Intraoperative blood pressure serialization generation method for multivariable relation perception local enhancement

Abstract

The invention relates to the technical field of blood pressure prediction, in particular to a multivariable relation sensing local enhanced intraoperative blood pressure serialization generation method, which comprises the following steps of: establishing a multivariable relation sensing local enhanced intraoperative blood pressure serialization generation model which comprises a sequence representation encoder, a sequence generation decoder and a full connection layer; the sequence representation encoder comprises a first convolution attention layer, a first multivariable relation sensing module, a first standardization layer and a self-attention focusing layer. The sequence generation decoder comprises a second convolution attention layer, a second multivariable relation sensing module, a second standardization layer, a third convolution attention layer, a third multivariable relation sensing module and a third standardization layer. According to the method, the systolic pressure, diastolic pressure and average arterial pressure of the patient in an operation can be continuously subjected to serialization generation prediction, so that an anesthetist is assisted to pre-judge whether the patient has a hypotension risk in a period of time in the future.

Description

Technical Field

[0001] The present invention relates to the technical field of blood pressure prediction, and in particular to a method for generating intraoperative blood pressure serialization with multivariate relationship-aware local enhancement. Background Art

[0002] Globally, approximately 300 million surgeries are performed each year, and the 30-day mortality rate and risk of complications after surgery have always been one of the issues of great concern in the medical field. Intraoperative hypotension (IOH) is a common risk event during surgery, which is closely related to the incidence of postoperative complications and postoperative mortality. If intraoperative hypotension events can be accurately predicted, it will help clinical anesthesiologists to intervene in the patient's risk status in a timely manner, which will have a positive significance in reducing the patient's risk of postoperative complications and mortality.

[0003] At present, the prediction of intraoperative hypotension risk is mainly achieved through retrospective analysis of the multivariate vital sign monitoring data of patients during surgery, which only considers the association between hypotension risk events at a certain moment and limited historical data within a period of time. That is, a binary classification model is established to predict whether intraoperative hypotension events will occur at a certain moment in the future by learning existing data. This defines intraoperative hypotension events as instantaneous events with only two states: occurrence and non-occurrence, which is obviously inconsistent with the facts. This ignores the clinical significance of the development of hypotension during these time intervals, increases the burden of intraoperative decision-making by anesthesiologists, and limits the clinical application scenarios of intraoperative hypotension prediction models.

[0004] In addition, the traditional binary classification model is based on the implementation of statistical machine learning algorithms. These studies mainly focus on the effectiveness verification of the application of statistical machine learning models in the prediction of intraoperative hypotension and the analysis of the characteristics of intraoperative hypotension. It is impossible to achieve real-time prediction of intraoperative hypotension, which limits its clinical application. As deep learning models have shown excellent performance in prediction tasks in general fields, researchers have recently begun to introduce neural network models such as LSTM, CNN, and Transformer in deep learning algorithms into the study of intraoperative hypotension prediction methods. Compared with statistical machine learning models, they have better performance in the task of intraoperative hypotension risk prediction.

[0005] However, from a clinical perspective, when predicting intraoperative hypotension, clinical anesthesiologists usually combine the patient's preoperative basic examination data and intraoperative vital sign monitoring data, and judge whether to actively administer fluids, vasopressors or inotropic drugs to deal with the possible risk of intraoperative hypotension based on clinical evidence within a certain time frame. The main drawback of this method is the delayed reaction time, which means that the patient may suffer from the adverse effects of hypotension or even injury before receiving appropriate intervention.

[0006] In addition, from a technical point of view, the use of a binary classification model based on a statistical machine learning algorithm will, on the one hand, lead to the model's reliance on the definition of hypotension, that is, to distinguish positive and negative samples based on a specific blood pressure threshold to predict the occurrence of intraoperative hypotension, and the definition of intraoperative hypotension is not unified in medical literature. Although the most widely used definition of intraoperative hypotension is MAP below 65 mmHg. However, a literature review revealed that there were 140 different definitions of hypotension in 130 studies. Some of these definitions are fixed thresholds, such as considering systolic blood pressure below 100 mmHg or 80 mmHg as hypotension, while others are percentages of decrease relative to the patient's baseline blood pressure, such as a decrease of more than 20% or 30%. The baseline is a single blood pressure measurement before surgery or before anesthesia induction. Therefore, it is obviously not enough to rely solely on static blood pressure thresholds to predict intraoperative hypotension. On the other hand, the binary classification model based on a statistical machine learning algorithm cannot continuously predict the dynamic changes of blood pressure. The occurrence of intraoperative hypotension is a certain process. Relying solely on the prediction results at a certain moment, there is no sufficient evidence to assist anesthesiologists in making decisions, which will increase the decision-making burden of anesthesiologists.

[0007] To address this practical problem, the present invention proposes and designs a multivariate relationship-aware locally enhanced intraoperative blood pressure serialization generation method. Summary of the invention

[0008] The present invention provides a method for serializing and generating intraoperative blood pressure with multivariate relationship perception and local enhancement. According to high-definition intraoperative multivariate vital sign data continuously collected by invasive monitoring equipment during the patient's operation, the method continuously generates and predicts the patient's systolic arterial pressure (SAP), diastolic arterial pressure (DAP) and mean arterial pressure (MAP) during the operation, thereby assisting anesthesiologists in predicting whether the patient will have a risk of hypotension in the future.

