Intraoperative hypotension prediction method based on double attention mechanism and related products

Abstract

The present application relates to the medical clinical technical field, disclose a kind of intraoperative hypotension prediction method and related product based on double attention mechanism, using intraoperative monitoring physiological index, utilize the deep learning model based on double attention mechanism to carry out early prediction to intraoperative hypotension, using multi-head attention mechanism different weight is assigned to different indicators;Also using step attention mechanism is weighted for each time step, let different time step play different roles.The present application adopts physiological index time step multi-head attention mechanism and physiological index multi-head attention mechanism double attention mechanism, consider the influence and effect of physiological index time step and physiological index on intraoperative hypotension prediction simultaneously, is favorable to improve intraoperative hypotension prediction precision, so that doctor takes antihypertensive measure in advance, reduces the adverse risk and death risk of patient after operation, solves the correlation problem between physiological index and physiological index time step in prior art.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of medical clinical technology, and in particular to an intraoperative hypotension prediction method based on a double attention mechanism and related products. BACKGROUND

[0002] Intraoperative hypotension can bring a series of postoperative adverse symptoms to patients, and early prediction of intraoperative hypotension can help doctors take measures as soon as possible, thereby reducing the frequency of patients experiencing intraoperative hypotension. Intraoperative hypotension prediction is to predict whether intraoperative hypotension will occur in a future period of time by using historical intraoperative monitoring data. With the continuous monitoring of patients in the intensive care unit (ICU), a large amount of data is generated, which brings better convenience for predicting intraoperative hypotension. Intraoperative monitoring of heart rate, mean arterial pressure, systolic pressure, diastolic pressure, etc. are common physiological indicators for predicting intraoperative hypotension. These indicators are typical time series, so intraoperative hypotension prediction is generally classified as a multivariate time series binary classification task.

[0003] At present, some intraoperative hypotension predictions are based on arterial pressure waveform data. Features are extracted from the original arterial pressure waveform data by manually designing a feature extraction algorithm, and then a machine learning method such as logistic regression (LR), naive Bayes (NB), and support vector machine (SVM) is used to classify the extracted features. The above method needs to manually design features, while deep learning can automatically learn and identify features and is often used for intraoperative hypotension prediction.

[0004] According to the type of data used, it can be divided into two categories. One is to downsample the data from the arterial waveform data at a certain frequency (such as 100HZ), and then use a deep learning method for classification. However, blood pressure changes are closely related to other physiological indicators, and some researchers have begun to try the second method, which combines blood pressure indicators and other physiological indicators for intraoperative hypotension prediction. LSTM in deep learning methods can solve the long-time dependency problem of time series and is widely used in intraoperative hypotension prediction. CNN can extract local spatial features of multivariate time series and capture the correlation between variables through convolution operations, and some researchers have begun to try to use it for intraoperative hypotension prediction.

[0005] However, in the process of implementing the technical solutions of the embodiments of the present application, the present inventors have found that the above technical solutions at least have the following technical problems:

[0006] (1) Correlation between variables. No one has used attention mechanisms in the field of intraoperative hypotension prediction, and all related researches treat all indicators as equally important when using multiple physiological indicators for intraoperative hypotension prediction.

[0007] (2) The correlation problem between time steps. Time series are sorted in chronological order, and the data of the previous step and the data of the next step are closely related and play different roles. SUMMARY

[0008] To solve the above technical problems, the present application uses intraoperative physiological indicator data for monitoring, uses a deep learning model based on a double attention mechanism to predict intraoperative hypotension in advance, so that the doctor can take measures to lower blood pressure in advance, and reduce the risk of adverse reactions and death of the patient after the operation.

