A method for monitoring anesthesia depth based on neural network

Abstract

The disclosed embodiment provides a method for monitoring the depth of anesthesia based on a neural network, including: collecting physiological signal data of a patient during surgery; preprocessing the physiological signal data to reduce the influence of noise and distortion; constructing a neural network model adapted to a non-stationary state based on the preprocessed physiological signal data; using the neural network model to predict the physiological signal data in real time to determine the patient's depth of anesthesia; and adjusting the dosage of anesthetic drugs according to the predicted result of the depth of anesthesia. The solution of the disclosed embodiment can solve the problem of inaccurate model prediction caused by complex changes in physiological signals during surgery under a non-stationary state.

Description

Technical Field

[0001] The present application relates to the field of anesthesia technology, and in particular to an anesthesia depth monitoring method based on a neural network. Background Art

[0002] The neural network-based anesthesia depth monitoring method uses advanced machine learning technology to evaluate and monitor the depth of anesthesia during surgery by analyzing the patient's physiological signals (such as heart rate, blood pressure, brain waves, etc.), thereby improving the safety and effectiveness of anesthesia. However, this method faces a major problem in practical applications, that is, during surgery under non-stationary conditions, due to the complexity and variability of the patient's physiological signals, the model's prediction results may be biased, affecting the accurate judgment of the depth of anesthesia. This is mainly because in non-stationary conditions, such as stress responses during surgery or dynamic changes in drug effects, the pattern of physiological signals will change rapidly, exceeding the data range during model training, thereby affecting its prediction performance. Summary of the invention

[0003] In view of this, an embodiment of the present disclosure provides an anesthesia depth monitoring method based on a neural network, which at least partially solves the problems existing in the prior art.

[0004] The present application provides a method for monitoring anesthesia depth based on a neural network, comprising:

[0005] Collect the patient's physiological signal data during surgery;

[0006] Preprocessing the physiological signal data to reduce the effects of noise and distortion;

[0007] Based on the preprocessed physiological signal data, a neural network model that adapts to non-stationary states is constructed;

[0008] Using the neural network model to predict the physiological signal data in real time to determine the anesthesia depth of the patient;

[0009] The dosage of the anesthetic drug is adjusted according to the predicted result of the anesthesia depth.

[0010] In a specific embodiment, constructing a neural network model adapted to a non-stationary state based on the preprocessed physiological signal data includes:

[0011] Select the appropriate time window size based on the characteristics of non-stationary signals;

[0012] Perform wavelet transform on physiological signals within the selected time window to extract multi-scale features;

[0013] Model building using dynamic recurrent neural networks;

[0014] The training process of the neural network model is optimized through an adaptive learning rate algorithm.

[0015] In a specific implementation, selecting a suitable time window size based on the characteristics of the non-stationary signal includes:

[0016] Obtain the data length L of the current physiological signal;

[0017] Determine the initial time window size T0;

[0018] Adjust the time window size based on the following formula: T = T0 * (L / L0), where L0 is the reference signal length;

[0019] If T is less than the preset minimum time window Tmin, then T = Tmin; otherwise, T remains unchanged;

[0020] Where L represents the data length of the current physiological signal, L0 represents the reference signal length, T0 represents the initial time window size, T represents the adjusted time window size, and Tmin represents the minimum time window.

[0021] In a specific embodiment, performing wavelet transform on the physiological signal within the selected time window to extract multi-scale features includes:

[0022] Select the wavelet basis function w;

[0023] Perform multi-level wavelet decomposition on the signal within the time window;

[0024] Extract detail coefficients and approximation coefficients at each level;

[0025] The eigenvector F(i) is selected based on the following formula: |d(i)| / (a(i) + ε), where d(i) is the i-th level detail coefficient, a(i) is the i-th level approximation coefficient, and ε is a constant to prevent the denominator from being zero;

[0026] In this formula, d(i) represents the i-th level detail coefficient, a(i) represents the i-th level approximation coefficient, ε represents a constant to prevent the denominator from being zero, and F(i) represents the i-th level eigenvector.

[0027] In a specific embodiment, the model construction using a dynamic recurrent neural network includes:

[0028] Initialize the number of hidden layer nodes H of the dynamic recurrent neural network;

[0029] Set the number of recursive layers R and the learning rate αr of each recursive layer;

[0030] Adjust the learning rate based on the following formula α = αr / (1 + γr * epoch), where epoch is the current training round and γr is the recursive decay factor;

[0031] Update the weight parameter of the dynamic recurrent neural network in each iteration by Wt+1 = Wt + α * ∇W, where ∇W is the gradient of the weight parameter;

[0032] Where αr represents the learning rate of each recursive layer, epoch represents the current training round, γr represents the recursive decay factor, α represents the adjusted learning rate, Wt represents the current weight parameter, ∇W represents the gradient of the weight parameter, and Wt+1 represents the updated weight parameter.

[0033] In a specific embodiment, the training process of optimizing the neural network model by an adaptive learning rate algorithm includes:

[0034] Get the current iteration error E;

[0035] Calculate the rate of change of error ΔE = (E - Eprev) / Eprev, where Eprev is the error of the previous iteration;

[0036] Adjust the learning rate η = η0 * (1 + β * ΔE) based on the following formula, where η0 is the initial learning rate and β is the error sensitivity coefficient;

[0037] If the error change rate ΔE is greater than the preset threshold θ, the current learning rate is maintained; otherwise, the learning rate is adjusted according to the formula;

[0038] Where E represents the current iteration error, Eprev represents the error of the previous iteration, ΔE represents the rate of change of the error, η0 represents the initial learning rate, β represents the error sensitivity coefficient, θ represents the preset threshold, and η represents the adjusted learning rate.

[0039] In a specific embodiment, constructing a neural network model adapted to a non-stationary state based on the preprocessed physiological signal data includes:

[0040] Perform time-frequency analysis on the preprocessed physiological signal data to extract multi-scale feature information;

[0041] Input the extracted multi-scale feature information into the multi-layer perceptron network;

[0042] Dynamically adjust the structure of the neural network based on the sliding window technology to adapt to the changes of non-stationary states;

[0043] Cross-validation techniques are used to evaluate and optimize the performance of neural network models to ensure the predictive accuracy of the models under non-stationary conditions.

