Deep learning-based anesthesia depth real-time assessment method and system thereof

Abstract

The present application relates to the medical technical field, and particularly to a method and system for real-time evaluation of anesthesia depth based on deep learning, which comprises the following steps: acquiring multi-modal physiological signals of a patient, the multi-modal physiological signals including electroencephalogram, electrocardiogram, photoplethysmography, blood pressure and respiratory signals; preprocessing the multi-modal physiological signals; fusing the preprocessed multi-modal physiological signals by using a hierarchical attention fusion network to obtain fusion features; wherein the hierarchical attention fusion network comprises a signal-level attention layer, a time-scale attention layer and a frequency component attention layer; adjusting the time scale of the fusion features based on a dynamic time window to obtain adjusted features; through multi-modal signal fusion and deep learning technology, the accuracy of anesthesia depth evaluation is improved to 95%, which is 10 percentage points higher than the prior art, and the risk of insufficient or excessive anesthesia can be effectively reduced to improve the safety of surgery.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of medical technology, in particular to a method for real-time evaluation of anesthesia depth based on deep learning and a system thereof. BACKGROUND

[0002] Real-time evaluation of anesthesia depth is crucial for surgical safety and patient outcomes. With the advancement of medical technology, anesthesia depth monitoring methods have evolved from single indicators to multi-modal analysis.

[0003] Early anesthesia depth evaluation mainly relied on clinical signs such as blood pressure, heart rate, and pupillary response. However, these indicators often lag behind actual changes in anesthesia status, making it difficult to achieve precise control. In the 1990s, electroencephalogram (EEG)-based monitoring techniques, such as the BIS index, began to be widely used. These methods analyze the spectral characteristics of EEG to evaluate anesthesia depth, significantly improving monitoring accuracy.

[0004] In recent years, with the development of machine learning technology, researchers have begun to try to combine multiple physiological signals to obtain a more comprehensive evaluation of anesthesia status. For example, some studies have combined EEG with electrocardiogram (ECG) and used neural network models for analysis, achieving certain results. However, these methods still have the following technical problems:

[0005] 1. Insufficient signal fusion: Most existing methods still rely on a single signal (usually EEG), failing to fully utilize the complementary information of multi-modal signals.

[0006] 2. Lack of real-time performance: Complex algorithms often result in processing delays, making it difficult to meet the needs of rapid changes during surgery.

[0007] 3. Poor adaptability to individual differences: Existing models are mostly static models, making it difficult to adapt to individual differences in patients during anesthesia.

[0008] 4. Lack of interpretability: Many machine learning models are like black boxes, making it difficult for anesthesiologists to provide intuitive and understandable decision-making basis.

[0009] 5. Weak anti-interference ability: The operating room environment is complex, and existing methods lack resistance to factors such as electrosurgical interference and body movement.

[0010] In view of the above problems, there is an urgent need for an anesthesia depth evaluation method that can comprehensively utilize multi-modal physiological signals, have real-time processing capability, adapt to individual differences, provide interpretable results, and have strong anti-interference ability. SUMMARY

[0011] The purpose of the present application is to provide a deep learning-based anesthesia depth real-time evaluation method and system to solve the above problems in the prior art.

[0012] The present application provides a deep learning-based anesthesia depth real-time evaluation method, comprising the following steps:

[0013] Obtain the multi-modal physiological signals of the patient, including electroencephalogram (EEG), electrocardiogram (ECG), photoplethysmogram (PPG), blood pressure (BP), and respiratory signal (RESP); preprocess the multi-modal physiological signals; the method further comprises:

[0014] Fusing the preprocessed multi-modal physiological signals using a hierarchical attention fusion network (HAFN) to obtain fusion features;

[0015] The hierarchical attention fusion network includes a signal-level attention layer, a time-scale attention layer, and a frequency component attention layer.

[0016] Adjusting the fusion features based on a dynamic time window (DTW) to obtain adjusted features;

[0017] The dynamic time window uses a reinforcement learning method to adaptively adjust the size of the observation window based on the current signal features and historical states.

[0018] Input the adjusted features into a multi-task learning framework to simultaneously predict anesthesia depth classification, electroencephalogram frequency band energy ratio, autonomic nervous system balance index, and drug concentration.

[0019] According to the output of the multi-task learning framework, determine the anesthesia depth of the patient, and generate an anesthesia depth staging chart.

[0020] Specifically, the preprocessing step of the multi-modal physiological signals comprises:

[0021] Adaptive wavelet denoising is performed on the multi-modal physiological signals; wherein the best wavelet basis function and threshold are automatically selected by an unsupervised learning method; the denoised signals are compressed and sampled; wherein the sparsity of the signals is utilized to reduce the dimensionality of the sampling through a random projection matrix; online anomaly detection and correction is performed on the sampled signals; wherein the local outlier factor (LOF) algorithm is used to detect and correct signal anomalies in real time.

[0022] Specifically, the processing steps of the hierarchical attention fusion network (HAFN) include applying signal-level attention to the input multi-modal physiological signals to obtain signal-level fusion features; applying time-scale attention to the signal-level fusion features to obtain time-scale fusion features; and applying frequency component attention to the time-scale fusion features to obtain the final fusion features.

[0023] Specifically, the implementation step of the dynamic time window DTW comprises: defining a state space including current signal features and a history window size; defining an action space including the amplitude of adjusting the window size; defining a reward function based on signal stability and information quantity; using a reinforcement learning model with an Actor-Critic architecture to learn the optimal window adjustment strategy by interacting with the environment.

[0024] Specifically, the processing step of the multi-task learning framework comprises: setting the main task as anesthesia depth classification; setting the auxiliary tasks to include electroencephalogram frequency band energy ratio prediction, autonomic nervous system balance index prediction, and drug concentration estimation; constructing a shared feature extraction network and task-specific output layers; defining a multi-task loss function considering the loss and weight of each task; and optimizing the performance of all tasks simultaneously through a backpropagation algorithm.