[0009] The above invention purposes are mainly achieved through the following technical solutions:

[0010] In a first aspect, the method for predicting intraoperative hypotension based on a double attention mechanism includes:

[0011] Monitoring intraoperative physiological indicator data;

[0012] Inputting the monitored physiological indicator data into a deep learning model based on a double attention mechanism to output an intraoperative hypotension prediction result;

[0013] The training method of the deep learning model based on a double attention mechanism includes:

[0014] Step 1, taking the intraoperative physiological indicator data for monitoring as the original sample data X, and processing it to obtain the input matrix that the model can receive

[0015] Step 2, using a multi-head attention mechanism to assign different weights to each time step of the physiological indicators in the input matrix to obtain a time step weight matrix;

[0016] Step 3, using a multi-head attention mechanism to assign different weights to each physiological indicator in the input matrix to obtain a physiological indicator weight matrix;

[0017] Step 4, extracting the spatial features of the physiological indicators from the time step weight matrix and the physiological indicator weight matrix respectively to obtain a time step spatial feature matrix and a physiological indicator spatial feature matrix;

[0018] Step 5, fusing the time step spatial feature matrix and the physiological indicator spatial feature matrix to obtain a new output matrix;

[0019] Step 6, calculating and outputting the probability of occurrence and non-occurrence of intraoperative hypotension according to the new output matrix.

[0020] In a second aspect, an intraoperative hypotension prediction device based on a dual attention mechanism comprises:

[0021] a physiological index module configured to monitor intraoperative physiological index data;

[0022] an intraoperative hypotension prediction module configured to input the monitored physiological index data into a deep learning model based on a dual attention mechanism, and output an intraoperative hypotension prediction result;

[0023] the deep learning model based on the dual attention mechanism comprises:

[0024] a preprocessing module configured to process the intraoperative physiological index data monitored as raw sample data X, and obtain an input matrix that can be received by the model

[0025] a Step Attention module configured to use a multi-head attention mechanism to assign different weights to each time step of the physiological index in the input matrix , and obtain a time step weight matrix;

[0026] a Variable Attention module configured to use a multi-head attention mechanism to assign different weights to each physiological index in the input matrix , and obtain a physiological index weight matrix;

[0027] an FCN module configured to extract spatial features of the physiological index from the time step weight matrix and the physiological index weight matrix respectively, and obtain a time step spatial feature matrix and a physiological index spatial feature matrix;

[0028] a Gate module configured to fuse the time step spatial feature matrix and the physiological index spatial feature matrix to obtain a new output matrix;

[0029] an output module configured to calculate and output the probability of occurrence and non-occurrence of intraoperative hypotension according to the new output matrix.

[0030] In a third aspect, a computer readable storage medium stores a computer program, and the computer program is executed by a processor to implement the steps of the above-mentioned intraoperative hypotension prediction method based on a dual attention mechanism.

[0031] In a fourth aspect, an electronic device comprises a memory, a processor, and a computer program stored in the memory and executable on the processor, and the processor implements the steps of the above-mentioned intraoperative hypotension prediction method based on a dual attention mechanism when executing the computer program.

[0032] Compared with the prior art, when using multiple physiological indicators for intraoperative hypotension prediction, in order to make important indicators play a greater role, the application uses a multi-head attention mechanism to assign different weights to different indicators. A step attention mechanism is also used to weight each time step, allowing different time steps to play different roles. Physiological indicators are typical time series, and their data are strictly arranged in chronological order. The data of the previous step and the data of the next step are closely related and play different roles, so the multi-head attention learns the time-dependent relationship of the time series. Different physiological indicators play different roles in predicting hypotension, such as indicators directly related to blood pressure (systolic pressure, diastolic pressure, mean arterial pressure), which play a more important role. Compared with the self-attention mechanism, the multi-head attention mechanism can learn different information from different representation spaces, so the multi-head attention mechanism is used to assign weights to each physiological indicator. In summary, the application adopts a dual attention mechanism of physiological indicator time step multi-head attention mechanism and physiological indicator multi-head attention mechanism, taking into account the influence and role of physiological indicator time steps and physiological indicators on intraoperative hypotension prediction, which is beneficial to improve the prediction accuracy of intraoperative hypotension, so that doctors can take measures to lower blood pressure in advance, reduce the risk of postoperative adverse events and death of patients, and solve the correlation problem between physiological indicators and between physiological indicator time steps in the prior art. BRIEF DESCRIPTION OF DRAWINGS

[0033] Figure 1 The schematic diagram of the overall architecture of the deep learning model based on the dual attention mechanism provided by the embodiment one of the application is provided.