[0044] In a specific embodiment, the method of dynamically adjusting the structure of the neural network based on the sliding window technology includes:

[0045] The physiological signal data is segmented according to preset time intervals to form multiple time windows;

[0046] Perform feature extraction on the data in each time window to obtain the window feature vector;

[0047] Adjusting the number of hidden layer nodes of the neural network based on the statistical characteristics of the window feature vector;

[0048] If the coefficient of variation VC within the time window satisfies the condition VC > λ, then a hidden layer node is added, otherwise a hidden layer node is reduced, where λ is a pre-set threshold, indicating the threshold of feature variability, and VC =σ / μ, σ indicates the standard deviation of the feature vector, and μ indicates the mean of the feature vector.

[0049] In a specific implementation, when the total number N of feature vectors in all time windows is less than a certain threshold T, ie, N < T, the total number of feature vectors in all time windows is recalculated, where T is a preset threshold.

[0050] The disclosed embodiment provides a method for monitoring the depth of anesthesia based on a neural network, including: collecting physiological signal data of a patient during surgery; preprocessing the physiological signal data to reduce the influence of noise and distortion; constructing a neural network model adapted to a non-stationary state based on the preprocessed physiological signal data; using the neural network model to predict the physiological signal data in real time to determine the patient's depth of anesthesia; and adjusting the dosage of anesthetic drugs according to the predicted result of the depth of anesthesia. The solution of the disclosed embodiment can solve the problem of inaccurate model prediction caused by complex changes in physiological signals during surgery under a non-stationary state. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] In order to more clearly illustrate the technical solutions of the exemplary implementation methods of the embodiments of the present disclosure, the drawings required for use in the embodiments will be briefly introduced below. It should be understood that the following drawings only show certain embodiments of the embodiments of the present disclosure and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without paying creative work.

[0052] Figure 1is a flowchart of a neural network-based anesthesia depth monitoring method of the present application;

[0053] Figure 2 It is a flowchart of building a neural network model adapted to non-stationary state based on preprocessed physiological signal data according to the first embodiment of the present application;

[0054] Figure 3 This is a flowchart of the second embodiment of the present application for constructing a neural network model that adapts to non-stationary states based on preprocessed physiological signal data. DETAILED DESCRIPTION

[0055] In the following, only some exemplary embodiments are briefly described. As those skilled in the art will appreciate, the described embodiments may be modified in various ways without departing from the spirit or scope of the present application. Therefore, the drawings and descriptions are considered to be exemplary and non-restrictive in nature.

[0056] Next, refer to Figure 1 , describes a method for monitoring the depth of anesthesia based on a neural network of the present invention. The method mainly includes the following five steps:

[0057] S101: collecting physiological signal data of the patient during the operation;

[0058] S102: preprocessing the physiological signal data to reduce the influence of noise and distortion;

[0059] S103: constructing a neural network model adapted to non-stationary states based on the preprocessed physiological signal data;

[0060] S104: using the neural network model to perform real-time prediction on the physiological signal data to determine the anesthesia depth of the patient;

[0061] S105: Adjusting the dosage of the anesthetic drug according to the prediction result of the anesthesia depth.

[0062] Specifically, the first step is to collect the patient's physiological signal data during the operation. In the operating room, various physiological monitoring devices are used to collect the patient's vital signs, such as electrocardiogram (ECG), electroencephalogram (EEG), blood oxygen saturation (SpO2), blood pressure (BP) and other data. These data can reflect the patient's current physiological state and are crucial for subsequent anesthesia depth assessment. For example, in a cardiac surgery, the medical team uses a variety of sensors and monitoring equipment to continuously record the patient's heart rate, brain waves, respiratory rate and other key physiological parameters. These data provide an important source of information for anesthesia depth monitoring.

[0063] Next, the second step is to preprocess the collected physiological signal data to reduce the impact of noise and distortion. During the operation, various external interferences and equipment noise may introduce a lot of noise, which will affect the quality of the data and the accuracy of the model. Preprocessing methods include filtering, baseline drift correction, outlier processing and other techniques to ensure the cleanliness and stability of the data. In one embodiment, by applying adaptive noise filters and wavelet transforms to remove power frequency interference and distortion caused by muscle activity in ECG and EEG signals, the quality of the data is significantly improved, providing a more reliable basis for subsequent analysis.

[0064] The third step is to build a neural network model that adapts to non-stationary states based on the preprocessed physiological signal data. In non-stationary states, traditional static models are difficult to cope with the drastic fluctuations of physiological signals during surgery. To this end, neural network architectures such as long short-term memory networks (LSTM) or convolutional neural networks (CNN) are used to capture the dynamic characteristics of time series data. For example, in the specific implementation, a model based on a hybrid structure of LSTM and convolutional layers is designed, which can effectively capture the physiological change trends of patients in different time periods, and adaptively adjust the parameters of the prediction model to improve the prediction accuracy in non-stationary states.

[0065] The fourth step is to use the neural network model to predict the physiological signal data in real time to determine the patient's anesthesia depth. The model analyzes the physiological signal data transmitted in real time and combines historical records to output the patient's anesthesia depth score, such as the BIS value or the AAI index. These scores can intuitively reflect the current anesthesia effect. Specifically, in a complex liver resection operation, the system monitored the patient's sudden increase in heart rate and changes in brain waves in real time, and responded quickly through the neural network model, predicting that the patient may have a risk of mild awakening, and immediately issued an alarm.

[0066] Finally, the fifth step is to adjust the dosage of the anesthetic drug according to the predicted results of the anesthesia depth. When the system detects that the patient's anesthesia depth is insufficient or too deep, it will automatically suggest that the doctor adjust the treatment plan to ensure that the patient maintains the best anesthesia state throughout the operation. For example, in the above-mentioned liver resection surgery, the doctor increased the dose of propofol in time according to the system's prompts, avoiding the patient's pain reaction that may be caused by insufficient anesthesia.