[0009] To achieve the above object, the present invention adopts the following technical solutions: A method for generating intraoperative blood pressure serialization with multivariate relationship perception and local enhancement is provided. A model for generating intraoperative blood pressure serialization with multivariate relationship perception and local enhancement is established. The model comprises a sequence representation encoder, a sequence generation decoder and a fully connected layer. The sequence representation encoder comprises a first convolutional attention layer, a first multivariate relationship perception module, a first normalization layer and a self-attention focusing layer. The sequence generation decoder comprises a second convolutional attention layer, a second multivariate relationship perception module, a second normalization layer, a third convolutional attention layer, a third multivariate relationship perception module and a third normalization layer. The specific steps are as follows: Collect various vital sign data collected by the patient through the vital sign monitoring equipment during the operation to form an input sequence; Add the input sequence, position code and global timestamp code to get the time-coded input vector; In the first convolutional attention layer, the time-encoded input vector is used to generate the corresponding query matrix, key matrix, and value matrix through different weight matrices, and then input into the first convolutional attention layer for feature extraction to obtain the first self-attention. In the first multivariate relationship perception module, the first self-attention is firstly globally averaged pooled in the time dimension, and then the dependency between input channels is captured through a linear layer and an activation function to obtain the first relationship matrix; In the first normalization layer, the first relationship matrix is ​​concatenated with the time-coded input vector, and the concatenated matrix is ​​normalized using Z-Score normalization to obtain a first normalized output; In the self-attention focusing layer, the first normalized output is convolved, activated, and max-pooled, and the sequence representation encoder is Layer passed to Layer, after traversal, the final hidden representation of the sequence representation encoder is obtained; Get the input vector for the sequence generation decoder , the formula is as follows: ; in, represents the position code, It indicates that the output prediction is guided by the starting sequence, which is extended to a generation method, representing the length of the target blood pressure sequence before the prediction The sequence slice of is used as the starting mark of prediction. is a placeholder for the target sequence to be predicted; is the feature dimension after input representation, To predict the target blood pressure sequence length, here , represents the diastolic blood pressure, mean arterial pressure, and systolic blood pressure that need to be predicted, Yes and Perform splicing operations.

[0010] In the second convolutional attention layer, the input vector , through different weight matrices, the corresponding query matrix, key matrix, and value matrix are generated respectively, and input into the second convolutional attention layer for feature extraction to obtain the second self-attention; In the second multivariate relationship perception module, the second self-attention is first globally averaged pooled in the time dimension, and then the dependencies between input channels are captured through linear layers and activation functions to obtain the second relationship matrix; In the second normalization layer, the second relationship matrix is ​​compared with the input vector Perform splicing, and standardize the spliced ​​matrix using Z-Score standardization to obtain a second standardized output; In the third convolutional attention layer, the final hidden representation of the sequence representation encoder is used to generate the corresponding query matrix and key matrix through different weight matrices, and the second standardized output is used to generate the value matrix through the value weight matrix. Then, the three matrices are input into the third convolutional attention layer for feature extraction to obtain the third self-attention. In the third multivariate relationship perception module, the third self-attention is first globally averaged pooled in the time dimension, and then the dependencies between input channels are captured through linear layers and activation functions to obtain the third relationship matrix; In the third standardization layer, the third relationship matrix is ​​concatenated with the second standardized output, and the concatenated matrix is ​​standardized using Z-Score standardization to obtain a third standardized output; In the fully connected layer, the third standardized output is processed using the following formula to output the target blood pressure sequence; ; in, is the weight matrix of the fully connected layer, is the bias vector, is the activation function, Generate the output of the decoder for the sequence, i.e., the third normalized output; is the target blood pressure sequence.

[0011] In this specification, the sequence refers to the encoder input sequence , Positional encoding and global timestamp encoding To realize the sequence representation function, Sequence and The feature dimensions of different global timestamps after input representation are ; Use fixed position encoding to retain the relative position information in the sequence, as shown in the following formula: ; ; in, , in order to adjust the dimension, a one-dimensional convolution filter is used and the input sequence is projected onto Dimension vector , and finally get the time-coded input vector , as shown below: ; in , is a learnable embedding, is a one-dimensional convolution operation, Encode the position.

[0012] In this specification, in the first convolutional attention layer, the second convolutional attention layer, and the third convolutional attention layer, the multi-head attention sublayer converts the hidden state Converted into Different query matrices , key matrix , value matrix ; , The conversion method uses a kernel size of , a one-dimensional convolution with a stride of 1, that is, , ,and Use a convolution kernel size of 1 to construct, , which is equivalent to the linear transformation, is the parameter matrix, , , ; Calculate ProbSparse self-attention , is the ProbSparse self-attention calculation, is the softmax function, is the scaling factor, as shown below; .

[0013] In this specification, in the first multivariate relationship perception module, the second multivariate relationship perception module and the third multivariate relationship perception module, the squeezed excitation block is embedded in each convolutional attention layer as a supplement to the attention mechanism, and the output after the convolutional attention layer is In order to further extract the correlation of input channels, we first aggregate the information in the time dimension to obtain the average value of the input channels. , and then captures the dependencies between input channels , as shown in the following formula; ; ; in, express No. Input channels, represents the length of the input sequence, Represents the aggregation time dimension after The average value of the input channels, and Refers to the activation function and ,in , Represents learning the weight coefficients of the input channels using two linear layers, is the hyperparameter that needs to be set; the input channel dimension is reduced to the original through the first linear layer times to achieve dimensionality simplification, reduce the number of model parameters and computational complexity, and compress and integrate the channel information of the input feature map to extract the correlation of the input channels; the number of channels is restored to the original number of channels through the second linear layer, which can better utilize the original channel information, ensure feature diversity and expression ability, and effectively learn the weight coefficients of different input channels through the joint action of the fully connected descending channel and the fully connected ascending channel.