[0025] Specifically, the feature extraction step further comprises:

[0026] A phase synchronization index PSI is calculated to quantify the synchronization between different brain regions or different physiological signals.

[0027] A multi-scale entropy MSE is calculated to evaluate the complexity of signals at different time scales.

[0028] Symbolic dynamics features SDF are extracted to convert continuous signals into symbolic sequences and extract the dynamics features of the system.

[0029] Specifically, the multi-task learning framework adopts the following optimized model architecture:

[0030] A dynamic depth separable convolution is used to dynamically adjust the size and grouping number of the convolution kernel according to the complexity of the input features; an attention mechanism is introduced in the residual connection to realize attention-guided skip connection; and an adaptive activation function is used to automatically adjust the parameters of the activation function according to the needs of different layers.

[0031] Specifically, the post-processing step further comprises: calculating the probability distribution of anesthesia depth; detecting abnormal states and generating alerts; performing explainability analysis to quantify the contribution of different input features to the prediction results; and generating anesthesia adjustment recommendations based on the prediction results and explainability analysis.

[0032] Specifically, the generation step of the anesthesia depth staging chart comprises: dividing anesthesia depth into six stages of wakefulness, mild sedation, moderate sedation, deep sedation, general anesthesia, and over-deep anesthesia; using different colors to represent different anesthesia stages, with the depth of the color representing the reliability of the stage; and plotting a dynamic bar chart with time as the horizontal axis and anesthesia depth stage as the vertical axis.

[0033] The method is implemented by an anesthesia depth real-time evaluation system, which comprises:

[0034] A multi-source signal acquisition module for acquiring multi-modal physiological signals of a patient;

[0035] A signal preprocessing module for denoising, compressed sensing sampling and anomaly detection of the acquired signals;

[0036] A feature extraction module for calculating time-frequency features, phase synchronization index, multi-scale entropy and symbolic dynamics features;

[0037] A deep learning inference module comprising a hierarchical attention fusion network, a dynamic time window adjustment unit and a multi-task learning framework;

[0038] A result analysis and display module for generating anesthesia depth staging chart and decision suggestions;

[0039] The deep learning inference module is implemented by an edge computing device to ensure real-time performance; the output of the result analysis and display module is used to guide the multi-source signal acquisition module to adjust the acquisition parameters, forming a closed loop control.

[0040] The present application has the following beneficial effects:

[0041] 1. High precision evaluation: through multi-modal signal fusion and deep learning technology, the anesthesia depth evaluation accuracy is improved to 95%, which is 10 percentage points higher than the prior art. This high precision evaluation can effectively reduce the risk of insufficient or excessive anesthesia and improve the safety of surgery.

[0042] 2. Real-time performance: edge computing and FPGA acceleration are adopted, and the system response time is less than 100ms, which can timely capture the rapid changes of anesthesia state. This feature is particularly important for emergency handling in surgery.

[0043] 3. Personalized adaptation: the dynamic time window technology can adaptively adjust according to the individual characteristics of the patient and the anesthesia stage, improving the adaptability to different patients and different surgery stages.

[0044] 4. Comprehensive evaluation: the multi-task learning framework simultaneously predicts anesthesia depth, EEG frequency band energy ratio, autonomic nervous system balance index and drug concentration, providing comprehensive patient state information for anesthesiologists.

[0045] 5. Explainability: the anesthesia depth staging chart visually displays the changes of anesthesia state, enhancing the explainability of decision-making, which helps doctors make more reasonable clinical decisions.

[0046] 6. Strong anti-interference ability: adaptive signal preprocessing and multi-modal fusion significantly improve the system's resistance to common operating room interference, ensuring the stability and reliability of the evaluation results.

[0047] 7. Assistant decision support: the system not only provides anesthesia depth evaluation, but also gives anesthesia drug adjustment suggestions based on the results of multi-task learning, providing strong support for precise anesthesia management.

[0048] In summary, the anesthesia depth real-time evaluation method provided by the present application effectively solves the problems faced by the prior art through innovative algorithm design and system integration, and makes important contributions to improving anesthesia safety, optimizing anesthesia management processes, and promoting the development of personalized precise anesthesia. BRIEF DESCRIPTION OF DRAWINGS

[0049] Figure 1 The system logic block diagram of the present application.

[0050] Figure 2 The hierarchical attention fusion network HAFN logic block diagram of the present application.

[0051] Figure 3 The dynamic time window DTW technology logic block diagram of the present application.

[0052] Figure 4 The multi-task learning framework logic block diagram of the present application.

[0053] Figure 5 The anesthesia depth staging diagram generation logic block diagram of the present application.

[0054] Figure 6 The feature extraction module logic block diagram of the present application.

[0055] Figure 7 The phase synchronization index PSI calculation logic block diagram of the present application.

[0056] Figure 8 The multi-scale entropy MSE calculation logic block diagram of the present application.

[0057] Figure 9 The symbolic dynamics feature SDF extraction logic block diagram of the present application.

[0058] Figure 10 The anti-interference ability line graph of the present application and Comparative Examples 1-3.

[0059] Figure 11 The anesthesia depth-time-electroencephalogram frequency 3D surface graph of the present application and Comparative Examples 1-3. DETAILED DESCRIPTION

[0060] In order to further illustrate the technical means and effects taken by the present application to achieve the predetermined inventive purpose, the following describes in detail the specific implementation, structure, features and effects of the anesthesia depth real-time evaluation method and system based on deep learning according to the present application, in combination with the accompanying drawings and preferred embodiments. In the following description, different "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. In addition, the specific features, structures or characteristics in one or more embodiments can be combined in any suitable form.

[0061] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs.

[0062] The present application will be further described in detail below in combination with the accompanying drawings and specific embodiments.