[0034] Figure 2 The schematic diagram of the flow of the deep learning model training method based on the dual attention mechanism provided by the embodiment one of the application is provided.

[0035] Figure 3 The schematic diagram of the sliding window filling missing data provided by the embodiment one of the application is provided.

[0036] Figure 4 The specific flow schematic diagram of step 2 of the deep learning model training method based on the dual attention mechanism provided by the embodiment one of the application is provided.

[0037] Figure 5 The architecture schematic diagram of the Step Attention layer provided by the embodiment one of the application is provided.

[0038] Figure 6 The architecture schematic diagram of the Variable Attention layer provided by the embodiment one of the application is provided.

[0039] Figure 7A structural schematic diagram of a deep learning model training device based on a double attention mechanism is provided for the second embodiment of the present application.

[0040] Figure 8 A flowchart of an intraoperative hypotension prediction method based on a double attention mechanism is provided for the third embodiment of the present application.

[0041] Figure 9 A structural schematic diagram of an intraoperative hypotension prediction device based on a double attention mechanism is provided for the fourth embodiment of the present application. DETAILED DESCRIPTION

[0042] In order to enable personnel in the technical field to better understand the present application scheme, the technical solutions in the embodiments of the present application will be described clearly and completely below in conjunction with the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor should fall within the scope of protection of the present application.

[0043] Embodiment one

[0044] A model training method is used to train a deep learning model based on a double attention mechanism, and the model is used for intraoperative hypotension prediction, as shown in the following formula: Figure 1 The model includes a preprocessing layer, a Step Attention layer, a Variable Attention layer, an FCN layer, a Gate layer and an output layer, and the Step Attention layer and the Variable Attention layer are both connected to an FCN layer.

[0045] As shown in the following formula: Figure 2 The training method includes the following steps:

[0046] Step 1: The physiological index data monitored during the operation is taken as the original sample data X through the preprocessing layer, and the input matrix X' that can be received by the model is obtained by processing. The error data in the original sample data X is retained, and the missing data is filled with -1, and then a new sample data is generated by the sliding window method, and the data in the observation gap is labeled according to the definition of intraoperative hypotension in the gap window, and the label includes 0 and 1, wherein the label 1 represents that the corresponding data is hypotension data, and the label 0 represents that the corresponding data is normal blood pressure data, and in the present application, the definition of intraoperative hypotension is: "the mean arterial pressure is lower than 65mmHg for one minute".

[0047] Specifically, as shown in the following formula: Figure 3As shown, because there are data missing and errors in medical data (e.g. -34 for systolic blood pressure), in order to make the finally trained model have the ability to handle error data and missing data, the error data is retained, and -1 is used to fill in the missing data, because the error data is mostly negative. Then the original data is divided into a window of T minutes in size by the sliding window method, with a sliding step of 1 minute, where each window of size T contains a 20-minute observation window, an N-minute gap window and a 1-minute prediction window, with T = 20 + N + 1, N can be 5 minutes, 10 minutes and 15 minutes. The gap window is used to realize the prediction of intraoperative hypotension in advance of N minutes.

[0048] Step 2, using the multi-head attention mechanism of the Step Attention layer to assign different weights to each time step of the physiological indicators in the input matrix to obtain a time step weight matrix;

[0049] Step 3, using the multi-head attention mechanism of the Variable Attention to assign different weights to each physiological indicator in the input matrix to obtain a physiological indicator weight matrix;

[0050] Step 4, extracting the spatial features of the physiological indicators from the time step weight matrix and the physiological indicator weight matrix respectively to obtain a time step spatial feature matrix and a physiological indicator spatial feature matrix;

[0051] Step 5, fusing the time step spatial feature matrix and the physiological indicator spatial feature matrix to obtain a new output matrix;

[0052] Step 6, calculating and outputting the probabilities of intraoperative hypotension occurrence and non-occurrence according to the new output matrix.

[0053] The probabilities of the two categories are calculated by using the Softmax function, where the probability of category i is C is 2.