[0067] Through the above steps, especially by constructing a neural network model that adapts to non-stationary states, the entire monitoring method effectively solves the problem of inaccurate model prediction caused by complex changes in physiological signals during surgery, thereby improving the safety and success rate of surgery.

[0068] Next, refer to Figure 2, describes the first embodiment of the present invention, based on the preprocessed physiological signal data, to construct a neural network model that adapts to non-stationary states. The method mainly includes the following steps: selecting a suitable time window size; performing wavelet transform on the physiological signal in the selected time window to extract multi-scale features; using a dynamic recurrent neural network (DRNN) to construct the model; and optimizing the training process of the neural network model through an adaptive learning rate algorithm.

[0069] First, select an appropriate time window size based on the characteristics of non-stationary signals. Physiological signals often exhibit non-stationary characteristics during anesthesia, so it is crucial to select an appropriate time window to capture these changes. When selecting the time window size, it is necessary to consider the rate of change of the physiological signal and the stability of the characteristics. For example, in anesthesia depth monitoring, if the time window is too large, important transient information may be smoothed out; if the time window is too small, it may lead to an increase in noise. After a large number of experimental verifications, window sizes between 10 seconds and 30 seconds perform best in most cases.

[0070] Next, wavelet transform is performed on the physiological signal in the selected time window to extract multi-scale features. Wavelet transform is a method suitable for non-stationary signal analysis, which can provide time and frequency information at the same time. By performing wavelet transform on the physiological signal in the selected time window, features at different scales can be extracted, which can better reflect the dynamic changes of physiological signals. For example, specifically, assuming that the electrocardiogram (ECG) signal collected during anesthesia is subjected to wavelet transform, high-frequency, medium-frequency and low-frequency parts can be extracted, corresponding to different physiological events, such as heart beats and respiratory changes.

[0071] Then, a dynamic recurrent neural network (DRNN) is used to build the model. DRNN is a neural network that can process time series data and has a memory function. It is suitable for analyzing the non-stationarity and dynamics of physiological signals. In DRNN, the state of each neuron depends not only on the current input, but also on the state at the previous moment. This mechanism enables DRNN to extract complex temporal dependencies in a continuous data stream. In one embodiment, a long short-term memory network (LSTM) or a gated recurrent unit (GRU) can be used as a specific implementation of DRNN. These networks show excellent performance when processing long sequence data.

[0072] Finally, the training process of the neural network model is optimized through an adaptive learning rate algorithm. Adaptive learning rate algorithms such as Adam and RMSprop can automatically adjust the learning rate, speed up the convergence speed and improve the generalization ability of the model. Specifically, the Adam algorithm combines the advantages of momentum gradient and adaptive learning rate, and can dynamically adjust the step size during the optimization process to avoid falling into the local minimum. For example, the initial learning rate is set to 0.001, the momentum parameter β1 is 0.9, and the decay parameter β2 is 0.999. With these settings, the model can converge quickly in the early stage of training, and fine-tune the parameters in the later stage to achieve better results.

[0073] Through the above steps, the method can more effectively solve the problem of inaccurate model prediction caused by complex changes in physiological signals during surgery in a non-stationary state. For example, in one embodiment, by real-time monitoring and analysis of the electrocardiogram, electroencephalogram and respiratory signals of multiple patients during anesthesia surgery, the DRNN model constructed using this method can accurately predict the patient's anesthesia depth at different anesthesia stages, thereby providing reliable decision support for clinicians.

[0074] Next, the method of selecting a suitable time window size based on the characteristics of the non-stationary signal of the present invention is described. Specifically, the method includes obtaining the data length L of the current physiological signal, which means reading the physiological signals of the current patient from the monitoring device during anesthesia and measuring the time or number of data points of these signals. This is to ensure that the specific characteristics and length of the physiological signal can be taken into account during the analysis, because the data length of different time periods or different patients may vary greatly.

[0075] Determine the initial time window size T0. This step is to pre-set an initial value based on experience or historical data without any specific input or conditions. The value of T0 is usually the optimal value obtained based on past experiments or clinical practice, and is used as the time reference for subsequent calculations. For example, in one embodiment, T0 may be set to 5 seconds, which is a reasonable starting time window that is generally applicable to most situations.

[0076] The time window size T is adjusted based on the following formula: T = T0 * (L / L0), where L0 is the reference signal length. The core idea of ​​this formula is to dynamically adjust the time window size according to the actual length of the current physiological signal. By introducing the reference signal length L0, the adjusted T can be more in line with the characteristics of the actual signal and improve the accuracy of subsequent processing. The parameter L represents the data length of the current physiological signal, which usually ranges from a few seconds to a few minutes; L0 is the reference signal length, which is usually an empirical value, such as 10 seconds; T0 is the initial time window size, which usually ranges from 1 to 10 seconds; T is the adjusted time window size, and the optimal value depends on the specific situation. The meaning of this formula is that if the current signal is longer than the reference signal, the time window will increase accordingly; otherwise, it will decrease. This is because longer signals may contain more important information, and a larger time window is required to capture this information.

[0077] If T is less than the preset minimum time window Tmin, then T = Tmin; otherwise T remains unchanged. This step is to ensure that the time window size is not too small, so as to avoid missing important information due to the window being too small. The setting of Tmin is based on the system's minimum requirements for signal processing, and is usually a fixed small value, such as 1 second. Specifically, in one embodiment, if the adjusted T is 0.8 seconds, and Tmin is set to 1 second, then T will eventually be set to 1 second.

[0078] Through the above steps, the time window size can be dynamically adjusted for non-stationary physiological signals of different lengths and characteristics, so as to more effectively monitor the depth of anesthesia. For example, in actual operation, when the heartbeat signal of the patient during anesthesia suddenly becomes dense, resulting in a significant increase in the data length L, the system will automatically increase the time window size to ensure the integrity of the signal and the accuracy of the analysis.

[0079] Next, the present invention describes the method of performing wavelet transform on physiological signals within a selected time window to extract multi-scale features.