[0063] As shown in Figures 1-11 the anesthesia depth real-time evaluation method based on deep learning provided by the present application comprises the following steps:

[0064] Step S1: acquiring multi-modal physiological signals of a patient.

[0065] In the operating room or intensive care unit, the multi-modal physiological signals of the patient are synchronously collected by multiple medical sensors. Specifically, the multi-modal physiological signals include electroencephalogram (EEG), electrocardiogram (ECG), photoplethysmogram (PPG), blood pressure (BP) and respiratory signal (RESP).

[0066] Preferably, in one embodiment of the present application, the EEG signal adopts 10-20 system electrode placement, focusing on the frontal lobe region (Fp1, Fp2, F3, F4), and the sampling rate is 250 Hz. This is because the EEG signal of the frontal lobe region is most sensitive to changes in anesthesia depth. The ECG signal adopts standard 12-lead, with a sampling rate of 500 Hz, to capture subtle changes such as heart rate variability. The PPG signal adopts a finger clip sensor, with a sampling rate of 100 Hz, for assessing changes in peripheral vascular resistance. The BP signal measures non-invasive blood pressure every 3 minutes, for monitoring the effects of anesthesia on the cardiovascular system. The RESP signal adopts a chest strap sensor, with a sampling rate of 25 Hz, for monitoring changes in respiratory rate and depth.

[0067] Step S2: pre-processing the multi-modal physiological signals.

[0068] In order to improve the accuracy of subsequent analysis, the original signals collected in this step are pre-processed. This step includes the following sub-steps:

[0069] S2.1 adaptive wavelet denoising, first, an adaptive wavelet denoising technique is applied to each physiological signal. The specific implementation is as follows:

[0070] a) Calculate the wavelet coefficients of the signal:

[0071] W = WaveletTransform(x)

[0072] where x is the original signal, and WaveletTransform represents the wavelet transform function. The present application preferably uses the Discrete Wavelet Transform (DWT) because of its excellent performance in computational efficiency and signal reconstruction.

[0073] b) Adaptively select the threshold value according to the local characteristics of the signal

[0074]

[0075] where σ is the estimate of the noise standard deviation, and N is the length of the signal. The present application uses the MAD (Median Absolute Deviation) method to estimate the noise standard deviation because it is more robust to outliers.

[0076] c) Soft threshold processing of wavelet coefficients:

[0077] W denoised = sign(W) · max(0, |W| - λ)

[0078] Soft threshold processing can produce smoother results than hard threshold processing, reducing artificial artifacts.

[0079] d) Inverse wavelet transform to obtain the denoised signal:

[0080] x denoised = InverseWaveletTransform(W denoised )

[0081] Preferably, in one embodiment of the present application, the sym8 wavelet is selected as the base wavelet for EEG signals, with 5 layers of decomposition. This selection can effectively remove high-frequency muscle interference and 50Hz power frequency interference, while retaining the δ, θ, α, β band information related to anesthesia depth in the EEG signal. For ECG signals, db4 wavelet is selected, with 4 layers of decomposition, which can effectively remove baseline drift and high-frequency noise, and retain the morphological characteristics of QRS complexes.

[0082] S2.2 Compressive sensing sampling In order to reduce the pressure of data storage and transmission while retaining the key information of the signal, the present application uses the compressive sensing technology to sample the denoised signal. The specific steps are as follows:

[0083] a) Construct a random projection matrix Φ:

[0084]

[0085] where M is the dimension of the sampled signal, N is the dimension of the original signal, usually M « N. In the present application, M is chosen based on experimental results, usually 20%~30% of the length of the original signal.

[0086] b) multiply the signal x with Φ, to get the sampled result:

[0087] y = Φx

[0088] c) recover the original signal by L1 optimization:

[0089]

[0090] Here, the orthogonal matching pursuit (OMP) algorithm is used to solve the above optimization problem, because it achieves a good balance between computational speed and reconstruction accuracy.

[0091] Preferably, in one embodiment of the present application, for EEG signals, the sampling rate can be reduced from 250 Hz to 62.5 Hz; for ECG signals, it can be reduced from 500 Hz to 125 Hz. Such reduction of sampling rate can significantly reduce the amount of data, while ensuring that the key features of the signal are not lost. In particular for EEG signals, a sampling rate of 62.5 Hz is still able to retain frequency components up to 30 Hz, which is sufficient to cover the main frequency band related to depth of anesthesia.

[0092] S2.3 Online anomaly detection and correction, in order to identify and process outliers in the signal in real time, such as artifacts caused by body movement, the present application uses the local outlier factor (LOF) algorithm for online anomaly detection. The specific steps are as follows:

[0093] a) calculate the local outlier factor of each data point:

[0094]

[0095] where N k (p) represents the k-nearest neighbors of point p.

[0096] b) if the LOF value exceeds the preset threshold, mark the point as an anomaly:

[0097] if LOF k (p) > threshold:

[0098] mark p as anomaly

[0099] c) use interpolation or prediction methods to correct the outlier:

[0100] x corrected = Interpolate(x anomaly)

[0101] Preferably, in one embodiment of the present application, the value of k is selected as 5 and the threshold is set as 2.0. The selection of these parameters is based on statistical analysis of a large amount of clinical data, which can effectively identify abnormal values caused by sudden movement of the patient or electrode detachment, while not misjudging normal physiological fluctuations. For the data points marked as abnormal, a cubic spline interpolation method is used for correction to maintain the continuity of the signal.

[0102] Step S3: fuse the preprocessed signals using a hierarchical attention fusion network (HAFN).

[0103] In anesthesia depth assessment, different physiological signals reflect the anesthesia state with different emphasis. For example, EEG mainly reflects the activity of the cerebral cortex, while ECG and BP reflect the changes of the autonomic nervous system. In order to comprehensively and accurately assess the depth of anesthesia, the present application designs an innovative hierarchical attention fusion network (HAFN). The processing steps of HAFN include:

[0104] First, apply signal-level attention to the input multi-modal physiological signals X to obtain signal-level fusion features X s :

[0105] a) Calculate attention weights: apply signal-level attention

[0106]

[0107] b) Obtain signal-level fusion features:

[0108] X s =A s X

[0109] Where W s and V s are learnable parameter matrices.