[0014] In this specification, in the first standardization layer, the second standardization layer, and the third standardization layer, the mean and standard deviation of each feature are first calculated, and the transformation process shown in the following formula is performed on the observed value of each feature at a time point, ; in, It's time point The observed value of and The mean and standard deviation of each feature are calculated respectively; in the prediction stage of the model, the output is denormalized to the actual value using the standard deviation and mean during standardization.

[0015] In this specification, in the self-attention focusing layer, the formula for extracting the dominant feature is as follows: ; Among them Operations including convolutional attention layers and multivariate relational awareness modules; It is the maximum pooling operation; It is a one-dimensional convolution operation; As the activation function, nonlinear transformation is introduced to further optimize the representation of features. Finally, one-dimensional convolution and maximum pooling operations are performed on the input feature map to focus the feature map, downsample the feature map to half of its original size to reduce the computational complexity and retain key information, and represent the encoder from the sequence. Layer passed to The whole focusing operation process is repeated, and the number of self-attention focusing layers is gradually reduced by building multiple copies of the main stack. Finally, the final hidden representation of the sequence representation encoder is obtained by connecting the outputs of all stacks.

[0016] In summary, the present invention has at least the following beneficial effects: The present invention can assist clinical anesthesiologists in predicting the risk of intraoperative hypotension in patients in advance, thereby achieving timely intervention risks, avoiding intraoperative death and reducing the risk of postoperative complications, thereby effectively reducing the physical damage to patients caused by surgery and the medical expenses of patients during treatment. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0018] Figure 1 It is a schematic diagram of the multivariate relationship-aware locally enhanced intraoperative blood pressure serialization generation method involved in the present invention.

[0019] Figure 2 It is a schematic diagram of the processing process of the sequence representation function involved in the present invention.

[0020] Figure 3 Schematic diagram of the convolutional attention mechanism of the convolutional attention layer involved in the present invention.

[0021] Figure 4 Schematic diagram of the multivariate relationship perception module involved in the present invention.

[0022] Figure 5 It is a schematic diagram of the sequence generation decoder architecture involved in the present invention.

[0023] Figure 6 A schematic diagram of the prediction of intraoperative hypotension events involved in the present invention.

[0024] Figure 7 It is a schematic diagram of the intraoperative hypotension sequence prediction involved in the present invention. DETAILED DESCRIPTION

[0025] 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 embodiments of the present invention. Therefore, the drawings and descriptions are considered to be exemplary and non-restrictive in nature.

[0026] The disclosure below provides many different embodiments or examples to implement different structures of the embodiments of the present invention. In order to simplify the disclosure of the embodiments of the present invention, the parts and settings of specific examples are described below. Of course, they are only examples, and the purpose is not to limit the embodiments of the present invention. In addition, the embodiments of the present invention can repeat reference numbers and / or reference letters in different examples, and this repetition is for the purpose of simplification and clarity, which itself does not indicate the relationship between the various embodiments and / or settings discussed.

[0027] The embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0028] like Figure 1 As shown, this embodiment provides a method for generating intraoperative blood pressure serialization with multivariable relationship perception and local enhancement, and establishes a multivariable relationship perception and local enhancement intraoperative blood pressure serialization generation model, which includes a sequence representation encoder, a sequence generation decoder and a fully connected layer, wherein the sequence representation encoder includes a first convolutional attention layer, a first multivariable relationship perception module, a first normalization layer and a self-attention focusing layer, and the sequence generation decoder includes a second convolutional attention layer, a second multivariable relationship perception module, a second normalization layer, a third convolutional attention layer, a third multivariable relationship perception module and a third normalization layer; the specific steps are as follows: Collect various vital sign data collected by the patient through the vital sign monitoring equipment during the operation to form an input sequence; Add the input sequence, position code and global timestamp code to get the time-coded input vector; In the first convolutional attention layer, the time-encoded input vector is used to generate the corresponding query matrix, key matrix, and value matrix through different weight matrices, and then input into the first convolutional attention layer for feature extraction to obtain the first self-attention. In the first multivariate relationship perception module, the first self-attention is firstly globally averaged pooled in the time dimension, and then the dependency between input channels is captured through a linear layer and an activation function to obtain the first relationship matrix; In the first normalization layer, the first relationship matrix is ​​concatenated with the time-coded input vector, and the concatenated matrix is ​​normalized using Z-Score normalization to obtain a first normalized output; In the self-attention focusing layer, the first normalized output is convolved, activated, and max-pooled, and the sequence representation encoder is Layer passed to Layer, after traversal, the final hidden representation of the sequence representation encoder is obtained; Get the input vector for the sequence generation decoder , the formula is as follows: ; in, represents the position code, It indicates that the output prediction is guided by the starting sequence, which is extended to a generation method, representing the length of the target blood pressure sequence before the prediction The sequence slice of is used as the starting mark of prediction. is a placeholder for the target sequence to be predicted; is the feature dimension after input representation, To predict the target blood pressure sequence length, here , represents the diastolic blood pressure, mean arterial pressure, and systolic blood pressure that need to be predicted, Yes and Perform splicing operations.

