the calibrated multimodal fusion features output at the previous moment and the state transfer matrix;

[0047] The difference between the current fusion feature prediction state value and the multimodal fusion feature value is calculated in matrix form one by one to obtain the current fusion feature state prediction error value. The fusion feature state prediction error value is then weighted corrected using the real-time updated Kalman gain matrix to obtain the calibrated multimodal fusion feature estimation value.

[0048] Calculate and record the fluctuation amplitude and fluctuation rate of the multimodal fusion feature estimation value at multiple consecutive moments to identify abnormal fluctuations in the fusion feature caused by clinical interference factors;

[0049] The abnormal fluctuation suppression mechanism is used to robustly correct the abnormal fluctuation of the real-time identified fusion features to obtain the calibrated multimodal fusion features.

[0050] Optionally, the calibrated multimodal fusion feature is used to calculate the predicted value of the subject's anesthesia depth and consciousness state, and obtain an anesthesia warning signal, specifically:

[0051] Inputting the calibrated multimodal fusion features into a real-time anesthesia depth and consciousness state calculation unit, wherein the real-time anesthesia depth and consciousness state calculation unit adopts a multi-layer feedforward neural network with a fully connected structure;

[0052] The output layer of the multi-layer feedforward neural network is equipped with two independent output neurons:

[0053] The first output neuron outputs the predicted value of the anesthesia depth at the current moment in real time.

[0054] The second output neuron outputs the predicted value of the current state of consciousness in real time;

[0055] A real-time threshold determination method is used to analyze the predicted values ​​of anesthesia depth and consciousness state in real time, and the consciousness recovery risk threshold and anesthesia depth safety threshold are obtained respectively;

[0056] The anesthesia depth prediction value is numerically compared with the anesthesia depth safety threshold in real time, and the consciousness state prediction value is numerically compared with the consciousness recovery risk threshold in real time. When at least one of the anesthesia depth prediction value exceeds the anesthesia depth safety threshold or the consciousness state prediction value exceeds the consciousness recovery risk threshold, a detailed anesthesia warning signal is generated including the subject number, the current abnormal prediction value, the specific time when the prediction value exceeds the threshold and the corresponding threshold value, and anesthesia warning is performed.

[0057] The beneficial effects of the present invention are:

[0058] (1) The present invention realizes the real-time feature fusion and prediction of multi-source physiological signals through a dynamic fusion strategy combining a multimodal Transformer attention network and a long short-term memory neural network, effectively improving the real-time and accuracy of anesthesia status monitoring, and enhancing the stability and anti-interference ability of anesthesia status warning under clinical interference conditions.

[0059] (2) The present invention realizes the real-time and accurate identification and dynamic calibration of abnormal fluctuations of fusion features through the real-time dynamic update mechanism of the Kalman filter noise covariance matrix and the observation covariance matrix, significantly improves the reliability of the anesthesia status warning signal, and shows better adaptability in complex clinical anesthesia environments.

[0060] (3) In terms of real-time analysis of anesthesia consciousness state, the present invention effectively solves the problems of early warning delay and high misjudgment rate in the existing technology through real-time calculation of prediction error and abnormal fluctuation suppression mechanism based on residual statistical analysis, breaks through the limitations of static and insufficient self-adaptation of fusion weights in the existing technology, and realizes specific and significant improvements in real-time monitoring of anesthesia depth and consciousness state, effectively improving the level of clinical anesthesia safety control. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0062] Figure 1 This is an overall flow chart of an anesthesia early warning method based on multi-source signal fusion proposed by the present invention. DETAILED DESCRIPTION

[0063] The present invention will now be described in further detail with reference to the accompanying drawings, which are simplified schematic diagrams that illustrate the basic structure of the present invention in a schematic manner.

[0064] refer to Figure 1 , an anesthesia early warning method based on multi-source signal fusion, comprising:

[0065] The system collects the subjects' EEG signals, ECG signals, blood pressure signals and blood oxygen saturation signals in real time during anesthesia, and performs preprocessing such as filtering, removing baseline drift and removing physiological interference on each channel of collected physiological signals to obtain time-series physiological signal data for feature fusion.

[0066] Each channel of pre-processed time-series physiological signal data is input into the multimodal Transformer attention network to extract the temporal correlation features within each physiological signal and the cross-correlation features between EEG signals, ECG signals, blood pressure signals, and blood oxygen saturation signals, thereby obtaining dynamically updated multimodal fusion features.

[0067] A long short-term memory neural network is used to predict the fusion features of the next moment based on the multimodal fusion features at the current moment to obtain a fusion feature prediction value, and the fusion feature prediction value is compared with the actual multimodal fusion features at the next moment to obtain a prediction error, which is used to identify mutations or abnormal interference in the fusion features;

[0068] Based on the prediction error, the noise covariance matrix and the observation covariance matrix of the Kalman filter are updated in real time, and the fusion weight of the Kalman filter is dynamically adjusted to improve the sensitivity and suppression ability of abnormal fusion features;

[0069] The multimodal fusion features output by the multimodal Transformer attention network are calibrated in real time using a dynamically adjusted Kalman filter to generate robustly calibrated multimodal fusion features to overcome fluctuations in fusion features caused by clinical interference factors.

[0070] Based on the robustly calibrated multimodal fusion features, the predicted values ​​of the subject's anesthesia depth and consciousness state are calculated in real time. When the predicted values ​​exceed the pre-set anesthesia depth safety threshold or consciousness recovery risk threshold, an anesthesia warning signal is output.