[0080] First, a wavelet basis function w is selected. This choice depends on the characteristics of the signal and the purpose of the study. Common wavelet basis functions include Haar wavelet, Db wavelet, Mexican Hat wavelet, etc. In one embodiment, Db4 wavelet is selected as the basis function because of its good localization characteristics. Specifically, this choice enables the wavelet transform to effectively capture the instantaneous changes and frequency characteristics in physiological signals.

[0081] Secondly, the signal in the time window is subjected to multi-level wavelet decomposition. This step involves decomposing the signal into components of different scales, and each level of decomposition obtains the first-level approximation coefficient and the first-level detail coefficient. In a specific example, four-level wavelet decomposition can be selected to obtain signal features from fine-grained to coarse-grained. This can more comprehensively capture the multi-scale characteristics of the signal.

[0082] Then, the detail coefficients and approximation coefficients of each level are extracted. During the wavelet decomposition process, each level will generate a set of detail coefficients and approximation coefficients, which respectively represent the high-frequency components and low-frequency components of the signal at that level. For example, the detail coefficient d(1) and approximation coefficient a(1) obtained from the first level of decomposition may mainly reflect high-frequency noise, while the approximation coefficient a(4) of the last level mainly reflects the overall trend of the signal.

[0083] Finally, the eigenvector F(i) = |d(i)| / (a(i) + ε) is selected based on the following formula, where d(i) is the i-th level detail coefficient, a(i) is the i-th level approximation coefficient, and ε is a constant to prevent the denominator from being zero. The range of the parameter d(i) is a real number, representing the i-th level wavelet detail coefficient, the range of a(i) is also a real number, representing the i-th level wavelet approximation coefficient, and ε is usually a very small positive number (such as 0.001). The selection of the optimal value depends on the actual application scenario and is generally determined by experiments. The significance of the formula is that the relationship between the detail coefficient and the approximation coefficient is normalized in the form of a ratio so as to maintain the consistency of the features under different signal strengths. The formula is set in this way because the high and low frequency components of physiological signals often have different importance. Taking the ratio of the detail coefficient to the approximation coefficient as the eigenvector can better reflect the multi-scale characteristics of the signal. For example, in anesthesia depth monitoring, physiological signals in certain frequency ranges may be more sensitive. This formula can highlight these important features and improve the accuracy of monitoring.

[0084] Next, the model building using a dynamic recurrent neural network (DRNN) of the present invention is described.

[0085] First, initialize the number of hidden layer nodes H of DRNN; this step is to determine the size of the hidden layer in DRNN, that is, the number of intermediate units in the network that process information. The number of nodes H is usually determined based on the complexity of the input data and the performance of the expected model. For example, in anesthesia depth monitoring, H may be set to an integer between 64 and 256 to effectively capture the patient's physiological signal characteristics.

[0086] Set the number of recursive layers R and the learning rate αr of each recursive layer; the number of recursive layers R represents the number of layers in DRNN used to process sequence data, and each layer can capture information at different time scales. The learning rate αr controls the step size of the model weight update. A higher learning rate can accelerate convergence, but may skip the optimal solution; a lower learning rate can more accurately find the solution that minimizes the loss function, but the training speed will be slower. In anesthesia depth monitoring, R can be set to 3 to 5, and the initial learning rate αr of each layer is usually set to 0.001.

[0087] The learning rate is adjusted based on the following formula α = αr / (1 + γr * epoch); this formula is used to gradually reduce the learning rate as training progresses. The gradual reduction of the learning rate helps the model converge quickly in the early stages of training and fine-tune the parameters in the later stages to achieve better results. Among them, γr is the recursive decay factor that determines the speed at which the learning rate decreases as the number of training rounds increases, and is usually between 0 and 0.1. In the anesthesia depth monitoring application, γr can be set to 0.01.

[0088] In each iteration, the weight parameters of DRNN are updated by Wt+1 = Wt + α * ∇W; the weight update step here is based on the back-propagation algorithm, by calculating the gradient of the loss function with respect to the weight and adjusting the weight according to the learning rate. Wt is the current weight, ∇W is the gradient of the weight parameter, and α is the adjusted learning rate. For example, in the tth iteration, if the current weight Wt is an n×m matrix, the calculated gradient ∇W is also a matrix of the same shape, and the weight parameter will be updated according to the above formula.

[0089] In one embodiment, suppose that a neural network-based anesthesia depth monitoring study is being conducted, and a DRNN model needs to be built to predict the patient's anesthesia state during surgery. First, the number of hidden layer nodes H of the DRNN is initialized to 128, the number of recursive layers R is 4, and the initial learning rate αr of each layer is 0.001. The recursive decay factor γr is set to 0.01 to ensure that the learning rate gradually decreases during training. Subsequently, during the training process, each round of iteration updates the weight parameters of the DRNN according to the adjusted learning rate α to ensure that the model can gradually approach the optimal state. After multiple training rounds, the DRNN can effectively identify changes in the patient's anesthesia depth and provide accurate monitoring results.

[0090] Next, the training process of optimizing the neural network model by the adaptive learning rate algorithm of the present invention is described.

[0091] First, obtain the current iteration error E. In this step, the system needs to calculate the gap between the predicted output of the neural network and the actual output in the current iteration round. The mean square error or cross entropy loss function is usually used to represent this gap. For example, in the anesthesia depth monitoring method based on a neural network, E can represent the error between the anesthesia depth value predicted by the model and the true value of the actual patient anesthesia depth.

[0092] Calculate the rate of change of the error ΔE = (E - Eprev) / Eprev, where Eprev is the error of the previous iteration. This formula is used to calculate the rate of change of the current error compared to the previous error in order to understand how fast the model's error is decreasing during training. If ΔE is a positive number, it means the error is increasing; if ΔE is a negative number, it means the error is decreasing. Assume that in a certain training, the current error E = 0.05, and the previous error Eprev = 0.06, then ΔE = (0.05- 0.06) / 0.06 ≈ -0.167.