[0029] In the second convolutional attention layer, the input vector , through different weight matrices, the corresponding query matrix, key matrix, and value matrix are generated respectively, and input into the second convolutional attention layer for feature extraction to obtain the second self-attention; In the second multivariate relationship perception module, the second self-attention is first globally averaged pooled in the time dimension, and then the dependencies between input channels are captured through linear layers and activation functions to obtain the second relationship matrix; In the second normalization layer, the second relationship matrix is ​​compared with the input vector Perform splicing, and standardize the spliced ​​matrix using Z-Score standardization to obtain a second standardized output; In the third convolutional attention layer, the final hidden representation of the sequence representation encoder is used to generate the corresponding query matrix and key matrix through different weight matrices, and the second standardized output is used to generate the value matrix through the value weight matrix. Then, the three matrices are input into the third convolutional attention layer for feature extraction to obtain the third self-attention. In the third multivariate relationship perception module, the third self-attention is first globally averaged pooled in the time dimension, and then the dependencies between input channels are captured through linear layers and activation functions to obtain the third relationship matrix; In the third standardization layer, the third relationship matrix is ​​concatenated with the second standardized output, and the concatenated matrix is ​​standardized using Z-Score standardization to obtain a third standardized output; In the fully connected layer, the third standardized output is processed using the following formula to output the target blood pressure sequence; ; in, is the weight matrix of the fully connected layer, is the bias vector, is the activation function, Generate the output of the decoder for the sequence, i.e., the third normalized output; is the target blood pressure sequence.

[0030] The technical concept of the present invention is as follows: The present invention is mainly used in the field of clinical medicine. According to the high-definition intraoperative multivariate vital signs data continuously collected by invasive monitoring equipment during the patient's operation, the patient's systolic arterial pressure (SAP), diastolic pressure (DAP) and mean arterial pressure (MAP) during the operation are continuously sequenced and predicted, so as to assist the anesthesiologist in predicting whether the patient will have the risk of hypotension in the future. In response to this practical problem, the present invention proposes and designs a method for generating intraoperative blood pressure serialization with local enhancement of multivariate relationship perception. The present invention can assist clinical anesthesiologists in predicting the risk of intraoperative hypotension in patients in advance, thereby achieving timely intervention risks, avoiding intraoperative death and reducing the risk of postoperative complications, thereby effectively reducing the patient's physical damage caused by surgery and the patient's medical expenses during treatment.

[0031] The present invention utilizes the Transformer model's advantage in time series prediction, based on the "encoding-decoding" framework, and adopts the method of adding convolutional attention mechanism and multivariate relationship perception module from the inside of the model to establish a multivariate relationship perception locally enhanced intraoperative blood pressure serialization generation model (Multivariate Relationship Perception Transformer, MRPformer). The architecture of the model is as follows Figure 1 As shown in the figure, the architecture integrates the traditional encoder-decoder framework and embeds multiple feature extraction layers. It mainly includes six types of components, namely: (1) sequence representation encoder (referred to as encoder); (2) convolutional attention layer (including the first to third convolutional attention layers); (3) multivariate relationship perception module (including the first to third multivariate relationship perception modules); (4) normalization layer (including the first to third normalization layers); (5) self-attention focusing layer; (6) sequence generation decoder (referred to as decoder).

[0032] The input data of the encoder is the various vital signs data collected by the patient through the vital signs monitoring equipment during the operation (vital signs data generated over time during the operation), including but not limited to (systolic blood pressure, diastolic blood pressure, mean arterial pressure). In addition, other vital signs data can also be added to the input data, because what we need to do is multivariate, multivariate means multiple vital signs data (each variable is time series data), and the collected data is input into the model, that is, the input sequence .

[0033] It should be noted that the functions implemented by the modules in the encoder and decoder, the convolutional attention layer, the multivariate relationship perception module and the normalization layer are consistent, and the only difference is that the input and output data are different, that is, the encoder includes a convolutional attention layer, a multivariate relationship perception module, a normalization layer and a self-attention focusing layer; the decoder includes two convolutional attention layers, two multivariate relationship perception modules and two normalization layers.

[0034] exist Figure 1 In the framework shown, first, the input sequence After position encoding and global timestamp encoding, the position information and global time information of the sequence are integrated. Subsequently, the convolutional attention layer constructs convolutional attention to enhance the extraction of local features of the sequence. In order to efficiently process long sequence data, a sparse self-attention strategy is introduced in the attention calculation to extract important queries and keys in the convolutional attention mechanism, reducing the spatiotemporal complexity of the attention mechanism to At the same time, after building a multivariate relationship perception module and embedding it into the attention layer, the model's ability to mine dependencies on multivariate data is further enhanced, enabling the model to effectively adjust the weight distribution between different variable input channels and more accurately perceive the dynamic interactions between channels. The self-attention focusing layer is responsible for strengthening the key information in the sparsely processed feature map, removing redundant values, further reducing the time dimension, and allowing the model to pay more attention to the dominant features. Decoder input To predict the representation of the target blood pressure sequence, the input passes through two multi-head convolutional attention layers, one of which is a cross-attention layer that receives information from the encoder and associates the representation of the decoder's current time step with the encoder output representation, thereby capturing relevant information in the source sequence that is relevant to the target blood pressure prediction. Finally, a fully connected layer is used to output the predicted target sequence at one time. .

[0035] Sequence representation encoder: Figure 2Provides an intuitive overview that the sequence representation function consists of three independent parts, namely the input sequence, the position encoding, and the global timestamp encoding (hours, minutes, seconds).