[0110] Preferably, in one embodiment of the present application, the initial values of W s and V s are generated by the Xavier initialization method, and the initial learning rate is set to 0.001. This initialization method can make the network easier to converge at the beginning of training, especially when dealing with high-dimensional signals such as EEG. The signal-level attention mechanism can automatically learn the importance of different physiological signals at different anesthesia stages, for example, in the light anesthesia stage, the alpha wave activity of EEG may be more important, while in deep anesthesia, the delta wave of EEG and heart rate variability of ECG may be more indicative.

[0111] S3.2 Time-scale attention. During the anesthesia process, signal changes at different time scales have different implications for the indication of anesthesia depth. For example, short-term EEG burst suppression can indicate deep anesthesia, while long-term heart rate variability reduction can indicate a sustained anesthesia state. Therefore, we apply time-scale attention to the signal-level fusion feature X s to obtain the time-scale fusion feature X t

[0112] a) Calculate attention weights:

[0113]

[0114] b) Obtain the time-scale fusion feature:

[0115] X t = A t X s

[0116] Preferably, in an embodiment of the present application, the time-scale attention mechanism considers three time scales: short-term (5 seconds), medium-term (30 seconds), and long-term (2 minutes). This design can capture features at various time scales, from rapid EEG changes to slow blood pressure trends, thereby comprehensively assessing anesthesia depth.

[0117] S3.3 Frequency component attention. Different frequencies of brain electrical activity correspond to different anesthesia depths. For example, the enhancement of alpha waves (8-13 Hz) usually indicates light anesthesia or impending awakening, while the dominance of delta waves (0.5-4 Hz) indicates deep anesthesia. To this end, we apply frequency component attention to the time-scale fusion feature X t to obtain the final fusion feature X f :

[0118] a) Fourier transform X t

[0119] F(X t ) = FFT(X t

[0120] b) Calculate attention weights:

[0121]

[0122] c) Obtain the frequency component fusion feature:

[0123] X f = A f F(X t )

[0124] ​Preferably, in one embodiment of the present application, the EEG signal is focused on four frequency bands: delta (0.5-4 Hz), theta (4-8 Hz), alpha (8-13 Hz), and beta (13-30 Hz). The frequency component attention mechanism can automatically learn the importance of different frequency bands in anesthesia depth assessment and dynamically adjust the weights according to different anesthesia stages.

[0125] Step S4: Perform additional feature extraction.

[0126] In order to more comprehensively describe the anesthesia state of the patient, the present application further introduces the following feature extraction steps:

[0127] S4.1 Calculate the phase synchronization index (PSI), which is used to quantify the synchronization between different brain regions or different physiological signals, and the calculation formula is as follows:

[0128]

[0129] wherein, and are the phases of the two signals.

[0130] Preferably, in one embodiment of the present application, the PSI between the left and right hemispheres of the frontal lobe (Fp1 and Fp2) in the EEG signal is calculated. This indicator can reflect the functional connectivity strength between the brain hemispheres and show characteristic changes under different anesthesia depths. For example, under deep anesthesia, the PSI value usually decreases, indicating the weakening of brain functional connectivity.

[0131] S4.2 Calculate the multi-scale entropy (MSE). Multi-scale entropy is used to evaluate the complexity of the signal at different time scales, and the calculation formula is as follows:

[0132]

[0133] wherein, τ is the scale factor, m is the embedding dimension, and r is the similarity tolerance.

[0134] Preferably, in one embodiment of the present application, the MSE of the EEG signal is calculated at 20 time scales, the embedding dimension m = 2, and the similarity tolerance r = 0.15 × std(signal). The multi-scale entropy can capture the changes in the complexity of the electroencephalogram during anesthesia. Generally, as the depth of anesthesia increases, the MSE value decreases, reflecting the simplification of brain activity.

[0135] S4.3 Extract the symbolic dynamics feature (SDF), which converts continuous signals into symbolic sequences to extract the dynamics of the system. The specific steps are as follows:

[0136] a) Convert the signal to a symbolic sequence S = {s1, s2, …, s n},s i ∈A (symbol set);

[0137] b) Calculate the probability of the subsequence w is a subsequence of length m.

[0138] Preferably, in one embodiment of the present invention, symbolic dynamics analysis is performed on the RR interval sequence of the ECG signal. Three symbols are used (0 represents decrease, 1 represents no change, and 2 represents increase), and the sequence length m = 3. This analysis can reveal the fine structure of heart rate variability, reflecting the regulatory state of the autonomic nervous system. For example, under deep anesthesia, patterns such as "222" (continuous increase) and "0000" (continuous decrease) may decrease, while alternating patterns such as "101" and "010" may increase, reflecting a reduction in heart rate variability.

[0139] By introducing these additional features, the present invention can describe the patient's physiological state from multiple perspectives, improving the comprehensiveness and accuracy of anesthesia depth assessment. These features, combined with the previous fusion feature X... f Together, these elements are passed as input to subsequent dynamic time window adjustments and multi-task learning frameworks, further enhancing the system's performance.

[0140] Step S5: Adjust the time scale of the fused features based on the Dynamic Time Window (DTW) technique.

[0141] The rate of change in anesthesia depth varies at different stages of surgery. For example, the depth changes rapidly during induction, while the change is relatively slow during maintenance. To adapt to this dynamic characteristic, this invention employs Dynamic Time Window (DTW) technology to adjust the time scale of the fusion features. The specific steps are as follows: a) Define the state space s t This includes the current signal characteristics and the historical window size. b) Define the action space a t This includes adjusting the window size (increasing / decreasing / keeping it constant). c) Define the reward function r. t Based on signal stability and information content:

[0142] r t =α·StabilityMeasure(X) f )+β·InformationGain(X f )

[0143] Among them, StabilityMeasure measures the stability of the signal, and InformationGain measures the information gain brought by the addition of new data. α and β are trade-off coefficients, determined experimentally.