[0054] Compared with the technical effect achieved by the prior art: when using multiple physiological indicators for intraoperative hypotension prediction, in order to make important indicators play a greater role, the application uses a multi-head attention mechanism to assign different weights to different indicators. A step attention mechanism is also used to weight each time step, allowing different time steps to play different roles. Physiological indicators are typical time series, and their data is strictly arranged in chronological order. The data of the previous step and the data of the next step are closely related and play different roles, so the multi-head attention learns the time-dependent relationship of the time series. Different physiological indicators play different roles in predicting hypotension, such as indicators directly related to blood pressure (systolic pressure, diastolic pressure, mean arterial pressure) that play a more important role. Compared with the self-attention mechanism, the multi-head attention mechanism can learn different information from different representation spaces, so the multi-head attention mechanism is used to assign weights to each physiological indicator. In summary, the application adopts a double attention mechanism of physiological indicator time step multi-head attention mechanism and physiological indicator multi-head attention mechanism, and considers the influence and role of physiological indicator time step and physiological indicator on intraoperative hypotension prediction, which is beneficial to improve the prediction accuracy of intraoperative hypotension, so that doctors can take measures to lower blood pressure in advance, reduce the risk of adverse reactions and death of patients after surgery, and solve the correlation problem between physiological indicators and between physiological indicator time steps in the prior art.

[0055] In a preferred embodiment, as shown in Figure 4 Step 2 specifically includes:

[0056] Step 2-1, map the input matrix into three matrices Q, K and V, where V is an input feature vector, Q and K are feature vectors for calculating attention weights;

[0057] Step 2-2, calculate the similarity of Q and K, and then reduce the result by times to get the weight k , where d

[0058] Step 2-3, use the softmax function to normalize by row or column to get the weight vector

[0059] Step 2-4, weight the input vector V and to get

[0060] Step 2-5, multiple attention heads simultaneously calculate steps 2-1 to 2-4, concatenate the output calculated by each attention head, and then pass through a linear transformation to get the final matrix containing different time step weights

[0061] The physiological indicators are typical time series, and the data is strictly arranged according to the time sequence. The data of the previous step and the data of the next step are closely related and play different roles, so the multi-head attention is used to learn the time dependence of the time series. The architecture of the Step Attention layer is as shown in the following figure, and the shape of the input matrix X is 1200x11, containing 1200 time steps, and the multi-head attention mechanism is used to weight different time steps. Figure 5

[0062] The steps of a single attention mechanism are as follows:

[0063] (1) Map the input matrix X into three matrices Q, K and V.

[0064] (2) Calculate the similarity of Q and K, and then reduce the result by 8 times to get the weight a, where d is the dimension of the time series. k

[0065] (3) Use the softmax function to normalize by row or column to get

[0066] The multi-head attention mechanism is to calculate multiple attention heads at the same time.

[0067] Preferably, it further comprises a dropout layer. The dropout is used to simplify the complexity of the network and reduce the possibility of overfitting. The input matrix X is added to the output of the Dropout layer to avoid gradient disappearance, according to the idea of the residual network. Finally, the Batch Normalization is used to normalize each batch of data, so that the input of each batch is more similar, so that the network can focus on learning the differences between categories. The input matrix X is input into the Step Attention layer, and the output of the Step Attention layer is a matrix

[0068] In a preferred embodiment, the step 3 specifically comprises:

[0069] The transpose matrix of the input matrix X is input into the Variable Attention layer, and the multi-head attention mechanism is used to learn from different representation spaces to obtain a matrix According to the matrix W, different weights are assigned to each physiological indicator in the input matrix X. ​​​​​​​​​​​

[0070] Different physiological indicators play different importance in predicting hypotension, such as indicators directly related to blood pressure (systolic pressure, diastolic pressure, mean arterial pressure) play a more important role. Compared with self-attention mechanism, multi-head attention mechanism can learn different information from different representation spaces, so multi-head attention mechanism is used to assign weights to each physiological indicator. The architecture of the Variable Attention layer is shown in Figure 6 The input of this layer is the transpose matrix of the input matrix with a shape of 11x1200, and the processing procedure is referred to the processing procedure of the Step Attention layer described above. The transpose matrix After the Variable Attention layer, a matrix containing different physiological indicator weights is obtained.