[0036] Assume that Sequence and Different global timestamps, the feature dimension after input representation is First, a fixed position encoding is used to retain the relative position information in the sequence. The specific operation is shown in Formula 1: ; ; (1) In formula 1, , this relative position encoding gives the model the ability to capture the internal dependencies of the sequence. For each global timestamp, it is represented by an embedding with a limited vocabulary size (up to 60, i.e., seconds as the finest granularity) and a learnable embedding By embedding global time information, the similarity calculation of self-attention can access the global context, helping the model understand the positional relationship of the input data in the entire time series. In order to adjust the dimension, a one-dimensional convolution filter is used, and the input sequence is projected onto Dimension vector Finally, add the above three parts to get the input vector of time coding ,in , is a learnable embedding, is a one-dimensional convolution operation, is the position code. As shown in formula 2: ; (2) Convolutional attention layer: This paper constructs a convolutional attention mechanism such as Figure 3 As shown, in the convolutional attention layer, the multi-head attention sublayer converts the hidden state Converted into Different query matrices , key matrix , value matrix ; , The conversion method uses a kernel size of , a one-dimensional convolution with a stride of 1 (using 0 to properly fill in to maintain the dimension size), that is , ,and Use a convolution kernel size of 1 to construct, , which is equivalent to the linear transformation, is the parameter matrix, , , . This is generated and The similarity can be calculated based on local context information instead of point-by-point values, thus better capturing the local variation pattern in the blood pressure sequence. Then, a ProbSparse self-attention method is used to consider only the relationship between each key and the selected previous The interaction between queries is used to extract important , by considering only "important" queries, the computational overhead is reduced, and the key information is better captured. ProbSparse self-attention is calculated , is the ProbSparse self-attention calculation, is the softmax function, is the scaling factor, as shown in Formula 3.

[0037] ; (3) Multivariate Relationship Perception Module: The self-attention mechanism does not explicitly model the dependencies between multivariate input channels, which limits the predictive power of multivariate physiological sequence inputs in the blood pressure sequence generation task. Inspired by the excellent performance of the Squeeze-and-Excitation Block (SE_Block) in image classification tasks, this paper proposes a multivariate relationship perception module for time series data, embedding SE_Block into each convolutional attention layer as a supplement to the attention mechanism to further improve the model's ability to model the importance of input channels. Specifically, the output after the convolutional attention layer is In order to further extract the correlation of input channels, we first aggregate the information in the time dimension to obtain the average value of the input channels. , and then captures the dependencies between input channels .

[0038] In formula 4, express No. Input channels, represents the length of the input sequence, Represents the aggregation time dimension after The average value of the input channels, Perform global average pooling in the time dimension, that is, by All time steps of Sum and divide by , the channel Aggregate to a scalar value and get the channel average by aggregating the time information Finally, the next step aims to capture the dependencies between channels. and Refers to the activation function and ,in , Represents learning the weight coefficients of the input channels using two linear layers, are the hyperparameters that need to be set.

[0039] ; ; (4) The multivariate relationship perception module framework is as follows Figure 4 As shown, the first linear layer reduces the input channel dimension to the original times, which can reduce the dimension, reduce the number of model parameters and computational complexity, and compress and integrate the channel information of the input feature map, thereby extracting the correlation of the input channels. The second linear layer restores the number of channels to the original number of channels, and the model can better utilize the original channel information, ensuring feature diversity and expression ability. Through the joint action of the fully connected descending channel and the fully connected ascending channel, the model can effectively learn the weight coefficients of different input channels, where RELU and Sigmod are activation functions and Scale is the scaling factor.

[0040] Standardization layer: The Z-Score standardization adopted in the present invention has the following specific principles: first, the mean and standard deviation of each feature are calculated, and the transformation process shown in Formula 5 is performed on the observed values ​​of each feature at a time point, where: It's time point The observed value of and The mean and standard deviation of each feature are calculated respectively. This method ensures that the model will not be affected by the scale of the original data when extracting features, and it is also conducive to faster convergence of the model during training and improves prediction accuracy. In the prediction stage of the model, the standard deviation and mean during standardization are used to denormalize the output to the actual value.

[0041] ; (5) The output obtained by the multivariate relationship perception module And the input vector The spliced ​​matrix was normalized using Z-Score standardization.

[0042] Self-attention focusing layer: The present invention uses the self-attention focusing layer to extract the dominant features, aiming to extract the most influential features, thereby generating a more refined self-attention feature map. This not only reduces the demand for memory, but also improves processing efficiency. Its definition is shown in Formula 6: ; (6) Among them Operations including convolutional attention layers and multivariate relational awareness modules; It is the maximum pooling operation; It is a one-dimensional convolution operation; As the activation function, nonlinear transformation is introduced to further optimize the representation of features. Finally, one-dimensional convolution and maximum pooling operations are performed on the input feature map to focus the feature map, downsample the feature map to half of its original size to reduce the computational complexity and retain key information, and obtain the feature map from the encoder. Layer passed to The whole focusing operation process is repeated, and the number of self-attention focusing layers is gradually reduced by building multiple copies of the main stack. Finally, the final hidden representation of the encoder is obtained by connecting the outputs of all stacks. That is, the attention focusing module is to pass the output of the normalization layer through (1d convolutional layer), (activation function), (Max Pooling) A series of operations, and then from the encoder Layer passed to Layer, after repeated multiple times, the final hidden representation of the encoder is obtained .