[0144] d) Reinforcement learning model using Actor-Critic architecture, learn the optimal window adjustment policy by interacting with the environment:

[0145] Update value function:

[0146] V(s t )=V(s t )+η[r t +γV(s t+1 )-V(s t )]

[0147] Update policy function:

[0148]

[0149] Preferably, in one embodiment of the present application, the initial time window is set to 30 seconds, allowing dynamic adjustment between 10 seconds and 2 minutes. The learning rate η is set to 0.01, and the discount factor γ is set to 0.99. This setting can strike a balance between maintaining evaluation stability and responsiveness to rapid changes.

[0150] Step S6: input the adjusted features into the multi-task learning framework.

[0151] In order to comprehensively evaluate the anesthesia state, the present application adopts a multi-task learning framework to simultaneously predict multiple related indicators. Specifically, it includes:

[0152] a) Construct a shared feature extraction network f shared and task-specific output layer f taski ;

[0153] b) Define the multi-task loss function:

[0154] L=λ1L depth +λ2L eeg +λ3L ans +λ4L drug +R(θ)

[0155] Where L depth is the anesthesia depth classification loss, L eeg is the EEG frequency band energy ratio prediction loss,

[0156] L ans is the autonomic nervous system balance index prediction loss, L drug is the drug concentration estimation loss, R(θ)

[0157] is the regularization term.

[0158] c) Optimize the performance of all tasks simultaneously through the backpropagation algorithm:

[0159]

[0160] Preferably, in one embodiment of the present invention, the shared network adopts an improved MobileNetV3 architecture to balance computational efficiency and feature extraction capability. The weight ratios λ1 to λ4 are set to 2:1:1:1, reflecting the dominance of the anesthesia depth classification task. The learning rate η adopts a cosine annealing strategy with an initial value of 0.001. This multi-task learning framework can utilize the correlation between tasks to improve the model's generalization ability and robustness.

[0161] Step S7: Based on the output of the multi-task learning framework, determine the patient's depth of anesthesia and generate an anesthesia depth staging map.

[0162] Based on the output of the multi-task learning framework, this step performs the final assessment and visualization of anesthesia depth. Specifically, this includes:

[0163] a) Calculate the probability distribution of the depth of anesthesia:

[0164] P(depth|X) = softmax(f depth (X))

[0165] Among them, f depth It is an anesthesia depth classifier, and X is the input feature.

[0166] b) Divide the depth of anesthesia into six stages: awake, mild sedation, moderate sedation, deep sedation, general anesthesia, and excessively deep anesthesia.

[0167] c) For each time point t, determine the anesthesia depth stage based on the output of the multi-task learning framework. t and the corresponding probability p t .

[0168] d) Generate an anesthesia depth stage map: with time t as the horizontal axis, the anesthesia depth stages are represented by... t The vertical axis uses different colors to represent different stages of anesthesia, with the intensity of the color indicating the reliability of that stage. The calculation method is as follows:

[0169] color_intensity=base_color×p t

[0170] Wherein, base_color is a predefined base color. Preferably, in one embodiment of the present invention, the color coding is as follows: conscious: blue (#0000FF); mild sedation: light green (#90EE90); moderate sedation: green (#008000); deep sedation: dark green (#006400); general anesthesia: yellow (#FFFF00); excessively deep anesthesia: red (#FF0000).

[0171] This color coding scheme intuitively reflects the changes in anesthesia depth, helping anesthesiologists quickly identify the patient's status. For example, when the color gradually changes from green to yellow, it prompts the doctor to possibly reduce the dose of anesthetic drugs to prevent over-anesthesia.

[0172] In addition, the present application also combines the output of other tasks to provide a more comprehensive assessment of the anesthesia state:

[0173] e) EEG frequency band energy ratio: display the energy ratio of δ, θ, α, β four frequency bands in the form of pie chart, reflecting the brain activity pattern.

[0174] f) Autonomic nervous system balance index: use a dashboard-style display to indicate the relative activity levels of the sympathetic and parasympathetic nervous systems.

[0175] g) Estimated drug concentration: display the estimated values of plasma and effect-site concentrations of main anesthetic drugs (such as propofol, remifentanil) in the form of a line chart.

[0176] By integrating these information, anesthesiologists can have a more comprehensive understanding of the patient's anesthesia state, and thus make more accurate drug decisions.

[0177] In order to realize the above method, the present application also provides an anesthesia depth real-time evaluation system. As shown in the figure, the system includes the following modules: Figure 2

[0178] 1. Multi-source signal acquisition module 1, which is responsible for acquiring the patient's multi-modal physiological signals, including EEG, ECG, PPG, BP and RESP.

[0179] Preferably, in an embodiment of the present application, the module includes: an 8-lead EEG acquisition device with a sampling rate of 250Hz; a 12-lead ECG monitor with a sampling rate of 500Hz; a finger clip PPG sensor with a sampling rate of 100Hz; a non-invasive BP monitor measuring every 3 minutes; a chest strap respiratory monitor with a sampling rate of 25Hz; all devices are connected to the central processing unit through a medical-grade data acquisition interface, ensuring the synchronization and integrity of the data.

[0180] Signal preprocessing module 2, which denoises, compressively samples and detects anomalies on the collected signals.

[0181] Preferably, in an embodiment of the present application, the module is implemented using FPGA (Field Programmable Gate Array) to ensure real-time performance. The FPGA model is Xilinx Artix-7 with a clock frequency of 200MHz, which can process multiple signals in parallel.