[0071] In a preferred embodiment, the step 4 specifically comprises:

[0072] Each FCN layer contains three stacked convolution modules, and each convolution module contains a one-dimensional convolution. The time step weight matrix and the physiological indicator weight matrix are standardized by using BN (Batch Normalization) to make the input distribution of each batch of data more similar, so that the network can focus on learning the differences between categories. The standardized data are respectively input into the three stacked convolution modules, and the one-dimensional convolution of each convolution module is used for convolution operation to extract the spatial features of the physiological indicators. The ReLU function is used as the activation function to increase the nonlinearity of the neural network, and the time step spatial feature matrix and the physiological indicator spatial feature matrix and

[0073] If a nonlinear activation function is not used, the output of each layer is a linear combination of the input of the upper layer, so a nonlinear factor needs to be added to the neural network. The ReLU function is selected as the activation function in the present application, and its equation is: y = max{0, x}. This function has good sparsity and nonlinearity, which can better mine relevant features and fit training data.

[0074] In a preferred embodiment, the convolution operation by the one-dimensional convolution of each convolution module to extract the spatial features of the physiological indicators specifically comprises: in the one-dimensional convolution of each convolution module, a sliding window is used to slide from top to bottom in the time step weight matrix and the physiological indicator weight matrix, and the data in the window are multiplied with the data in the convolution kernel one by one and then summed, and the sliding step is 1. The size of the sliding window and the convolution kernel is 3*n, where n is the number of physiological indicators and is a positive integer.

[0075] Taking the above example of using 11 physiological indicators, the convolution operation uses a sliding window that slides from top to bottom in the time step weight matrix and the physiological indicator weight matrix, and multiplies the data in the window with the data in the convolution kernel one by one and then sums them. The sliding step size is 1, and the size of the sliding window and the convolution kernel is 3×11, with a quantity of 128.

[0076] Using three stacked convolutional modules, hypotension can be predicted for multiple physiological indicators. By extracting the spatial features of physiological indicators and their time steps, the correlation between variables can be captured, which helps to improve the accuracy of intraoperative hypotension prediction.

[0077] In a preferred embodiment, adding a Global Average Pooling operation after the last convolutional module can not only summarize the spatial information of the extracted features, but also reduce the number of parameters and computational cost, thereby reducing the possibility of overfitting.

[0078] In a preferred embodiment, step 5 specifically includes:

[0079] The SoftMax function is used to calculate the weights of the time step spatial feature matrix and the physiological index spatial feature matrix, respectively.

[0080] The time step spatial feature matrix is ​​multiplied by its corresponding weights to obtain a first matrix, and the physiological index spatial feature matrix is ​​multiplied by its corresponding weights to obtain a second matrix. The first matrix and the second matrix are then concatenated to obtain a new output matrix.

[0081] The role of the Gate layer is to fuse features. and First use the SoftMax function for and Calculate the weights g1 and g2 separately, and then let and Multiply by the corresponding weights g1 and g2, and finally concatenate them to form a new output matrix.

[0082] Integrating two features and The simplest method is to splice. However, simple concatenation treats both features as equally important. Since features extracted using different methods are often different, it's impossible to intuitively determine which feature is more beneficial for predicting intraoperative hypotension. Therefore, this invention uses a gate mechanism to treat both features... and By redistributing weights and multiplying the two features with their corresponding weights before concatenating them, the accuracy of intraoperative hypotension prediction can be further improved.

[0083] To illustrate the advantages of the technical solutions of the present application, a set of experimental data is used as follows:

[0084] Eleven common physiological indicators were used to predict intraoperative hypotension, namely systolic blood pressure (ART Systolic), diastolic blood pressure (ART Diastolic), mean arterial pressure (ART Mean), heart rate (HR), SpO2 1 pulse rate (SpO2 1 Pulse Rate), SpO2 1 saturation (SpO2 1 Saturation), SpO2 1 perfusion index (SpO2 1 Perfusion Index), carbon dioxide respiration rate (CO2 Respiration Rate), end-tidal CO2 (End-Tidal CO2), inspired O2 (Inspired O2), and end-tidal O2 (End-Tidal O2).