[0043] Sequence generation decoder: In the scenario of continuous blood pressure prediction, the present invention adopts a sequence generation decoder design to solve the problems of decreasing speed and error accumulation in traditional prediction methods. The definition of is shown in Formula 7: ; (7) in Represents position encoding, here It indicates that the output prediction is guided by the starting sequence, which is extended to a generation method, representing the length of the target blood pressure sequence before the prediction The sequence slice of is used as the starting mark of prediction. is a placeholder (set to 0) for the target sequence to be predicted, is the feature dimension after input representation, To predict the target blood pressure sequence length, here , represents the diastolic blood pressure, mean arterial pressure, and systolic blood pressure that need to be predicted, Yes and Perform splicing operations. Figure 5 As shown, the input Finally, the sequence generation decoder directly generates the target blood pressure sequence through the fully connected layer, avoiding the error accumulation caused by the step-by-step reasoning of the traditional autoregressive decoder. and They are all inputs to the decoder, but the input positions are different. It is input from the decoder head and passes through all modules in the decoder, and The input starts from the second convolutional attention layer (cross attention layer) in the decoder.

[0044] For the task of intraoperative blood pressure prediction, the performance of different models with blood pressure sequence prediction lengths of 5, 10, and 15 minutes is compared (see Table 1). The prediction results of diastolic pressure, mean arterial pressure, and systolic pressure are used for comprehensive evaluation, which helps to fully understand the performance of each model in different prediction length ranges. The present invention uses the mean absolute error (MAE), root mean square error (RMSE), and correlation coefficient (CORR) commonly used in time series prediction as evaluation indicators to evaluate the prediction effect of the MRPformer model on the task of intraoperative blood pressure prediction. The comparison model uses 7 mainstream time series prediction models based on deep learning, namely: LSTNet, Dlinear, Autoformer, SCINet, Reformer, PatchTST, and Informer. All experiments use the 5-fold cross-validation method to ensure the reliability of the results, and take the average value as the final experimental result. As shown in Table 1, our proposed method shows excellent performance in all three indicators, surpassing other comparison models. In addition, compared with Informer, the average MAE of the three predicted blood pressures was reduced by 7.18%, 11.3%, and 11.44%, respectively, and the error increment was smaller as the prediction length increased, indicating that the model has better long-term prediction performance.

[0045]

[0046] In order to study the impact of the multivariate relationship perception module and convolutional attention mechanism on intraoperative blood pressure prediction, we designed an ablation experiment. For intuitive display, the average error index of diastolic pressure, mean arterial pressure, and systolic pressure is used as the result. MRPformer is the model of the present invention, Informer+c represents the replacement of the original canonical attention with convolutional attention, Informer+m represents the addition of a multivariate relationship perception module, and Informer is the unimproved model. The ablation experiment results reveal the contribution of each component of the model and its effectiveness in the task of intraoperative blood pressure prediction, as shown in Table 2.

[0047]

[0048] After replacing the canonical attention with convolutional attention, MAE decreased by 2.42%, which shows that introducing the locality of convolutional attention to extract attention to focus on the local similarity of the sequence can improve the prediction performance of the model to a certain extent. In addition, by adding only the multivariate relationship perception module, MAE decreased by 6.98%, indicating that by modeling the importance of the input channel dimension, the model's perception of the deep interactive relationship of the multivariate physiological sequence can be enhanced, which can effectively enhance the model's prediction ability. The above experimental results all demonstrate the superiority of our proposed method for generating intraoperative blood pressure serialization with multivariate relationship perception and local enhancement.

[0049] The present invention proposes to utilize the advantages of the Transformer-based model in time series prediction, adopt a locally enhanced convolutional attention mechanism to replace the canonical attention and a multivariate relationship perception module embedding model, and establish a multivariate relationship perception locally enhanced intraoperative blood pressure serialization generation method. This method uses the locally enhanced convolutional attention mechanism to capture long-sequence dependencies while introducing local time correlation and trend, thereby achieving a more accurate prediction of intraoperative blood pressure changes. And through the multivariate relationship perception module, the model's ability to model the relationship between channels of time series multivariate input is enhanced, so that the model can effectively learn the weights between different variable input channels and better perceive the correlation between variable input channels.

[0050] On the one hand, from a clinical perspective, the method proposed in the present invention adopts a time series prediction model, defines the continuous prediction task of intraoperative blood pressure as a time series generation task, and predicts the continuous change trend of the patient's blood pressure series in the future. This can intuitively present the development process of intraoperative hypotension, allowing anesthesiologists to observe the blood pressure change trend of patients at various stages of the operation in real time, providing sufficient evidence to assist anesthesiologists in making decisions, and plays a significant role in reducing the burden on anesthesiologists and the surgical risks of patients. On the other hand, from a technical perspective, the method proposed in the present invention uses a locally enhanced convolutional attention mechanism and a multivariate relationship perception module to enhance the model's perception of the local correlation and trend of long time series and the correlation between multivariate vital sign input channels, greatly improving the performance of the model in the task of continuous blood pressure prediction.

[0051] like Figure 6 , Figure 7 As shown, predicting the value and change trend of blood pressure series in the future has more important practical application value.

[0052] The key points of the invention are: (1) To address the problem that traditional binary classification models based on statistical machine learning algorithms cannot achieve real-time blood pressure prediction, we designed a method for serializing intraoperative blood pressure. We defined the prediction of intraoperative blood pressure as a time series generation task, so that the model can predict the value and change trend of the blood pressure series over a period of time in the future, eliminating the impact of inconsistent definitions of hypotension threshold standards, and significantly improving the effectiveness of assisting anesthesiologists in predicting the risk of intraoperative hypotension.

[0053] (2) The standard Transformer attention mechanism has shortcomings when processing physiological time series data, that is, it only captures global information through point-to-point operations and fails to fully consider the importance of local features. At the same time, the Transformer's self-attention mechanism does not explicitly model the dependencies between multivariate input channels, which limits its ability in multivariate blood pressure prediction tasks. To address these problems, we designed a multivariate relation-aware local enhanced Transformer model (MRPformer), which introduced a convolutional attention mechanism for local temporal correlation and trend, and also introduced a multivariate relation-aware module that can enhance the model's ability to model the relationship between input channels.