[0182] ​Feature extraction module 3, responsible for calculating time-frequency features, phase synchronization index, multi-scale entropy, and symbolic dynamics features.

[0183] Preferably, in one embodiment of the application, the module is accelerated using a GPU, using a NVIDIA Jetson Xavier NX module with 384 CUDA cores, which can efficiently handle large-scale parallel computing tasks.

[0184] Deep learning inference module 4, which includes a hierarchical attention fusion network, a dynamic time window adjustment unit, and a multi-task learning framework.

[0185] Preferably, in one embodiment of the application, the module also uses a NVIDIA Jetson Xavier NX module, using its deep learning accelerator (DLA) for efficient inference. The model is optimized using TensorRT, with a quantization precision of FP16 to balance accuracy and speed.

[0186] Result analysis and display module 5, responsible for generating anesthesia depth staging charts and decision recommendations.

[0187] Preferably, in one embodiment of the application, the module uses a 12-inch high-resolution touch screen (1920x1080 resolution), and the user interface is developed using the Qt framework, supporting real-time data visualization and interactive operation. The interface refresh rate is 60Hz, ensuring a smooth user experience.

[0188] Through the coordinated work of the above modules, the system can achieve millisecond-level real-time evaluation of anesthesia depth, providing timely and accurate decision support for anesthesiologists, significantly improving the safety and effectiveness of anesthesia management.

[0189] The advantages and applications of the present application are as follows:

[0190] 1. High-precision evaluation: through multi-modal signal fusion and deep learning technology, anesthesia depth evaluation accuracy is improved to more than 95%, significantly surpassing existing technologies. This high-precision evaluation can effectively reduce the risk of insufficient or excessive anesthesia, improving surgical safety.

[0191] 2. Real-time performance: using edge computing and optimized algorithm design, the system response time is reduced to less than 100ms, which can timely capture the rapid changes of anesthesia state and provide timely support for emergency handling.

[0192] 3. Personalized adaptation: dynamic time window technology can adaptively adjust according to the individual characteristics of patients and anesthesia stages, significantly improving the adaptability to different patients and different surgical stages, with an individual difference adaptability improvement of more than 40%. This is particularly important for handling anesthesia for special populations such as elderly patients, children, or obese patients.

[0193] 4. Comprehensive Assessment: The multi-task learning framework simultaneously predicts anesthesia depth, EEG frequency band energy ratios, autonomic nervous system balance indices, and drug concentrations, providing a comprehensive patient status overview for anesthesiologists, supporting precision anesthesia management. For example, in patients with liver dysfunction, drug concentration predictions can help doctors more accurately adjust dosages.

[0194] 5. Strong Anti-interference Ability: Innovative signal preprocessing and multi-modal fusion techniques significantly improve the system's resistance to common operating room disturbances. In the presence of electrosurgical interference, accuracy decreases by no more than 5%. This is particularly important in electrosurgical procedures such as laparoscopic surgery or neurosurgery.

[0195] 6. Explainable Design: Anesthesia depth staging charts visually demonstrate changes in anesthesia status, enhancing decision-making explainability and helping doctors make more informed clinical decisions. This visualization tool can help train young anesthesiologists, especially in teaching hospitals.

[0196] 7. Drug Use Optimization: Based on accurate anesthesia depth assessment, this method can help optimize anesthetic drug use, reducing drug use by an average of 20%-30% and reducing the risk of drug-related adverse reactions. This not only reduces patient recovery time but also reduces healthcare costs.

[0197] 8. Intraoperative Awareness Prevention: By precisely controlling anesthesia depth, this method can effectively prevent intraoperative awareness events, with no intraoperative awareness cases in clinical trials. This is particularly important for high-risk patients such as those undergoing cardiac surgery.

[0198] 9. Economic Benefits: By optimizing drug use and reducing complications, this method can significantly reduce healthcare costs. Preliminary estimates show that each surgery can save about 5%-10% of total costs.

[0199] 10. Remote Monitoring Support: The system supports remote data transmission and real-time analysis, allowing experts to remotely assist in the anesthesia management of complex cases. This is particularly valuable in areas where medical resources are unevenly distributed.

[0200] The anesthesia depth real-time assessment method and system of the present invention can be widely applied to various surgical and medical scenarios, including but not limited to:

[0201] 1. General Surgery: Such as laparoscopic cholecystectomy, appendectomy, etc. In these surgeries, the system can help anesthesiologists accurately control anesthesia depth and reduce common complications such as postoperative nausea and vomiting.

[0202] 2. Cardiac surgery: such as coronary artery bypass graft, valve replacement, etc. This type of surgery requires very high control of anesthesia depth, and the high precision evaluation of this system can significantly improve the safety of the operation.

[0203] 3. Neurosurgery: such as brain tumor resection, spinal surgery, etc. In these surgeries, precise control of anesthesia depth is crucial for protecting neural function.

[0204] 4. Day surgery: such as cataract surgery, arthroscopy, etc. This system can help achieve rapid anesthesia induction and recovery, improving the efficiency of day surgery.

[0205] 5. Intensive Care Unit (ICU): For patients who need long-term sedation, this system can help healthcare workers better manage sedation depth and reduce complications such as ventilator-associated pneumonia.

[0206] 6. Obstetric anesthesia: In cesarean section surgery, this system can help anesthesiologists minimize the impact on the fetus while ensuring the safety of the mother.

[0207] 7. Pediatric surgery: The physiological characteristics of children make anesthesia management more challenging. The personalized adaptation of this system can provide more accurate guidance for pediatric anesthesia.

[0208] 8. Surgery for elderly patients: Elderly patients often have multiple underlying diseases and different sensitivities to anesthetic drugs. This system can help achieve safer anesthesia management for elderly patients.

[0209] 9. Robot-assisted surgery: such as prostatectomy assisted by the Da Vinci surgical system. In such high-precision surgeries, this system can provide stable anesthesia depth control.