[0093] Adjust the learning rate η = η0 * (1 + β * ΔE) based on the following formula, where η0 is the initial learning rate and β is the error sensitivity coefficient. The core of this formula is to dynamically adjust the learning rate according to the error change rate ΔE. When ΔE is a negative number, it indicates that the error is decreasing, so the learning rate can be appropriately increased to speed up the convergence; otherwise, the learning rate should be reduced to avoid oscillation caused by over-correction. Specifically, assuming that the initial learning rate η0 = 0.01, the error sensitivity coefficient β = 0.1, and ΔE = -0.167 in the previous step, the adjusted learning rate η = 0.01 * (1 + 0.1 * (-0.167)) ≈0.00833.

[0094] If the error change rate ΔE is greater than the preset threshold θ, the current learning rate is maintained; otherwise, the learning rate is adjusted according to the formula. The purpose of this step is to prevent frequent adjustments to the learning rate when the error change is not significant, which affects the stability of model training. Assuming the preset threshold θ = -0.05, the ΔE = -0.167 calculated in the previous step is less than -0.05, so the learning rate needs to be adjusted according to the formula.

[0095] In the above description, E represents the current iteration error, Eprev represents the error of the previous iteration, ΔE represents the rate of change of the error, η0 represents the initial learning rate, which is usually between 0.01 and 0.1, and the optimal value needs to be determined through experiments; β is the error sensitivity coefficient, which generally ranges from 0 to 0.1, and is used to control the sensitivity of the learning rate to the error change; θ represents the preset threshold, which is used to control the sensitivity of the learning rate adjustment, and the value range is between -0.1 and 0, and the specific value needs to be adjusted according to the actual situation; η represents the adjusted learning rate, which is used to guide the step size of the parameter update in the next iteration of the model. The purpose of this formula is to optimize the model training process by adaptively adjusting the learning rate, and to improve the convergence speed and final performance of the model. In one embodiment, when the model is just starting to train, due to the large initial error, the setting of the error sensitivity coefficient β and the preset threshold θ can help the system quickly reduce the error. As the training gradually approaches convergence, the system will automatically reduce the learning rate to avoid over-adjustment and cause the model performance to deteriorate.

[0096] Next, refer to Figure 3 , describing the second embodiment of the present invention, based on the preprocessed physiological signal data, building a neural network model that adapts to non-stationary states.

[0097] S301: Perform time-frequency analysis on the preprocessed physiological signal data to extract multi-scale feature information. The purpose of this step is to convert the physiological signal data into a feature vector containing time domain and frequency domain information. By analyzing the signal strength at different time periods and frequencies, subtle changes in physiological signals can be captured. For example, in anesthesia depth monitoring, the EEG signal can be decomposed into multiple time-frequency sub-bands by wavelet transform or short-time Fourier transform (STFT) to obtain the characteristics of EEG activity at different stages of anesthesia. Specifically, if the sampling rate is 256 Hz, the window size is 2 seconds, and the step size is 0.5 seconds, a series of time-frequency matrices can be obtained, each of which represents the spectrum distribution in a specific time period.

[0098] S302: Input the extracted multi-scale feature information into a multi-layer perceptron (MLP) network. This step involves feeding the extracted feature information into a neural network for further processing. MLP is a feed-forward artificial neural network that can achieve complex tasks by learning the nonlinear relationship between input and output. In one embodiment, the input layer can have 50 neurons corresponding to 50 multi-scale features, the hidden layer can select 3 layers, each with 64, 32 and 16 neurons, and the output layer can output an estimated value of the depth of anesthesia. Specifically, the ReLU activation function can be used to introduce nonlinearity.

[0099] S303: Dynamically adjust the structure of the neural network based on the sliding window technology to adapt to changes in non-stationary states. In practical applications, physiological signal data is non-stationary, that is, the statistical characteristics of the data change over time. The sliding window technology allows the model to dynamically adjust its own structure based on the data in the current window. For example, during anesthesia, the characteristics of the EEG signal may change as the patient enters different stages of anesthesia. In order to adapt to these changes, a fixed window size (such as 5 minutes) can be set, the window can be moved every certain period of time (such as 1 minute), and the weights and structure of the neural network can be adjusted according to the statistical characteristics of the data in the window. This dynamic adjustment can be achieved through adaptive learning rate or structural optimization algorithm to improve the prediction accuracy of the model at different stages of anesthesia.

[0100] S304: Use cross-validation techniques to evaluate and optimize the performance of the neural network model to ensure the prediction accuracy of the model in a non-stationary state. Cross-validation is a common method for evaluating the performance of machine learning models, especially when the data set is small. By dividing the data set into multiple subsets and training and validating the model on each subset, the generalization ability of the model can be estimated more accurately. For example, a 5-fold cross-validation can be used, that is, the data set is divided into 5 equal parts, 4 parts are used as training sets each time, and the remaining 1 part is used as a validation set, and a total of 5 validations are performed. The specific optimization indicator can be the mean square error (MSE) or the classification accuracy. By minimizing these indicators, the model can obtain better prediction performance in a non-stationary state. Minimizing MSE can ensure that the model maintains a high prediction accuracy in a non-stationary state.

[0101] Next, the specific steps of dynamically adjusting the structure of the neural network based on the sliding window technology of the present invention are described.

[0102] First, the physiological signal data is segmented according to the preset time interval to form multiple time windows. Specifically, the physiological signal data is usually a set of sequence data that changes over time, such as heart rate, blood pressure or brain waves. By setting a fixed time interval (such as 10 seconds), this set of data is segmented into multiple small time windows, each of which contains a continuous physiological signal. For example, in an anesthesia depth monitoring system, the patient's heart rate and brain wave data can be intercepted every 10 seconds to form multiple 10-second time windows.

[0103] The second step is to extract features from the data in each time window to obtain a window feature vector. This step aims to extract useful feature information from the physiological signal data in each time window. These features may include time domain features, frequency domain features, and nonlinear features. By combining these features into a feature vector, it can provide input for subsequent neural network processing. In one embodiment, the mean, variance, and power spectral density can be extracted from each 10-second EEG data to form a feature vector.