[0071] In this embodiment, the subject's EEG signals, ECG signals, blood pressure signals, and blood oxygen saturation signals are collected in real time during anesthesia, and each of the collected physiological signals is filtered, baseline drift removed, and physiological interference removed to obtain time-series physiological signal data for feature fusion, specifically:

[0072] The method collects EEG signals from multiple preset electrode positions on the subject's scalp in real time, removes myoelectric artifacts from the EEG signals through a bandpass filter, and uses an adaptive mean filtering method based on signal statistical characteristics to remove slowly varying baseline drift in the EEG signals, thereby obtaining time-series EEG signal data free of myoelectric interference and with a stable baseline. The adaptive mean filtering method based on signal statistical characteristics comprises: calculating the mean and standard deviation of the currently collected EEG signal data within a sliding window in real time, dynamically adjusting the sliding window length and the mean filter coefficient based on the standard deviation, and updating the filter output value of the current data point, thereby adaptively eliminating slowly varying baseline drift.

[0073] The method collects ECG signals from standard lead positions on the subject's chest in real time, eliminates high-frequency environmental noise in the ECG signals through a bandpass filter with a limited frequency range, and uses a wavelet transform method optimized by R-wave adaptive positioning to remove nonlinear baseline drift in the ECG signals, thereby obtaining time-series ECG signal data with continuous waveforms and stable baselines. The wavelet transform method optimized by R-wave adaptive positioning specifically comprises: detecting and locating the R-wave position in the current ECG signal in real time, dynamically adjusting the wavelet decomposition scale and threshold parameters according to the time interval between two adjacent R-wave intervals, and decomposing and reconstructing the ECG signal by adaptively selecting wavelet scale coefficients, thereby eliminating nonlinear baseline drift in real time.

[0074] Invasive blood pressure signals from the radial or brachial artery of a subject are collected in real time, spike and burr interference in the blood pressure signal is eliminated in real time through a median filtering method with adaptive window length adjustment, and baseline drift in the blood pressure signal is removed in real time using a low-pass filter selected by a cutoff frequency, so as to obtain time-series blood pressure signal data without burr interference and with a stable baseline, wherein the median filtering method with adaptive window length adjustment specifically comprises: real-time calculation of the data change rate of the current blood pressure signal within a sliding window, adjustment of the length of the filter window according to the data change rate, shortening the window length to quickly respond to spike interference when the data change rate is large, and extending the window length to suppress random noise when the data change rate is small, and the low-pass filter with cutoff frequency selection specifically comprises: determining the frequency band range of the baseline drift of the current blood pressure signal based on spectrum analysis of the blood pressure signal, and setting the cutoff frequency of the low-pass filter within this frequency band;

[0075] The blood oxygen saturation signal at the fingertips of the subject is collected in real time. High-frequency environmental noise and motion artifact interference in the blood oxygen saturation signal are removed in real time using a bandpass filter with a limited bandwidth. A sliding average filtering method based on the signal periodicity characteristics is used to dynamically remove periodic baseline drift in the blood oxygen saturation signal, thereby obtaining time-series blood oxygen saturation signal data with low interference and a stable baseline. The sliding average filtering method based on the signal periodicity characteristics specifically comprises: detecting the pulse wave peak position of the blood oxygen saturation signal, calculating the average period between consecutive adjacent pulse wave peaks, and dynamically adjusting the window length of the sliding average filter based on the currently calculated pulse wave average period, so that the filter window adapts in real time to the actual periodic changes of the blood oxygen saturation signal.

[0076] The acquired time-series EEG signal data, time-series ECG signal data, time-series blood pressure signal data, and time-series blood oxygen saturation signal data are synchronized to a unified time series based on the EEG signal sampling time. The sampling frequency and data length of each physiological signal data are dynamically adjusted in real time through an interpolation algorithm based on a cubic spline function to obtain time-series physiological signal data.

[0077] In this embodiment, the obtained pre-processed temporal physiological signal data of each channel is input into the multimodal Transformer attention network respectively, and the time correlation features within each physiological signal and the cross-correlation features between the EEG signal, ECG signal, blood pressure signal and blood oxygen saturation signal are extracted respectively to obtain the dynamically updated multimodal fusion features, specifically:

[0078] The preprocessed time-series physiological signal data, including time-series EEG signal data, time-series ECG signal data, time-series blood pressure signal data, and time-series blood oxygen saturation signal data, are segmented using sliding windows with fixed lengths and fixed time overlap ratios to obtain physiological signal segment sequences with stable time-series structures.