[0054] The above-described embodiments are used to illustrate the present invention, not to limit the present invention, so changes in the exemplified values ​​or replacement of equivalent elements should still fall within the scope of the present invention.

[0055] From the above detailed description, it can be understood by those skilled in the art that the present invention can indeed achieve the aforementioned objectives and is in compliance with the provisions of the Patent Law.

[0056] Although the preferred embodiments of the present invention have been described, those skilled in the art may make additional changes and modifications to these embodiments once they are aware of the basic creative concepts. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present invention. The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. It should be pointed out that any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

[0057] It should be noted that the above description of the relevant process is only for example and explanation, and does not limit the scope of application of this specification. For those skilled in the art, various modifications and changes can be made to the process under the guidance of this specification. However, these modifications and changes are still within the scope of this specification.

[0058] The basic concepts have been described above. Obviously, for those of ordinary skill in the art who have read this application, the above invention disclosure is only for example and does not constitute a limitation of this application. Although not explicitly stated here, those of ordinary skill in the art may make various modifications, improvements and amendments to this application. Such modifications, improvements and amendments are suggested in this application, so such modifications, improvements and amendments still belong to the spirit and scope of the exemplary embodiments of this application.

[0059] At the same time, the present application uses specific words to describe the embodiments of the present application. For example, "one embodiment", "an embodiment", and / or "some embodiments" refer to a certain feature, structure or characteristic related to at least one embodiment of the present application. Therefore, it should be emphasized and noted that "one embodiment" or "an embodiment" or "an alternative embodiment" mentioned twice or more in different positions in this specification does not necessarily refer to the same embodiment. In addition, some features, structures or characteristics in one or more embodiments of the present application can be appropriately combined.

[0060] In addition, it will be understood by those skilled in the art that various aspects of the present application may be illustrated and described by a number of patentable categories or situations, including any new and useful process, machine, product or combination of substances, or any new and useful improvement thereof. Therefore, various aspects of the present application may be implemented entirely by hardware, entirely by software (including firmware, resident software, microcode, etc.), or by a combination of hardware and software. The above hardware or software may be referred to as a "unit", "module" or "system". In addition, various aspects of the present application may take the form of a computer program product embodied in one or more computer-readable media, wherein computer-readable program code is contained therein.

[0061] The computer program code required for the operation of each part of the present application can be written in any one or more programming languages, including object-oriented programming languages ​​such as Java, Scala, Smalltalk, Eiffel, JADE, Emerald, C++, C#, VB.NET, Python, etc., conventional procedural programming languages ​​such as C programming language, VisualBasic, Fortran2103, Perl, COBOL2102, PHP, ABAP, dynamic programming languages ​​such as Python, Ruby and Groovy, or other programming languages. The program code can be run entirely on the user's computer, or run on the user's computer as an independent software package, or run partially on the user's computer and partially on the remote computer, or run entirely on the remote computer or server. In the latter case, the remote computer can be connected to the user's computer through any network form, such as a local area network (LAN) or a wide area network (WAN), or connected to an external computer (e.g., via the Internet), or in a cloud computing environment, or used as a service such as software as a service (SaaS).

[0062] In addition, unless explicitly stated in the claims, the order of the processing elements and sequences described in this application, the use of alphanumeric characters, or the use of other names are not intended to limit the order of the processes and methods of this application. Although the above disclosure discusses some invention embodiments that are currently considered useful through various examples, it should be understood that such details are only for illustrative purposes, and the attached claims are not limited to the disclosed embodiments. On the contrary, the claims are intended to cover all modifications and equivalent combinations that are consistent with the essence and scope of the embodiments of this application. For example, although the implementation of the various components described above can be embodied in a hardware device, it can also be implemented as a pure software solution, for example, installation on an existing server or mobile device.

[0063] Similarly, it should be noted that in order to simplify the description disclosed in this application and thus help understand one or more embodiments of the invention, in the above description of the embodiments of the application, multiple features are sometimes combined into one embodiment, figure or description thereof. However, this method of the application should not be interpreted as reflecting the intention that the claimed object requires more features than those explicitly stated in each claim. On the contrary, the subject of the invention should have fewer features than the above single embodiment.