[0104] The third step is to adjust the number of hidden layer nodes of the neural network based on the statistical characteristics of the window feature vector. The goal of this step is to dynamically adjust the complexity of the neural network according to the changes in the feature vector within the time window. By calculating the statistical characteristics of the feature vector in each time window, such as the mean, variance, or coefficient of variation, it is decided whether to adjust the number of hidden layer nodes of the neural network. For example, the coefficient of variation (VC) can be used as an evaluation criterion, which is defined as the standard deviation of the feature vector divided by the mean, that is, VC = σ / μ. In this formula, σ represents the standard deviation of the feature vector, and μ represents the mean of the feature vector. The coefficient of variation is used to measure the degree of fluctuation of the feature vector, ranging from [0, +∞). The closer it is to zero, the smaller the data fluctuation, and the larger it is, the greater the fluctuation. The optimal value depends on the specific application scenario, and a reasonable threshold λ can usually be determined in advance through experiments.

[0105] In the fourth step, if the coefficient of variation (VC) within the time window satisfies the condition VC > λ, then add a hidden layer node, otherwise reduce a hidden layer node. This step is a continuation of the previous step, and the structure of the neural network is adjusted based on the evaluation results of the coefficient of variation. For example, if the coefficient of variation within a 10-second time window is greater than the preset threshold λ, it means that the characteristics within the time window fluctuate greatly, so it may be necessary to add a hidden layer node to improve the adaptability and predictive ability of the model; conversely, if the coefficient of variation is less than or equal to the threshold λ, then reduce a hidden layer node to reduce the complexity of the model and avoid overfitting. In this way, the structure of the neural network can be flexibly adjusted according to the dynamic characteristics of the data, so as to more accurately monitor the depth of anesthesia.

[0106] In general, through these four steps, the dynamic adjustment method of the neural network structure based on the sliding window technology can effectively cope with the fluctuating characteristics of physiological signal data and improve the performance and reliability of the anesthesia depth monitoring system.

[0107] Next, the step of adding a hidden layer node if the coefficient of variation (VC) within the time window satisfies the condition VC>λ, and reducing a hidden layer node otherwise, according to the present invention is described.

[0108] First, calculate the standard deviation σ of the feature vector in each time window. The goal of this step is to evaluate the degree of change of the feature vector in a given time window. The standard deviation is used to measure the degree of dispersion of the data distribution. The larger σ is, the more drastic the data fluctuation is.

[0109] Calculate the mean μ of the feature vector in each time window. The mean represents the average level of the feature vector in the time window. The calculation formula for the mean is ,in is each value in the feature vector, and N is the length of the feature vector in the time window.

[0110] Based on the above calculation results, the coefficient of variation VC = σ / μ is calculated for each time window. The coefficient of variation is a standardized measure used to measure the relative fluctuation of data. The larger the value of VC, the greater the change relative to the mean. The formula for the coefficient of variation is VC = σ / μ, where σ is the standard deviation and μ is the mean. This setting is to eliminate the influence of dimension so that data in different units can be compared.

[0111] If VC > λ, it means that the feature changes greatly in the current time window, and a hidden layer node is added, otherwise a hidden layer node is reduced, where λ represents the threshold of the coefficient of variation. This step determines whether to adjust the structure of the neural network by setting the threshold λ. If the change of the feature in the current time window is greater than the threshold λ, it means that the fluctuation of the feature is large, and the complexity of the model is increased to capture more change information; conversely, if the feature changes slightly, the hidden layer nodes are reduced to simplify the model and improve efficiency. The value of λ is usually adjusted according to the specific experimental results. The initial value can be selected between 0.1 and 1.0, and the optimal value needs to be verified through multiple experiments.

[0112] For example, in one embodiment, assuming that the time window length is 10 seconds, the collected feature vectors are [0.5, 0.6, 0.7, 0.8, 0.9, 1.0, 1.1, 1.2, 1.3, 1.4]. The mean μ = 0.95 and the standard deviation σ = 0.316 are calculated, thereby obtaining the coefficient of variation VC = σ / μ = 0.332). Assuming that the set threshold λ is 0.3, VC > λ at this time, so the system decides to add a hidden layer node. This shows that within the current 10-second time window, the feature has a large change amplitude, and adding nodes helps to capture these changes more accurately. On the contrary, if the feature vector is relatively stable and the calculated VC is less than λ, the system will reduce one hidden layer node to simplify the model. Specifically, this dynamic adjustment mechanism enables the neural network to automatically optimize the structure under different circumstances and improve the robustness and accuracy of the model.

[0113] Next, the method of calculating the mean μ of the feature vector in each time window of the present invention is described.

[0114] First, count the components of the feature vector in each time window. This means that in each specific time period (i.e., time window), each component of all collected feature vectors is accumulated. For example, in one embodiment, assuming that there is a time window of 30 seconds, and in this time period, a total of 5 feature vectors are collected, each of which contains 4 components (such as brain wave frequency, amplitude, waveform duration, and fluctuation range), then each component of these 5 feature vectors needs to be accumulated separately.

[0115] Next, calculate the sum of each component S. Here S represents the sum of each component over all feature vectors in the time window. For example, in the 5 feature vectors in the previous step, the sum of the first component is S1, the sum of the second component is S2, and so on, until the sum of the fourth component is S4.

[0116] Third, calculate the number of feature vectors n in each time window. This step is to determine how many feature vectors are collected in the time window. In the above example, the number is 5, that is, n = 5.

[0117] Finally, the mean μ = S / n is calculated based on S and n, where S represents the sum of the eigenvectors and n represents the number of eigenvectors in the time window. This step obtains the average value of each component in the time window by dividing the sum S of each component by the number of eigenvectors n. In the previous example, the mean of the first component is μ1 = S1 / 5, and the means of other components are calculated in the same way until μ4 is calculated.

[0118] The parameter S is the sum of each component in the time window, usually ranging from 0 to a positive number, depending on the data acquisition accuracy. The parameter n represents the number of eigenvectors in the time window, usually ranging from 1 to several integers, and the optimal value is determined by the specific system performance and data processing speed. The calculation formula of the mean μ μ = S / n represents the evaluation of the average situation of the eigenvector in the time window, the purpose of which is to smooth the data, reduce the impact of noise, and more accurately reflect the changing trend of the anesthesia state.