[0079] Each physiological signal segment sequence is input into the self-attention unit set up independently and in parallel in the multimodal Transformer attention network. A multi-head attention calculation mechanism is set up in each self-attention unit. The multi-head attention calculation mechanism performs multi-scale self-attention calculations on the data points in any physiological signal segment sequence simultaneously:

[0080]

[0081] in, is the attention weight of the i-th query data point to the j-th key data point in the h-th attention head, is the feature vector of the i-th query data point in the h-th attention head, is the feature vector of the jth key data point in the hth attention head, τi,j is the actual sampling time interval between data point i and data point j, T is the maximum time span in the entire window, γ h is the control factor used to adjust the strength of temporal correlation in the h-th attention head, λ h is the regularization factor in the h-th attention head used to prevent singular values ​​from appearing when the query is close to the key feature, thereby capturing and quantifying the correlation between distant time points within the same modal signal and obtaining the intra-modal attention feature with multi-scale temporal correlation information;

[0082] Using the modal interaction attention calculation unit set in the multimodal Transformer attention network, the EEG signal segment sequence is first used as the dominant modal reference sequence, and cross-modal attention calculation is performed one by one with the ECG signal, blood pressure signal, and blood oxygen saturation signal segment sequences:

[0083]

[0084] Among them, C X,Y (t) is the cross-modal attention correlation strength between the dominant modal signal X and the target modal signal Y at time t, X(i) and Y(i) are the values ​​of modal signals X and Y at time point i, respectively, Δ represents the time window width selected around the current time t for calculating cross-modal correlation, D X,Y (t) represents the time difference between the peak points of modal signal X and modal signal Y at the current time t, and η is the attenuation coefficient, which is used to control the peak time difference D X,Y (t) The degree of influence on the cross-modal attention intensity, thereby generating a cross-attention weight sequence that reflects the real-time dynamic interaction between EEG signals and other physiological signals. Then, using the ECG signal, blood pressure signal, and blood oxygen saturation signal fragment sequences as the dominant modality reference sequences, cross-modal attention calculations are performed one by one with other modal signals in the same way to obtain the cross-attention weight sequence between each modality;

[0085] In real time, the internal attention features of each modality and the cross-attention weight sequence between each modality calculated at the current moment are used, and an adaptive feature fusion method based on attention weight normalization is used to fuse the internal features of each modality and the cross-modal interaction features to obtain a fusion feature sequence that simultaneously reflects the internal long-range dependency relationship of the physiological signal and the cross-modal dynamic interaction relationship. The adaptive feature fusion method based on attention weight normalization is as follows: the weight value of the internal attention feature of each modality at the current moment and the cross-attention weight value between each modality are linearly weighted to obtain a mixed attention weight value, and the mixed attention weight value of each modal signal is normalized by dividing the mixed attention weight value of all modal signals by the sum of the mixed attention weight values ​​of all modal signals to obtain an adaptive fusion weight coefficient for each modal signal, and then the characteristic value of each modal signal is linearly weighted according to the fusion weight coefficient;

[0086] An adaptive feature selection mechanism based on feature correlation entropy value changes is adopted in real time to dynamically calculate the correlation entropy value of each feature in the fusion feature sequence in real time to obtain a multimodal fusion feature that is dynamically updated and stable and reliable in real time. The adaptive feature selection mechanism based on feature correlation entropy value changes is as follows: calculating the numerical distribution probability of each fusion feature in a preset sliding time window at the current moment, calculating the correlation entropy value of each fusion feature in real time according to the numerical distribution probability, determining the entropy value change amplitude according to the absolute value of the difference between the correlation entropy value at the current moment and the correlation entropy value at the previous moment, and automatically selecting the fusion feature whose entropy value change amplitude exceeds the preset entropy value change threshold as the effective feature according to the preset entropy value change threshold, so as to filter out redundant features with low contribution to the determination of anesthesia depth and consciousness state in real time.

[0087] In this embodiment, the multimodal Transformer attention network includes a modality-independent attention feature encoding unit, a self-attention unit, a modality interaction attention calculation unit, a cross-modal attention timing correction unit, and an attention weight adaptive update unit:

[0088] The modality-independent attention feature encoding unit receives the EEG signal, the ECG signal, the blood pressure signal, and the blood oxygen saturation signal, performs feature mapping on the data points in each modality signal sequence, and generates an independent initial feature vector sequence for each modality signal;

[0089] The self-attention unit uses multiple independent and parallel multi-head attention calculation mechanisms to perform self-attention calculations on the data points within the modality for the initial feature vector sequence of each modal signal output by the modality-independent attention feature encoding unit, thereby obtaining the time correlation features within each modal signal;

[0090] The modal interaction attention calculation unit receives the time correlation features within each modal signal output by the self-attention unit, and uses the time correlation features of each modal signal as the dominant modal reference sequence in a rotation manner to perform cross-modal attention calculations with the time correlation features of other modal signals one by one to obtain a real-time changing cross-modal attention weight;

[0091] The cross-modal attention timing correction unit monitors the change rate of the cross-modal attention weight in two consecutive time windows in real time, identifies the cross-modal attention weight whose change rate exceeds a preset threshold as an abnormal change, corrects the abnormal change in real time, and outputs the corrected cross-modal attention weight;

[0092] The attention weight adaptive update unit calculates the attention entropy value of the internal time correlation feature of each modal signal and the attention entropy value of the cross-modal cross-attention weight in real time, and adjusts the contribution weight of each modal signal to the overall fusion feature in real time according to the attention entropy value. Specifically, the contribution weight of the modal signal with a lower attention entropy value is increased, and the contribution weight of the modal signal with a higher attention entropy value is reduced, thereby outputting in real time a multimodal fusion feature that has both long-range dependency capture capabilities and cross-modal dynamic adaptive characteristics.

[0093] In this embodiment, the long short-term memory neural network is used to predict the fusion features of the next moment based on the multimodal fusion features at the current moment to obtain a fusion feature prediction value, and the fusion feature prediction value is compared with the actual multimodal fusion features at the next moment to obtain a prediction error, which is specifically:

[0094] The multimodal fusion features at the current moment are input into a long short-term memory neural network with a three-layer stacked structure, wherein the three-layer stacked long short-term memory neural network specifically includes a first layer of feature input long short-term memory units, a second layer of temporal enhancement long short-term memory units, and a third layer of prediction output long short-term memory units connected in series.