[0085] Because there are generally missing and erroneous data in medical data (for example, systolic blood pressure is -34), in order to enable the finally trained model to have the ability to process erroneous data and missing data, the erroneous data is retained, and -1 is used to fill in the missing data, and batch standardization is used to standardize each batch of samples, and the missing data filled in with -1 will be distinguished from other normal data after batch standardization.

[0086] Experiments were respectively conducted when the gap window size N was 5 minutes, 10 minutes and 15 minutes. Machine learning methods included Logistic Regression (LR), Naive Bayes (NB) and Support Vector Machine (SVM), and deep learning methods included LSTM, FCN and MultiAttention. The overall experimental results are shown in Table 1. The model of the present application is obviously superior to the three machine learning methods in all indicators. Compared with the LSTM model, the DoubleAttention_FCN model still performs more excellently in the intraoperative hypotension prediction task.

[0087] Table 1 Overall experimental results

[0088]

[0089]

[0090] In addition, the effectiveness of the DoubleAttention_FCN model is verified by ablation experiments, and the ablation experiment results are shown in Table 2. Although the performance of the model is reduced when the variable attention mechanism or the step attention mechanism is added alone on the basis of the FCN, especially the VariableAttention_FCN model performs poorly in various indicators, but the performance of the model is greatly improved when the variable attention mechanism and the step attention mechanism are added simultaneously on the basis of the FCN.

[0091] Table 2 Ablation experiment results

[0092]

[0093]

[0094] Embodiment two

[0095] The model training device is used for training a deep learning model based on a double attention mechanism, and the model is used for intraoperative hypotension prediction. The model includes a preprocessing layer, a Step Attention layer, a Variable Attention layer, an FCN layer, a Gate layer and an output layer. The Step Attention layer and the Variable Attention layer are both connected with an FCN layer.

[0096] As shown in Figure 7 , the device includes:

[0097] The preprocessing module is configured to process physiological index data monitored during an operation as original sample data X, and obtain an input matrix X' that can be received by the model.

[0098] The Step Attention module is configured to use a multi-head attention mechanism to assign different weights to each time step of the physiological indicators in the input matrix X' to obtain a time step weight matrix W.

[0099] The Variable Attention module is configured to use a multi-head attention mechanism to assign different weights to each physiological indicator in the input matrix X' to obtain a physiological indicator weight matrix V.

[0100] The FCN module is configured to extract spatial features of the physiological indicators from the time step weight matrix W and the physiological indicator weight matrix V respectively to obtain a time step spatial feature matrix and a physiological indicator spatial feature matrix.

[0101] The Gate module is configured to fuse the time step spatial feature matrix and the physiological indicator spatial feature matrix to obtain a new output matrix.​​

[0102] an output module configured to calculate and output probabilities of intraoperative hypotension occurrence and non-occurrence according to the new output matrix.

[0103] In a preferred embodiment, the Step Attention module is specifically configured to:

[0104] map the input matrix X into three matrices Q, K and V, where V is an input feature vector, Q and K are feature vectors for calculating attention weights; in step 2-2, similarity calculation is performed on Q and K, and the result is reduced by a factor of d k is the dimension of the time series; using a softmax function, normalization is performed by row or column to obtain a weight vector Multiple attention heads simultaneously perform steps 2-1 to 2-4, concatenate the outputs obtained by each attention head, and then perform linear conversion to obtain a final matrix containing weights of different time steps

[0105] In a preferred embodiment, the Variable Attention module is specifically configured to:

[0106] map the transpose of the input matrix X to a matrix input the matrix into the Variable Attention layer, and use a multi-head attention mechanism to learn from different representation spaces to obtain a matrix containing weights of different physiological indicators According to the matrix, different weights are assigned to each physiological indicator in the input matrix

[0107] In a preferred embodiment, the FCN module is specifically configured to:

[0108] Each FCN layer contains three stacked convolution modules, and each convolution module contains a one-dimensional convolution. BN is used to normalize the time step weight matrix and the physiological indicator weight matrix; the normalized data are respectively input into the three stacked convolution modules, and one-dimensional convolution is performed by the one-dimensional convolution of each convolution module to extract the spatial features of the physiological indicators, and a ReLU function is used as an activation function to increase the nonlinearity of the neural network, and time step spatial feature matrices and physiological indicator spatial feature matrices are extracted and

[0109] ​​​​​​​In a preferred embodiment, the Gate module is specifically used for:

[0110] The SoftMax function is used to calculate the respective weight of the time step space feature matrix and the physiological index space feature matrix respectively; the time step space feature matrix is multiplied by the corresponding weight to obtain a first matrix, and the physiological index space feature matrix is multiplied by the corresponding weight to obtain a second matrix; the first matrix and the second matrix are spliced to obtain a new output matrix.

[0111] The model training device of the embodiment of the application is corresponding to the model training method of the above embodiment, and realizes the corresponding functions. Since the steps of the model training method have been described in detail in the above embodiment, they will not be described here.

[0112] Embodiment three

[0113] The intraoperative hypotension prediction method based on the double attention mechanism, as shown in Figure 8 includes the following steps:

[0114] Monitoring intraoperative physiological index data;

[0115] Inputting the monitored physiological index data into a deep learning model based on a double attention mechanism to output an intraoperative hypotension prediction result;

[0116] The specific training method of the deep learning model based on the double attention mechanism can refer to embodiment one.

[0117] Embodiment four

[0118] The intraoperative hypotension prediction device based on the double attention mechanism, as shown in Figure 9 includes:

[0119] A physiological index module for monitoring intraoperative physiological index data;

[0120] An intraoperative hypotension prediction module for inputting the monitored physiological index data into a deep learning model based on a double attention mechanism to output an intraoperative hypotension prediction result;

[0121] The specific training method of the deep learning model based on the double attention mechanism can refer to embodiment one.

[0122] Embodiment five

[0123] A computer readable storage medium stores a computer program, and the computer program is executed by a processor to realize part or all steps of the above intraoperative hypotension prediction method based on the double attention mechanism.

[0124] The above has carried out the detailed introduction to the embodiment of the application, the principle and implementation mode of the application have been described by applying specific examples in this paper, the above embodiment description is only used for helping understanding the method of the application and its core idea; at the same time, for the general technical personnel in the art, according to the idea of the application, there will be changes in specific implementation mode and application range, and the above is described, the content of the specification should not be understood as the limitation of the application.

Claims

1. A method for predicting intraoperative hypotension based on a dual attention mechanism, characterized in that, include: Monitor intraoperative physiological data; The monitored physiological data are input into a deep learning model based on a dual attention mechanism, which outputs the prediction result of intraoperative hypotension. The model includes a preprocessing layer, a Step Attention layer, a Variable Attention layer, an FCN layer, a Gate layer, and an output layer. Both the Step Attention layer and the Variable Attention layer are connected to an FCN layer. The training method for the deep learning model based on the dual attention mechanism includes: Step 1: In the preprocessing layer, the physiological indicators monitored during the operation are used as the raw sample data. And process it to obtain the input matrix that the model can receive. ; Step 2, input matrix in Step Attention layer The mapping is done into three matrices Q, K, and V, where V represents the input feature vector, and Q and K are the feature vectors used to calculate the attention weights. Similarity is then calculated between Q and K, and the results are reduced. The weight is obtained by multiplying. ,in This refers to the time series dimension; the weight vector is obtained by normalizing the data by rows or columns using the softmax function. ; the input vector V and We obtain by weighted summation Multiple attention heads simultaneously compute the above steps. The outputs from each attention head are concatenated, and then a linear transformation is performed to obtain the final time-step weight matrix containing weights for different time steps. ; Step 3, convert the input matrix transpose matrix The input to the Variable Attention layer uses a multi-head attention mechanism to learn from different representation spaces, resulting in a physiological indicator weight matrix containing weights for different physiological indicators. According to the matrix The input matrix Different weights are assigned to each physiological indicator; Step 4: Each FCN layer contains three stacked convolutional modules, each containing a one-dimensional convolution, and Batch Normalization (BN) is used on the time step weight matrix. and physiological indicator weight matrix Standardization is performed; the standardized data is then input into three stacked convolutional modules, and convolution is performed through one-dimensional convolution in each module. The ReLU function is used as the activation function to increase the nonlinearity of the neural network, and the weight matrix at each time step is extracted. and physiological indicator weight matrix The spatial characteristics of physiological indicators were used to obtain the spatial feature matrix of time steps. and the spatial feature matrix of physiological indicators ; Step 5: At the Gate layer, the spatial feature matrix of the time step and the spatial feature matrix of the physiological indicators are fused to obtain a new output matrix; Step 6: Calculate and output the probability of intraoperative hypotension occurring and not occurring in the output layer based on the new output matrix.