[0119] In practical applications, suppose that during an anesthesia monitoring process, the system sets the time window to 30 seconds, and collects 6 feature vectors in every 30 seconds. The feature vectors contain brain wave frequency, amplitude, waveform duration, and fluctuation range. Specifically, in a 30-second time window, the component data of each feature vector are as follows:

[0120] - Eigenvector 1: [10, 5, 8, 2]

[0121] - Eigenvector 2: [12, 6, 9, 3]

[0122] - Eigenvector 3: [11, 5, 7, 4]

[0123] - Eigenvector 4: [10, 4, 8, 1]

[0124] - Eigenvector 5: [11, 5, 7, 3]

[0125] - Eigenvector 6: [12, 6, 9, 2]

[0126] Calculate the sum of the individual components S1 to S4:

[0127] - S1 = 10 + 12 + 11 + 10 + 11 + 12 = 66

[0128] - S2 = 5 + 6 + 5 + 4 + 5 + 6 = 31

[0129] - S3 = 8 + 9 + 7 + 8 + 7 + 9 = 48

[0130] - S4 = 2 + 3 + 4 + 1 + 3 + 2 = 15

[0131] The number of eigenvectors n = 6. Calculate the mean of the components μ1 to μ4:

[0132] - μ1 = S1 / 6 = 66 / 6 = 11

[0133] - μ2 = S2 / 6 = 31 / 6 ≈ 5.17

[0134] - μ3 = S3 / 6 = 48 / 6 = 8

[0135] - μ4 = S4 / 6 = 15 / 6 ≈ 2.5

[0136] In this way, the mean μ of the feature vector in each time window is calculated and can be used for further neural network model input to improve the accuracy of anesthesia depth monitoring.

[0137] Next, the calculation of the mean μ = S / N based on S and n of the present invention is described.

[0138] First, calculate the eigenvector and value S_i in each time window. This step means that in each preset time window, the eigenvector values ​​in the time period are added to obtain the eigenvector and value S_i of the window. Here, the eigenvector is the key data point representing the depth of anesthesia extracted from the neural network model, which can include the patient's physiological signals such as heart rate, brain wave activity, etc. The range of the eigenvector and value S_i depends on the specific value and number of the eigenvector, and the optimal value is the actual accumulation result in each window.

[0139] Then, the sum of all time windows is accumulated, S = ΣS_i. This step involves summing the feature vector and value S_i of each time window to obtain the sum value S in all time windows. Through this accumulation process, the overall data summary can be obtained, providing a basis for subsequent mean calculation. The range of S is also dynamically changed, determined by the distribution of actual data and the number of windows.

[0140] Next, calculate the total number of eigenvectors in all time windows N = Σn_i. The purpose of this step is to count the total number of eigenvectors N contained in all time windows. n_i represents the number of eigenvectors in each time window. By summing the n_i values ​​of all windows, the total number of eigenvectors N can be obtained. The range here ranges from 1 to the sum of the actual number of eigenvectors in each window, and the optimal value is the actual cumulative result.

[0141] Finally, based on the above accumulated values, the total mean μ = S / N is calculated. Here, the sum value S is divided by the total number of eigenvectors N to obtain the average eigenvector value μ. μ represents the average level of the eigenvectors in all time windows. This value can help evaluate the overall state of the depth of anesthesia and provide important reference information for doctors. For example, in one embodiment, if there are 10 time windows in a certain period of time, and the number of eigenvectors in each window is 100, 120, 110, 90, 110, 120, 130, 100, 90 and 110 respectively, then the sum S = 100 + 120 + 110 + 90 + 110 + 120 +130 + 100 + 90 + 110 = 1180. The total number of feature vectors N = 100 + 120 + 110 + 90 + 110 + 120 + 130 + 100 + 90 + 110 = 1090. Therefore, the average value μ = 1180 / 1090 ≈ 1.0825. Specifically, μ ranges from 0 to positive infinity, depending on the characteristics and distribution of the actual data. The formula is set to S / N because it can accurately reflect the average level of feature vectors in all time windows, thereby better describing the overall anesthetic state.

[0142] Next, the calculation of the total number of feature vectors N = Σn_i in all time windows of the present invention is described. First, for each time window i, the number of feature vectors n_i in the time window needs to be recorded. The purpose of this step is to ensure that the data of each time segment is accurately captured, and n_i represents the number of feature vectors in the i-th time window, and i ranges from 1 to the total number of time windows m. Specifically, in one embodiment, assuming there is an anesthesia depth monitoring system based on a neural network, the system divides the monitoring data into 10-second time windows, and each time window generates a number of feature vectors for training the neural network. 50 feature vectors are recorded in the first time window, and 60 feature vectors are recorded in the next time window, and so on.

[0143] After that, the number of feature vectors in all time windows is accumulated to obtain the total number of feature vectors N = Σn_i. This process is to summarize the number of feature vectors extracted in the previous time windows, with the purpose of providing a total reference for subsequent data processing and model training. The symbol Σ in the formula represents the summation operation, and n_i represents the number of feature vectors in the i-th time window. For example, in the monitoring system mentioned above, if there are 100 time windows in total, then the number of all feature vectors in these 100 time windows needs to be added to obtain the value of N.

[0144] Subsequently, based on the accumulated results, the calculated N is checked to ensure its accuracy. This step is mainly to eliminate any possible data recording or calculation errors and ensure the reliability of subsequent analysis and modeling. In one embodiment, the final N can be double-confirmed to be correct by repeatedly checking the n_i value of each time window and re-performing the summing operation.

[0145] Finally, if the calculated N is less than a preset threshold T, that is, N < T, the system will trigger the process of recalculating the total number of feature vectors in all time windows. The threshold T here defines the minimum allowable total number of feature vectors, and its range and optimal value should be determined according to the actual needs of the specific application field. In one embodiment, T may be set to 5000, indicating that at least 5000 feature vectors need to be collected during a complete surgical procedure to ensure the effectiveness of model training. The rationality of this setting is that by ensuring that there is enough data to support model training, the prediction accuracy and stability can be improved. If N < T, for example, N is only 4000, this indicates that the data is insufficient and needs to be re-collected and processed to ensure the effectiveness and reliability of the model.