[0095] The first layer of features is input into the long short-term memory unit, which is used to perform preliminary feature encoding on the input multimodal fusion feature sequence and generate a preliminary temporal feature representation sequence;

[0096] The second layer of time-series enhanced long short-term memory units takes the preliminary time-series feature representation sequence output by the first layer as input, and further captures the long-range temporal correlation information in the multimodal fusion feature sequence by setting higher-dimensional hidden layer feature vectors and state memory units, thereby enhancing the ability to capture time-series dependency information and generating an enhanced time-series feature representation sequence;

[0097] The third layer predicts and outputs a long short-term memory unit, which takes the enhanced temporal feature representation sequence output by the second layer as input, and adaptively controls the retention ratio of historical information, the update ratio of new feature information, and the output ratio of predicted features in each fused feature sequence by setting a forget gate threshold, an input gate threshold, and an output gate threshold, and predicts the fused feature prediction value at the next moment;

[0098] Performing corresponding difference calculations on the fusion feature prediction values ​​output by the long short-term memory neural network and the actual multimodal fusion features one by one to generate a feature prediction residual sequence corresponding to the actual multimodal fusion features, wherein the difference calculation uses Euclidean distance to measure the difference of each feature point to ensure that the residual value reflects the distance difference between the predicted value and the actual value;

[0099] A residual statistical analysis method based on a dynamic adaptive sliding window is adopted. With a fixed sliding window length as the benchmark, the sliding window length is dynamically adjusted according to the residual change rate calculated in real time. When the residual change rate exceeds the preset change rate threshold, the window length is shortened to improve the sensitivity to drastic residual fluctuations. When the residual change rate is lower than the preset change rate threshold, the window length is extended to improve the stability to low-amplitude residual fluctuations. The dynamic mean and dynamic standard deviation of each fusion feature residual in the sliding window are obtained in real time, specifically:

[0100]

[0101] Among them, RSI(t) represents the abnormal statistical index of the fusion feature prediction residual at time t, which is used to identify mutations or abnormal interference in the fusion feature. The larger the value of RSI(t), the greater the deviation of the current residual from the mean value in the window and the more significant the abnormality. r(t) represents the prediction residual between the fusion feature prediction value and the actual fusion feature value at time t, μ w (t) means the time t is the end point and the length is L w The mean of all residual values ​​in the dynamic sliding window of (t), σ w (t) means the time t is the end point and the length is L w The standard deviation of all residual values ​​in the dynamic sliding window of (t), L w (t) represents the sliding window length, which is dynamically adjusted with the rate of change of the residual. The larger the rate of change of the residual, the shorter the window length, and vice versa, so as to adapt to the changing trend of the fusion feature in real time;

[0102] Compare the absolute difference between the current fusion feature residual value and the residual mean in the corresponding dynamic sliding window in real time, and use three times the standard deviation of the residual in the dynamic sliding window as the abnormal threshold reference. When the absolute difference is greater than the abnormal threshold reference, identify the fusion feature residual value as a mutation or abnormal interference in the fusion feature sequence;

[0103] For mutations or abnormal interferences in the identified fusion features, the feature serial number corresponding to the abnormal feature, the specific time of occurrence, the residual value, and the specific multiple of exceeding the abnormal threshold benchmark are recorded in real time, so as to obtain the prediction error for identifying mutations or abnormal interferences in the fusion features in real time.

[0104] In this embodiment, the noise covariance matrix and observation covariance matrix of the Kalman filter are updated in real time based on the prediction error, and the fusion weight of the Kalman filter is dynamically adjusted to improve the sensitivity and suppression ability of abnormal fusion features. Specifically:

[0105] The absolute value of the prediction error corresponding to each fusion feature in the fusion feature prediction error is used as the basic data for real-time updating of the noise covariance matrix and the observation covariance matrix of the Kalman filter;

[0106] According to the real-time prediction error absolute value corresponding to each fusion feature in the fusion feature prediction error sequence, the prediction error absolute value mean and prediction error absolute value variance of each fusion feature prediction error absolute value in a sliding time window of fixed length are calculated respectively, and the prediction error absolute value mean is mapped to the diagonal elements of the observation covariance matrix of the Kalman filter in real time, and the prediction error absolute value variance is mapped to the diagonal elements of the noise covariance matrix of the Kalman filter in real time, so as to accurately reflect the measurement error amplitude and prediction error fluctuation amplitude of the real-time fusion feature;

[0107] According to the noise covariance matrix and the observation covariance matrix updated at the current moment, the state prediction equation and the state update equation of the Kalman filter are used to calculate the Kalman gain matrix between the fusion feature estimate and the actual fusion feature measurement value, wherein the state prediction equation is specifically: the fusion feature state prediction value at the current moment is obtained by the matrix product between the fusion feature estimate at the previous moment and the state transfer matrix; the state update equation is specifically: the fusion feature estimate at the current moment is obtained by real-time correction through the matrix product of the Kalman gain matrix calculated at the current moment and the difference between the actual fusion feature measurement value at the current moment and the fusion feature state prediction value at the current moment;