[0210] 10. Organ transplantation surgery: such as liver transplantation, kidney transplantation, etc. This type of surgery is long and risky, and this system can help maintain a stable anesthesia state for a long time.

[0211] Preferably, in one embodiment of the present application, the parameter settings of the system in different types of surgery can be fine-tuned. For example, in cardiac surgery, the weight of the autonomic nervous system balance index can be increased; in neurosurgery, more attention can be paid to the changes in specific frequency bands of EEG.

[0212] In summary, the anesthesia depth real-time evaluation method and system provided by the present application, through innovative algorithm design and system integration, effectively solves the problems faced by existing technologies, and makes important contributions to improving anesthesia safety, optimizing anesthesia management processes, and promoting the development of personalized precision anesthesia. It not only can improve the safety of surgery, but also can optimize resource utilization and reduce medical costs, and has important significance for promoting the development of modern anesthesiology.

[0213] To verify the superiority of the present application, an embodiment and three comparative examples were designed and detailed comparative tests were conducted. The test environment was the general surgery operating room of a certain third-grade hospital, and the test object was 100 patients receiving general anesthesia (age 18-65 years old, ASA classification I-II level).

[0214] Example 1: Anesthesia depth real-time evaluation method of the present application;

[0215] Comparative Example 1: Traditional BIS index monitoring method;

[0216] Comparative Example 2: Single EEG signal based on deep learning method;

[0217] Comparative Example 3: Multi-modal signal fusion method without using attention mechanism and dynamic time window.

[0218] Test indicators and methods:

[0219] 1. Anesthesia depth evaluation accuracy: three experienced anesthesiology experts evaluated the anesthesia depth according to the patient's clinical manifestations, physiological indicators and postoperative interview results, and the expert evaluation results were taken as the gold standard to calculate the evaluation accuracy of each method.

[0220] 2. System response time: a high-precision timer was used to record the time from signal input to output of anesthesia depth evaluation results.

[0221] 3. Individual difference adaptability: the evaluation accuracy standard deviation of different age groups (18-40 years old, 41-65 years old) and different ASA classification patients was calculated, and the smaller the standard deviation, the better the adaptability.

[0222] 4. Anti-interference ability: tests were conducted with and without electrotome interference, and the percentage decrease in evaluation accuracy was calculated.

[0223] 5. Intraoperative awareness prevention: the number of postoperative patient-reported intraoperative awareness events was recorded.

[0224] 6. Anesthetic drug dosage: the total amount of propofol and remifentanil used during the operation was recorded, and the drug dosage per kilogram of body weight was calculated.

[0225] The test results are shown in the following table:

[0226] Index Example 1 Comparative Example 1 Comparative Example 2 Comparative Example 3 Accuracy of anesthesia depth evaluation 95.2％ 82.5％ 88.7％ 90.1％ System response time 85 ms 1500 ms 210 ms 150 ms Individual difference adaptability (standard deviation) % 2.1％ 7.8％ 5.2％ 4.5％ Anti-interference ability (accuracy reduction) % 3.2％ 15.6％ 9.8％ 7.5％ Intraoperative awareness events 0 3 1 1 Propofol dosage (mg / kg / h) 4.2 5.8 5.1 4.8 Remifentanil dosage (μg / kg / h) 0.15 0.22 0.18 0.17

[0227] Based on the above test result analysis and discussion are as follows:

[0228] 1. Anesthesia depth assessment accuracy: The accuracy of Example 1 is significantly higher than other methods, reaching 95.2%. This demonstrates the effectiveness of the multi-modal signal fusion, hierarchical attention mechanism, and dynamic time window technology of the present invention. High accuracy means that anesthesia depth can be more accurately controlled, reducing the risk of complications.

[0229] 2. System response time: The response time of Example 1 is only 85ms, far superior to other methods. Fast response is crucial for timely capturing anesthesia state mutations, helping anesthesiologists make quick adjustments and improving surgical safety.

[0230] 3. Individual difference adaptability: The standard deviation of Example 1 is the smallest, at 2.1%, indicating that the method of the present invention can better adapt to individual differences of different patients. This adaptability is of great significance for achieving precise and personalized anesthesia management.

[0231] 4. Anti-interference ability: In the presence of an electrosurgical interference, the accuracy of Example 1 decreases the least, only 3.2%. This demonstrates that the present invention has stronger robustness in complex operating room environments and can provide more reliable anesthesia depth assessment. Please refer to Figure 10 , through this line chart, it can be clearly seen that the present invention (red line) maintains the highest accuracy under all interference levels, with the smallest slope of the line, indicating the strongest anti-interference ability. The lines of Comparative Examples 1-3 decrease significantly as the interference level increases, especially under moderate and severe interference, with a significant decrease in accuracy.

[0232] 5. Intraoperative awareness prevention: Example 1 did not have intraoperative awareness events, while other methods had varying degrees of occurrence. This indicates that the present invention can more effectively prevent intraoperative awareness, improving patient's surgical experience and satisfaction.

[0233] 6. Anesthetic drug dosage: The dosage of propofol and remifentanil in Example 1 is significantly lower than that of other methods. This shows that the present invention can help anesthesiologists more accurately control anesthesia depth, avoid excessive anesthesia, thereby reducing drug dosage, reducing the risk of drug-related adverse reactions, and possibly accelerating the patient's postoperative recovery.

[0234] Please refer to Figure 11 , through this 3D surface chart, it can be directly seen that the present invention provides more stable and accurate anesthesia depth assessment throughout the entire surgical process, with smoother surfaces and more reasonable frequency distribution. Based on the above test results, Example 1 is determined as the best embodiment. It performs best in all test indicators, especially in accuracy, response time, and anti-interference ability. The improvement of this comprehensive performance is due to the unique technical innovation of the present invention, including:

[0235] 1. The hierarchical attention fusion network (HAFN) effectively integrates the information of multi-modal physiological signals, improving the comprehensiveness and accuracy of the evaluation.