2. The intraoperative hypotension prediction method based on a dual attention mechanism as described in claim 1, characterized in that, The method of extracting spatial features of physiological indicators by performing convolution operations through one-dimensional convolution of each convolution module specifically includes: in the one-dimensional convolution of each convolution module, a sliding window is used to slide from top to bottom in the time step weight matrix and the physiological indicator weight matrix, and the data in the window is multiplied with the data in the convolution kernel one by one and then summed. The sliding step size is 1, and the size of the sliding window and the convolution kernel is 3*n, where n is the number of physiological indicators and is a positive integer.

3. The intraoperative hypotension prediction method based on a dual attention mechanism as described in claim 2, characterized in that, Step 5 specifically includes: The SoftMax function is used to calculate the weights of the time step spatial feature matrix and the physiological index spatial feature matrix, respectively. The time step spatial feature matrix is ​​multiplied by its corresponding weights to obtain a first matrix, and the physiological index spatial feature matrix is ​​multiplied by its corresponding weights to obtain a second matrix. The first matrix and the second matrix are then concatenated to obtain a new output matrix.

4. A device for predicting intraoperative hypotension based on a dual attention mechanism, characterized in that, include: The physiological indicators module is used to monitor intraoperative physiological indicators. The intraoperative hypotension prediction module is used to input the monitored physiological index data into a deep learning model based on a dual attention mechanism and output the intraoperative hypotension prediction result. The deep learning model based on the dual attention mechanism includes: The preprocessing module is used to use intraoperatively monitored physiological data as raw sample data. And process it to obtain the input matrix that the model can receive. ; The Step Attention module will input the matrix The mapping is done into three matrices Q, K, and V, where V represents the input feature vector, and Q and K are the feature vectors used to calculate the attention weights. Similarity is then calculated between Q and K, and the results are reduced. The weight is obtained by multiplying. ,in This refers to the time series dimension; the weight vector is obtained by normalizing the data by rows or columns using the softmax function. ; the input vector V and We obtain by weighted summation Multiple attention heads simultaneously compute the above steps. The outputs from each attention head are concatenated, and then a linear transformation is performed to obtain the final matrix containing weights for different time steps. ; The Variable Attention module processes the input matrix. transpose matrix The input to the VariableAttention layer uses a multi-head attention mechanism to learn from different representation spaces, resulting in a matrix containing weights for different physiological indicators. According to the matrix The input matrix Different weights are assigned to each physiological indicator; The FCN module comprises three stacked convolutional modules, each containing a one-dimensional convolution. Batch normalization (BN) is used to standardize the time-step weight matrix and the physiological indicator weight matrix. The standardized data are then input into the three stacked convolutional modules, where convolution is performed using the one-dimensional convolution of each module to extract the spatial features of the physiological indicators. The ReLU function is used as the activation function to increase the nonlinearity of the neural network, resulting in the extraction of the time-step spatial feature matrix and the physiological indicator spatial feature matrix. and ; The Gate module is used to fuse the time step spatial feature matrix and the physiological index spatial feature matrix to obtain a new output matrix. The output module is used to calculate and output the probability of intraoperative hypotension occurring and not occurring based on the new output matrix.

5. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the intraoperative hypotension prediction method based on the dual attention mechanism as described in any one of claims 1 to 3.

6. An electronic device, the device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the intraoperative hypotension prediction method based on a dual attention mechanism as described in any one of claims 1 to 3.

Images