[0108] Calculating the numerical change rate of the diagonal elements of the Kalman gain matrix between two consecutive moments, and judging whether the numerical change rate exceeds a preset Kalman gain change rate threshold, wherein the preset Kalman gain change rate threshold is specifically: by performing statistical analysis on the historical change rates of the diagonal elements of the Kalman gain matrix, calculating the mean and standard deviation of the change rates of the diagonal elements of the Kalman gain matrix within a preset time length, and taking the mean plus three times the standard deviation as the Kalman gain change rate threshold to ensure accurate identification of abnormal fluctuation changes in the fusion feature; when the numerical change rate exceeds the Kalman gain change rate threshold, identifying the corresponding fusion feature as an abnormally sensitive fusion feature, and increasing the fusion weight of the abnormally sensitive fusion feature in real time to enhance the real-time response capability of the Kalman filter to the abnormally sensitive fusion feature;

[0109] For the abnormal sensitive fusion features obtained by real-time identification, a dynamic weighted adjustment strategy for the fusion feature weights is adopted in real time. Specifically, according to the values ​​of the diagonal elements of the Kalman gain matrix calculated at the current moment and the size of the value change rate of the diagonal elements, the fusion weights of the abnormal sensitive fusion features with a larger value change rate are automatically increased to a preset weight upper limit, and at the same time, the fusion weights of the non-abnormal sensitive fusion features with a smaller value change rate are automatically adjusted to a preset weight lower limit, thereby dynamically improving the sensitivity and suppression ability of the Kalman filter to the abnormal fusion features in real time, so as to ensure that the fusion feature sequence obtained in real time has a stable and reliable response capability to abnormal interference fluctuations.

[0110] In this embodiment, the dynamically adjusted Kalman filter is used to perform real-time calibration on the multimodal fusion features output by the multimodal Transformer attention network to generate robustly calibrated multimodal fusion features to overcome fluctuations in fusion features caused by clinical interference factors. Specifically:

[0111] receiving, in real time, multimodal fusion features output by the multimodal Transformer attention network, and inputting the multimodal fusion features into a dynamically adjusted Kalman filter for state calibration, wherein the dynamically adjusted Kalman filter has completed real-time updates of a noise covariance matrix, an observation covariance matrix, and fusion weights based on a real-time prediction error;

[0112] Based on the real-time updated noise covariance matrix and observation covariance matrix, the state prediction equation of the Kalman filter is first used to generate the current fusion feature prediction state value by using the matrix product between the robust calibration multimodal fusion features output at the previous moment and the state transfer matrix;

[0113] Performing matrix-based difference calculations on the current fusion feature prediction state value and the multimodal fusion feature value one by one to obtain the current fusion feature state prediction error value, and then performing weighted correction on the fusion feature state prediction error value using the real-time updated Kalman gain matrix to obtain a calibrated multimodal fusion feature estimation value;

[0114] Calculate and record the fluctuation amplitude and fluctuation rate of the multimodal fusion feature estimate at multiple consecutive moments, specifically by calculating the Euclidean distance change rate of the fusion feature estimate at multiple consecutive moments, and identify abnormal fluctuations of the fusion feature caused by clinical interference factors in real time;

[0115] An abnormal fluctuation suppression mechanism is used to robustly correct abnormal fluctuations in fusion features identified in real time. The abnormal fluctuation suppression mechanism is specifically as follows: the Euclidean distance change rate between fusion features at adjacent moments in a multimodal fusion feature sequence that is robustly calibrated at multiple consecutive moments is calculated in real time, and a fusion feature abnormal fluctuation intensity index is generated in real time based on the Euclidean distance change rate. When the fusion feature abnormal fluctuation intensity index exceeds the dynamic abnormal fluctuation threshold calculated based on the statistical distribution of the historical abnormal fluctuation intensity of the fusion feature sequence, the abnormal fluctuation adaptive suppression strategy is automatically activated. By calculating the changing trend of the historical fusion feature estimation value sequence and the proportional relationship between the current abnormal fluctuation intensity and the dynamic abnormal fluctuation threshold in real time, the correction intensity of the fusion feature estimation value is determined based on the proportional relationship, and the weight ratio of the historical fusion feature estimation value sequence in the current fusion feature estimation value correction is dynamically adjusted, thereby outputting the robustly calibrated multimodal fusion feature in real time.

[0116] In this embodiment, the predicted values ​​of the subject's anesthesia depth and consciousness state are calculated in real time based on the robustly calibrated multimodal fusion features. When the predicted values ​​exceed the pre-set anesthesia depth safety threshold or consciousness recovery risk threshold, an anesthesia warning signal is output, specifically:

[0117] Acquire calibrated multimodal fusion features in real time and input the fusion features into a real-time anesthesia depth and consciousness state calculation unit, which uses a fully connected multi-layer feedforward neural network. The multi-layer feedforward neural network includes an input layer, multiple hidden layers, and an output layer. The number of neurons in the input layer is consistent with the fusion feature dimension, and each hidden layer uses a nonlinear activation function to perform layer-by-layer nonlinear mapping on the fusion feature vector.