[0236] 2. The dynamic time window (DTW) technology enhances the system's adaptability to rapid changes in anesthesia state, improving the real-time performance of the evaluation.

[0237] 3. The multi-task learning framework not only improves the performance of the main task (anesthesia depth evaluation), but also provides additional valuable information (such as EEG frequency band energy ratio, autonomic nervous system balance index, etc.), providing comprehensive support for anesthesia management.

[0238] In summary, these test results fully demonstrate the significant advantages of the present invention in the field of real-time anesthesia depth evaluation. It not only improves the accuracy and real-time performance of the evaluation, but also enhances the adaptability and anti-interference ability of the system. These improvements are of great significance for improving surgical safety, optimizing anesthesia management, reducing drug use, and preventing complications. The present invention provides strong technical support for the development of personalized precision anesthesia and is expected to play an important role in clinical practice. It should be noted that the above description is only a preferred embodiment of the present invention and does not limit the present invention. Any modifications, equivalent replacements, improvements, etc. within the principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A deep learning-based method for real-time anesthesia depth assessment, comprising the following steps: obtaining multi-modal physiological signals of a patient, including electroencephalogram (EEG), electrocardiogram (ECG), photoplethysmogram (PPG), blood pressure (BP), and respiratory signal (RESP); preprocessing the multi-modal physiological signals; characterized in that the method further comprises: fusing the preprocessed multi-modal physiological signals using a hierarchical attention fusion network (HAFN) to obtain fused features; wherein the hierarchical attention fusion network comprises a signal-level attention layer, a time-scale attention layer, and a frequency component attention layer; adjusting the fused features in time scale based on a dynamic time window (DTW) to obtain adjusted features; wherein the dynamic time window uses a reinforcement learning method to adaptively adjust the size of the observation window according to the current signal features and historical states; inputting the adjusted features into a multi-task learning framework to simultaneously predict anesthesia depth classification, EEG frequency band energy ratio, autonomic nervous system balance index, and drug concentration; determining the anesthesia depth of the patient according to the output of the multi-task learning framework, and generating an anesthesia depth staging diagram; the processing steps of the hierarchical attention fusion network (HAFN) include applying signal-level attention to the input multi-modal physiological signals to obtain signal-level fused features, applying time-scale attention to the signal-level fused features to obtain time-scale fused features, and applying frequency component attention to the time-scale fused features to obtain the final fused features; the implementation steps of the dynamic time window (DTW) include defining a state space including the current signal features and the historical window size, defining an action space including the amplitude of adjusting the window size, defining a reward function based on signal stability and information quantity, and using an Actor-Critic architecture reinforcement learning model to learn the optimal window adjustment strategy by interacting with the environment; the processing steps of the multi-task learning framework include setting the main task as anesthesia depth classification, setting the auxiliary tasks including EEG frequency band energy ratio prediction, autonomic nervous system balance index prediction, and drug concentration estimation, constructing a shared feature extraction network and task-specific output layer, defining a multi-task loss function considering the loss and weight of each task, and optimizing the performance of all tasks simultaneously through a backpropagation algorithm; further comprising the following feature extraction steps: calculating the phase synchronization index (PSI) to quantify the synchronization between different brain regions or different physiological signals; calculating the multi-scale entropy (MSE) to evaluate the complexity of the signal in different time scales; extracting the symbolic dynamics feature (SDF) to convert the continuous signal into a symbolic sequence and extract the dynamics features of the system; the multi-task learning framework uses the following optimized model architecture: using dynamic depth separable convolution to dynamically adjust the size and grouping number of the convolution kernel according to the complexity of the input features; introducing attention mechanism in residual connection to realize attention-guided skip connection; using an adaptive activation function to automatically adjust the parameters of the activation function according to the needs of different layers.

2. The method of claim 1, wherein, The preprocessing step of the multi-modal physiological signal includes: adaptive wavelet denoising of the multi-modal physiological signal; wherein the best wavelet basis function and threshold are automatically selected by an unsupervised learning method; compressed sensing sampling of the denoised signal; wherein the sparsity of the signal is utilized to perform dimensionality reduction sampling through a random projection matrix; Online anomaly detection and correction of the sampled signal; wherein the local outlier factor (LOF) algorithm is used to detect and correct signal anomalies in real time.

3. The method of claim 1, wherein, The following post-processing steps are also included: Calculate the probability distribution of the depth of anesthesia; Detect abnormal states and generate alarms; Perform interpretability analysis to quantify the contribution of different input features to the prediction results; Generate anesthesia adjustment recommendations based on the prediction results and interpretability analysis.

4. The method of claim 1, wherein, The generation step of the anesthesia depth staging chart includes: Divide the depth of anesthesia into six stages: awake, mild sedation, moderate sedation, deep sedation, general anesthesia, and over-deep anesthesia; Use different colors to represent different anesthesia stages, and the depth of color represents the reliability of the stage; Draw a dynamic bar chart with time as the horizontal axis and anesthesia depth stage as the vertical axis.

5. A system for real-time assessment of depth of anesthesia implementing the method of any one of claims 1 to 4, characterized in that, The system includes: A multi-source signal acquisition module for acquiring multi-modal physiological signals of a patient; A signal preprocessing module for denoising, compressed sensing sampling, and anomaly detection of the acquired signals; A feature extraction module for calculating time-frequency features, phase synchronization index, multi-scale entropy, and symbolic dynamics features; A deep learning inference module containing a hierarchical attention fusion network, a dynamic time window adjustment unit, and a multi-task learning framework; A result analysis and display module for generating anesthesia depth staging charts and decision recommendations; The deep learning inference module is implemented on an edge computing device to ensure real-time performance; the output of the result analysis and display module is used to guide the adjustment of the acquisition parameters of the multi-source signal acquisition module, forming a closed-loop control.

Images