[0146] The present invention provides a neural network-based anesthesia depth monitoring method that uses the above-mentioned neural network model that adapts to non-stationary states to predict physiological signal data in real time to determine the patient's anesthesia depth. According to the prediction results of the anesthesia depth, the doctor can adjust the dosage of the anesthetic drug in time to ensure the safety and smooth progress of the surgical process.

[0147] The exemplary systems and methods of the present invention have been specifically shown and described with reference to the above embodiments, which are merely examples of the best modes for implementing the present systems and methods. It will be appreciated by those skilled in the art that various changes may be made to the embodiments of the systems and methods described herein when implementing the present systems and / or methods without departing from the spirit and scope of the present invention as defined in the appended claims.

Claims

1. A method for monitoring the depth of anesthesia based on a neural network, characterized in that: include: Collect the patient's physiological signal data during surgery; Preprocessing the physiological signal data to reduce the effects of noise and distortion; Based on the preprocessed physiological signal data, a neural network model that adapts to non-stationary states is constructed; Using the neural network model to predict the physiological signal data in real time to determine the anesthesia depth of the patient; adjusting the dosage of the anesthetic drug according to the predicted result of the anesthesia depth; in The step of constructing a neural network model adapted to a non-stationary state based on the pre-processed physiological signal data comprises: Select the appropriate time window size based on the characteristics of non-stationary signals; Perform wavelet transform on physiological signals within the selected time window to extract multi-scale features; Model building using dynamic recurrent neural networks; Optimize the training process of the neural network model through an adaptive learning rate algorithm; and The selecting of a suitable time window size based on the characteristics of the non-stationary signal includes: Obtain the data length L of the current physiological signal; Determine the initial time window size T0; Adjust the time window size based on the following formula: T = T0 * (L / L0), where L0 is the reference signal length; If T is less than the preset minimum time window Tmin, then T = Tmin; otherwise, T remains unchanged; Where L represents the data length of the current physiological signal, L0 represents the reference signal length, T0 represents the initial time window size, T represents the adjusted time window size, and Tmin represents the minimum time window.

2. A method for monitoring depth of anesthesia based on a neural network according to claim 1, characterized in that: The performing wavelet transform on the physiological signal within the selected time window to extract multi-scale features comprises: Select the wavelet basis function w; Perform multi-level wavelet decomposition on the signal within the time window; Extract detail coefficients and approximation coefficients at each level; The eigenvector F(i) is selected based on the following formula: |d(i)| / (a(i) + ε), where d(i) is the i-th level detail coefficient, a(i) is the i-th level approximation coefficient, and ε is a constant to prevent the denominator from being zero; In this formula, d(i) represents the i-th level detail coefficient, a(i) represents the i-th level approximation coefficient, ε represents a constant to prevent the denominator from being zero, and F(i) represents the i-th level eigenvector.

3. The method for monitoring depth of anesthesia based on a neural network according to claim 1, characterized in that: The method of constructing a model using a dynamic recurrent neural network includes: Initialize the number of hidden layer nodes H of the dynamic recurrent neural network; Set the number of recursive layers R and the learning rate αr of each recursive layer; Adjust the learning rate based on the following formula α = αr / (1 + γr * epoch), where epoch is the current training round and γr is the recursive decay factor; Update the weight parameter Wt+1 = Wt + α * ∇W of the dynamic recurrent neural network in each iteration, where ∇W is the gradient of the weight parameter; Where αr represents the learning rate of each recursive layer, epoch represents the current training round, γr represents the recursive decay factor, α represents the adjusted learning rate, Wt represents the current weight parameter, ∇W represents the gradient of the weight parameter, and Wt+1 represents the updated weight parameter.

4. The method for monitoring depth of anesthesia based on a neural network according to claim 1, characterized in that: The training process of optimizing the neural network model by the adaptive learning rate algorithm includes: Get the current iteration error E; Calculate the rate of change of error ΔE = (E - Eprev) / Eprev, where Eprev is the error of the previous iteration; Adjust the learning rate η = η0 * (1 + β * ΔE) based on the following formula, where η0 is the initial learning rate and β is the error sensitivity coefficient; If the error change rate ΔE is greater than the preset threshold θ, the current learning rate is maintained; otherwise, the learning rate is adjusted according to the formula; Where E represents the current iteration error, Eprev represents the error of the previous iteration, ΔE represents the rate of change of the error, η0 represents the initial learning rate, β represents the error sensitivity coefficient, θ represents the preset threshold, and η represents the adjusted learning rate.

5. The method for monitoring depth of anesthesia based on a neural network according to claim 1, characterized in that: The step of constructing a neural network model adapted to a non-stationary state based on the pre-processed physiological signal data comprises: Perform time-frequency analysis on the preprocessed physiological signal data to extract multi-scale feature information; Input the extracted multi-scale feature information into the multi-layer perceptron network; Dynamically adjust the structure of the neural network based on the sliding window technology to adapt to the changes of non-stationary states; Cross-validation techniques are used to evaluate and optimize the performance of neural network models to ensure the predictive accuracy of the models under non-stationary conditions.

6. The method for monitoring depth of anesthesia based on a neural network according to claim 5, characterized in that: The structure of the neural network dynamically adjusted based on the sliding window technology includes: The physiological signal data is segmented according to preset time intervals to form multiple time windows; Perform feature extraction on the data in each time window to obtain the window feature vector; Adjusting the number of hidden layer nodes of the neural network based on the statistical characteristics of the window feature vector; If the coefficient of variation VC within the time window satisfies the condition VC > λ, then a hidden layer node is added, otherwise a hidden layer node is reduced, where λ is a pre-set threshold, indicating the threshold of feature variability, and VC = σ / μ, σ indicates the standard deviation of the feature vector, and μ indicates the mean of the feature vector.

7. The method for monitoring depth of anesthesia based on a neural network according to claim 6, characterized in that: When the total number of feature vectors N in all time windows is less than a certain threshold T, that is, N < T, the total number of feature vectors in all time windows is recalculated, where T is a pre-set threshold.

Images