[0118] The output layer of the multi-layer feedforward neural network is provided with two independent output neurons, wherein the first output neuron outputs the predicted value of the anesthesia depth at the current moment in real time, and the second output neuron outputs the predicted value of the consciousness state at the current moment in real time, so as to reflect the anesthesia depth and consciousness state of the subject in real time;

[0119] A real-time threshold determination method is used to perform real-time analysis on the predicted anesthesia depth value, wherein the anesthesia depth safety threshold is obtained through statistical analysis of previous clinical trial data. Specifically, a statistical analysis is performed on a large number of patient sample data with a risk of consciousness recovery during clinical anesthesia. Based on the anesthetic drug concentration and clinical physiological index change curve, a probability statistical method is used to obtain an anesthesia depth value that can clearly distinguish between a safe anesthesia state and an insufficient or excessively deep anesthesia state as the anesthesia depth safety threshold;

[0120] A real-time threshold determination method is used to perform real-time analysis on the predicted value of the state of consciousness, wherein the risk threshold for consciousness recovery is obtained by statistically analyzing the physiological indicators and clinical feedback data of a large number of patients who have recovered consciousness. Specifically, by analyzing the changing patterns of multi-source physiological signals such as EEG, ECG, blood pressure, and blood oxygen saturation during the transition from anesthesia to consciousness recovery in a large number of patients, a numerical feature commonly exhibited by clinical physiological signals before consciousness recovery occurs is determined, and this numerical feature is determined as the risk threshold for consciousness recovery;

[0121] performing a numerical comparison of the predicted anesthesia depth value with the anesthesia depth safety threshold in real time, and simultaneously performing a numerical comparison of the predicted consciousness state value with the consciousness recovery risk threshold in real time, and automatically triggering an anesthesia early warning signal generating unit when at least one of the following situations occurs: the predicted anesthesia depth value exceeds the anesthesia depth safety threshold, or the predicted consciousness state value exceeds the consciousness recovery risk threshold;

[0122] After being automatically triggered, the anesthesia warning signal generation unit generates a detailed anesthesia warning signal in real time, including the subject number, the current abnormal prediction value, the specific time when the prediction value exceeds the threshold, and the corresponding threshold value, and outputs it in real time through the system display unit and the alarm unit to complete the anesthesia warning.

[0123] Example 1:

[0124] To verify the feasibility of the present invention in practice, the present invention was applied to a clinical anesthesia monitoring platform of a medical institution to conduct real-time monitoring and early warning of the anesthesia depth and consciousness state of a group of patients undergoing clinical anesthesia. In this application scenario, traditional anesthesia monitoring technology usually uses a single EEG signal to judge the anesthesia state. This is easily affected by poor electrode contact, myoelectric interference, and physiological interference factors, resulting in delayed early warnings and a high rate of misjudgment. It is unable to accurately reflect the fluctuations in anesthesia depth and the risk of consciousness recovery, which seriously restricts the safety management of clinical anesthesia.

[0125] During the actual operation, four types of multi-source physiological signal data generated by patients during anesthesia were first collected in real time: EEG signals, ECG signals, blood pressure signals, and blood oxygen saturation signals. Each type of signal was preprocessed using a bandpass filter, a baseline drift removal algorithm, and a physiological interference elimination method to obtain high-quality time series signal data suitable for fusion. Subsequently, the preprocessed physiological signal data of each modality was input into a multimodal Transformer attention network, which extracted the temporal correlation features within the physiological signals and the cross-correlation features between different physiological signals, thereby obtaining a dynamically updated multimodal fusion feature representation.

[0126] During the fusion feature prediction phase, the multimodal fusion features are input into a long-short-term memory (LSTM) neural network to predict the fusion features for the next moment. The prediction error is calculated in real time and used to identify sudden changes or abnormal interference in the fusion features. The system utilizes a three-layer stacked LSTM neural network. The first layer initially encodes the fusion features, the second layer enhances the temporal dependencies of the features, and the third layer adaptively outputs the predicted feature values. Based on the fusion feature prediction error, the Kalman filter's noise covariance matrix and observation covariance matrix are then updated in real time, dynamically adjusting the Kalman filter's fusion weights to improve its sensitivity and ability to suppress abnormal fusion features.

[0127] During the real-time robust calibration phase of fusion features, the system utilizes a dynamically adjusted Kalman filter to calibrate the multimodal fusion features output by the multimodal Transformer network in real time, effectively overcoming fluctuations in fusion features caused by clinical interference. By dynamically coordinating the fusion feature prediction errors with the Kalman filter in real time, the system achieves robust and accurate predictions of anesthesia depth and state of consciousness.

[0128] In actual tests, the system uses a fully connected multi-layer feedforward neural network to calculate the predicted values ​​of anesthesia depth and consciousness state in real time based on robustly calibrated multimodal fusion features, and sets anesthesia depth safety threshold and consciousness recovery risk threshold to perform real-time threshold judgment. When the predicted value exceeds the preset threshold, the anesthesia warning signal is triggered in real time.

[0129] In this embodiment, the data of 5 actual clinical patients were used for evaluation, and the system-predicted anesthesia depth, consciousness state, warning delay time, abnormal interference recognition rate, and prediction error value were compared with the measured values ​​in detail. The specific data are shown in the following table:

[0130] Table 1 Comparison of the actual application results of clinical anesthesia state monitoring

[0131]

[0132]

[0133] It can be clearly seen from the data analysis of Table 1 above that the method proposed in the present invention has a significant improvement in the accuracy of anesthesia depth prediction and consciousness state monitoring. For example, the measured value of the anesthesia depth of sample A001 is 45.2, and the predicted value is 44.9, with an error of only 0.3; the error between the predicted value of consciousness state 0.23 and the measured value 0.22 is only 0.01, which shows the excellent predictive ability of the system. In addition, the system's recognition rate of abnormal interference also reflects a high accuracy, with an average recognition rate of more than 94%. Especially in terms of warning delay time, the system performs well, and the measured delay time is generally controlled within 1.5 seconds, which is significantly better than the delay performance index of traditional methods that usually exceeds 3 seconds.

[0134] In addition, the present invention reflects the accuracy of fusion feature prediction through real-time calculated prediction error values. The actual prediction error value of sample A003 is 0.33, and the system prediction value is 0.32, with an extremely small deviation. This shows that the system can keenly identify and accurately predict abnormal fluctuations in fusion features, further enhancing the reliability of monitoring and early warning.

[0135] In general, the multi-source signal fusion anesthesia warning method proposed in this embodiment shows obvious advantages in anesthesia status monitoring. It has the remarkable characteristics of high accuracy in predicting anesthesia depth and consciousness state, fast real-time response, and strong ability to identify abnormal interference. It can significantly improve the level of clinical anesthesia safety control, and has obvious clinical application value and broad application prospects.

[0136] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. An anesthesia early warning method based on multi-source signal fusion, characterized in that: include: Real-time collection of physiological signals during anesthesia of the subjects, and processing to obtain time-series physiological signal data; The multimodal Transformer attention network is used to extract the time-correlation features and cross-correlation features within the temporal physiological signal data to obtain multimodal fusion features. A long short-term memory neural network is used to predict the fusion features of the next moment based on the multimodal fusion features of the current moment, and the predicted fusion feature value is obtained. The predicted value is then compared with the actual multimodal fusion features of the next moment to obtain the prediction error. Based on the prediction error, update the noise covariance matrix and observation covariance matrix of the Kalman filter; The multimodal fusion features output by the multimodal Transformer attention network are calibrated using the adjusted Kalman filter to generate calibrated multimodal fusion features; The predicted values ​​of the subject's anesthesia depth and consciousness state are calculated based on the calibrated multimodal fusion features. When the predicted values ​​exceed the pre-set anesthesia depth safety threshold or consciousness recovery risk threshold, an anesthesia warning signal is output.

2. The anesthesia early warning method based on multi-source signal fusion according to claim 1, characterized in that: The real-time collection of physiological signals during anesthesia of the subject and the processing to obtain time-series physiological signal data are specifically: The EEG signals were collected from multiple preset electrode positions on the subject's scalp, and the myoelectric artifacts were removed by bandpass filtering. The adaptive mean filtering method was used to remove the slowly varying baseline drift to obtain time-series EEG signal data. The ECG signals of the subjects were collected at the standard lead position on their chests, high-frequency environmental noise was eliminated through a bandpass filter, and nonlinear baseline drift was removed using the wavelet transform method to obtain time-series ECG signal data; Invasive blood pressure signals are collected from the subjects' radial or brachial arteries. Spike interference is eliminated using a median filter, and baseline drift is removed using a low-pass filter to obtain time-series blood pressure signal data. The blood oxygen saturation signal at the fingertips of the subjects is collected in real time. High-frequency environmental noise and motion artifacts are removed through a bandpass filter, and periodic baseline drift is dynamically removed using a sliding average filter to obtain time-series blood oxygen saturation signal data. The time-series EEG signal data, time-series ECG signal data, time-series blood pressure signal data, and time-series blood oxygen saturation signal data are synchronized to a unified time series, and the physiological signal data of each channel are adjusted through an interpolation algorithm based on a cubic spline function to obtain the time-series physiological signal data.

3. The anesthesia early warning method based on multi-source signal fusion according to claim 1, characterized in that: The multimodal fusion features are obtained through the multimodal Transformer attention network, specifically: The time series physiological signal data are segmented using sliding windows with fixed length and time overlap ratio to obtain physiological signal segment sequences of each signal data; Perform multi-scale self-attention calculations simultaneously on data points in any sequence of physiological signal segments of each signal data to obtain the internal attention features of each modality; Using the EEG signal segment sequence as the dominant modality reference sequence, a cross-attention weight sequence between the EEG signal and other physiological signals is generated. Using the ECG signal, blood pressure signal, and blood oxygen saturation signal segment sequences as the dominant modality reference sequences, a cross-attention weight sequence between each modality is obtained. According to the internal attention features of each modality and the cross-attention weight sequence between each modality, the internal features of each modality and the cross-modal interaction features are fused to obtain a fused feature sequence; The correlation entropy value of each feature in the fusion feature sequence is calculated, and the fusion feature with higher entropy value is selected as the effective feature to obtain the multimodal fusion feature.

4. The anesthesia early warning method based on multi-source signal fusion according to claim 3, characterized in that: The multimodal Transformer attention network includes a modality-independent attention feature encoding unit, a self-attention unit, a modality-interaction attention calculation unit, a cross-modality attention timing correction unit, and an attention weight adaptive update unit: The modality-independent attention feature encoding unit performs feature mapping on the data points in each modality signal sequence to generate an initial feature vector sequence; The self-attention unit uses multiple independent and parallel multi-head attention calculation mechanisms to perform self-attention calculations on the data points within the modality for the initial feature vector sequence, thereby obtaining the time correlation features within each modal signal; The modal interaction attention calculation unit uses the time correlation features of each modal signal as the dominant modal reference sequence in a rotation manner, and performs cross-modal attention calculation with the time correlation features of other modal signals one by one to obtain cross-modal cross-attention weights; The cross-modal attention timing correction unit performs real-time correction on the cross-modal attention weight whose change rate exceeds a preset threshold through a timing correction method based on sliding window mean calculation, and outputs the cross-modal attention weight; The attention weight adaptive update unit calculates the attention entropy value of the internal time correlation feature of each modal signal and the attention entropy value of the cross-modal cross-attention weight, and adjusts the contribution weight of each modal signal to the overall fusion feature in real time according to the attention entropy value.

5. The anesthesia early warning method based on multi-source signal fusion according to claim 1, characterized in that: The prediction error obtained by the long short-term memory neural network is specifically: The multimodal fusion features at the current moment are input into a long short-term memory neural network with a three-layer stacked structure to predict the fusion feature prediction value at the next moment; Calculate the difference between the fusion feature prediction value and the actual multimodal fusion feature one by one to generate a feature prediction residual sequence that corresponds one to one with the actual multimodal fusion feature; A residual statistical analysis method based on a dynamic adaptive sliding window is used. The sliding window length is fixed as a benchmark and the sliding window length is dynamically adjusted according to the residual change rate calculated in real time. The absolute difference between the current fusion feature residual value and the residual mean value in the corresponding dynamic sliding window is compared in real time. When the absolute difference is greater than the abnormal threshold, the fusion feature residual value is identified as a mutation or abnormal interference in the fusion feature sequence. For mutations or abnormal interferences in the identified fusion features, the information corresponding to the abnormal features is recorded in real time to obtain the prediction error.

6. The anesthesia early warning method based on multi-source signal fusion according to claim 1, characterized in that: The updated Kalman filter is specifically: The absolute value of the prediction error corresponding to each fusion feature in the prediction error is used as the basic data of the noise covariance matrix and the observation covariance matrix of the Kalman filter; According to the real-time prediction error absolute value corresponding to each fusion feature in the prediction error sequence, the prediction error absolute value mean and prediction error absolute value variance of each fusion feature prediction error absolute value in a sliding time window of fixed length are calculated and mapped respectively; According to the noise covariance matrix and observation covariance matrix updated at the current moment, the Kalman filter's state prediction equation and state update equation are used to calculate the Kalman gain matrix between the fusion feature estimate and the actual fusion feature measurement value; Calculate the value change rate of the diagonal elements of the Kalman gain matrix between two consecutive moments, and determine whether the value change rate exceeds the preset Kalman gain change rate threshold, and identify the obtained abnormally sensitive fusion features; The Kalman filter is dynamically adjusted based on the abnormally sensitive fusion features obtained through identification.

7. The anesthesia early warning method based on multi-source signal fusion according to claim 1, characterized in that: The multimodal fusion features output by the multimodal Transformer attention network are calibrated using the adjusted Kalman filter to generate calibrated multimodal fusion features, specifically: Input the multimodal fusion features into the dynamically adjusted Kalman filter for state calibration; Based on the real-time updated noise covariance matrix and observation covariance matrix, the state prediction equation of the Kalman filter is used to generate the current moment's fusion feature prediction state value by using the matrix product between the calibrated multimodal fusion features output at the previous moment and the state transfer matrix; The difference between the current fusion feature prediction state value and the multimodal fusion feature value is calculated in matrix form one by one to obtain the current fusion feature state prediction error value. The fusion feature state prediction error value is then weighted corrected using the real-time updated Kalman gain matrix to obtain the calibrated multimodal fusion feature estimation value. Calculate and record the fluctuation amplitude and fluctuation rate of the multimodal fusion feature estimation value at multiple consecutive moments to identify abnormal fluctuations in the fusion feature caused by clinical interference factors; The abnormal fluctuation suppression mechanism is used to robustly correct the abnormal fluctuation of the real-time identified fusion features to obtain the calibrated multimodal fusion features.

8. The anesthesia early warning method based on multi-source signal fusion according to claim 1, characterized in that: The calibrated multimodal fusion feature is used to calculate the predicted value of the subject's anesthesia depth and consciousness state, and obtain an anesthesia warning signal, specifically: Inputting the calibrated multimodal fusion features into a real-time anesthesia depth and consciousness state calculation unit, wherein the real-time anesthesia depth and consciousness state calculation unit adopts a multi-layer feedforward neural network with a fully connected structure; The output layer of the multi-layer feedforward neural network is equipped with two independent output neurons: The first output neuron outputs the predicted value of the anesthesia depth at the current moment in real time. The second output neuron outputs the predicted value of the current state of consciousness in real time; A real-time threshold determination method is used to analyze the predicted values ​​of anesthesia depth and consciousness state in real time, and the consciousness recovery risk threshold and anesthesia depth safety threshold are obtained respectively; The anesthesia depth prediction value is numerically compared with the anesthesia depth safety threshold in real time, and the consciousness state prediction value is numerically compared with the consciousness recovery risk threshold in real time. When at least one of the anesthesia depth prediction value exceeds the anesthesia depth safety threshold or the consciousness state prediction value exceeds the consciousness recovery risk threshold, a detailed anesthesia warning signal is generated including the subject number, the current abnormal prediction value, the specific time when the prediction value exceeds the threshold and the corresponding threshold value, and anesthesia warning is performed.

Images