Claims

1. A method for generating intraoperative blood pressure serialization with multivariate relationship-aware local enhancement, characterized in that: A multivariate relationship-aware locally enhanced intraoperative blood pressure serialization generation model is established, which includes a sequence representation encoder, a sequence generation decoder and a fully connected layer. The sequence representation encoder includes a first convolutional attention layer, a first multivariate relationship perception module, a first normalization layer and a self-attention focusing layer. The sequence generation decoder includes a second convolutional attention layer, a second multivariate relationship perception module, a second normalization layer, a third convolutional attention layer, a third multivariate relationship perception module and a third normalization layer. The specific steps are as follows: Collect various vital sign data collected by the patient through the vital sign monitoring equipment during the operation to form an input sequence; Add the input sequence, position code and global timestamp code to get the time-coded input vector; In the first convolutional attention layer, the time-encoded input vector is used to generate the corresponding query matrix, key matrix, and value matrix through different weight matrices, and then input into the first convolutional attention layer for feature extraction to obtain the first self-attention. In the first multivariate relationship perception module, the first self-attention is firstly globally averaged pooled in the time dimension, and then the dependency between input channels is captured through a linear layer and an activation function to obtain the first relationship matrix; In the first normalization layer, the first relationship matrix is ​​concatenated with the time-coded input vector, and the concatenated matrix is ​​normalized using Z-Score normalization to obtain a first normalized output; In the self-attention focusing layer, the first normalized output is convolved, activated, and max-pooled, and the sequence representation encoder is Layer passed to Layer, after traversal, the final hidden representation of the sequence representation encoder is obtained; Get the input vector for the sequence generation decoder , the formula is as follows: ； in, represents the position code, It indicates that the output prediction is guided by the starting sequence, which is extended to a generation method, representing the length of the target blood pressure sequence before the prediction The sequence slice of is used as the starting mark of prediction. is a placeholder for the target sequence to be predicted; is the feature dimension after input representation, To predict the target blood pressure sequence length, here , represents the diastolic blood pressure, mean arterial pressure, and systolic blood pressure that need to be predicted, Yes and Perform splicing operations; In the second convolutional attention layer, the input vector , through different weight matrices, the corresponding query matrix, key matrix, and value matrix are generated respectively, and input into the second convolutional attention layer for feature extraction to obtain the second self-attention; In the second multivariate relationship perception module, the second self-attention is first globally averaged pooled in the time dimension, and then the dependencies between input channels are captured through linear layers and activation functions to obtain the second relationship matrix; In the second normalization layer, the second relationship matrix is ​​compared with the input vector Perform splicing, and standardize the spliced ​​matrix using Z-Score standardization to obtain a second standardized output; In the third convolutional attention layer, the final hidden representation of the sequence representation encoder is used to generate the corresponding query matrix and key matrix through different weight matrices, and the second standardized output is used to generate the value matrix through the value weight matrix. Then, the three matrices are input into the third convolutional attention layer for feature extraction to obtain the third self-attention. In the third multivariate relationship perception module, the third self-attention is first globally averaged pooled in the time dimension, and then the dependencies between input channels are captured through linear layers and activation functions to obtain the third relationship matrix; In the third standardization layer, the third relationship matrix is ​​concatenated with the second standardized output, and the concatenated matrix is ​​standardized using Z-Score standardization to obtain a third standardized output; In the fully connected layer, the third standardized output is processed using the following formula to output the target blood pressure sequence; ； in, is the weight matrix of the fully connected layer, is the bias vector, is the activation function, Generate the output of the decoder for the sequence, i.e., the third normalized output; is the target blood pressure sequence.

2. The method for generating intraoperative blood pressure serialization with multivariate relationship-aware local enhancement according to claim 1, characterized in that: The sequence represents the encoder through the input sequence , Positional encoding and global timestamp encoding To realize the sequence representation function, Sequence and The feature dimensions of different global timestamps after input representation are ; Use fixed position encoding to retain the relative position information in the sequence, as shown in the following formula: ； ； in, , in order to adjust the dimension, a one-dimensional convolution filter is used and the input sequence is projected onto Dimension vector , and finally get the time-coded input vector , as shown below: ； in , is a learnable embedding, is a one-dimensional convolution operation, Encode the position.

3. The method for generating intraoperative blood pressure serialization with multivariate relationship-aware local enhancement according to claim 1, characterized in that: In the first, second, and third convolutional attention layers, the multi-head attention sublayer converts the hidden state Converted into Different query matrices , key matrix , value matrix ; , The conversion method uses a kernel size of , a one-dimensional convolution with a stride of 1, that is, , ,and Use a convolution kernel size of 1 to construct, , which is equivalent to the linear transformation, is the parameter matrix, , , ; Calculate ProbSparse self-attention , is the ProbSparse self-attention calculation, is the softmax function, is the scaling factor, as shown below; 。 4. The method for generating intraoperative blood pressure serialization with multivariate relationship-aware local enhancement according to claim 1, characterized in that: In the first multivariate relationship perception module, the second multivariate relationship perception module, and the third multivariate relationship perception module, the squeeze excitation block is embedded into each convolutional attention layer as a supplement to the attention mechanism. The output after the convolutional attention layer is In order to further extract the correlation of input channels, we first aggregate the information in the time dimension to obtain the average value of the input channels. , and then captures the dependencies between input channels , as shown in the following formula; ； ； in, express No. Input channels, represents the length of the input sequence, Represents the aggregation time dimension after The average value of the input channels, and Refers to the activation function and ,in , Represents learning the weight coefficients of the input channels using two linear layers, is the hyperparameter that needs to be set; the input channel dimension is reduced to the original through the first linear layer times to achieve dimensionality simplification, reduce the number of model parameters and computational complexity, and compress and integrate the channel information of the input feature map to extract the correlation of the input channels; the number of channels is restored to the original number of channels through the second linear layer, which can better utilize the original channel information, ensure feature diversity and expression ability, and effectively learn the weight coefficients of different input channels through the joint action of the fully connected descending channel and the fully connected ascending channel.

5. The method for generating intraoperative blood pressure serialization with multivariate relationship-aware local enhancement according to claim 1, characterized in that: In the first, second, and third standardization layers, the mean and standard deviation of each feature are first calculated, and the following transformation process is performed on the observed values ​​of each feature at each time point: ； in, It's time point The observed value of and The mean and standard deviation of each feature are calculated respectively; in the prediction stage of the model, the output is denormalized to the actual value using the standard deviation and mean during standardization.

6. The method for generating intraoperative blood pressure serialization with multivariate relationship-aware local enhancement according to claim 1, characterized in that: In the self-attention focusing layer, the formula for extracting the dominant features is as follows: ； Among them Operations including convolutional attention layers and multivariate relational awareness modules; It is the maximum pooling operation; It is a one-dimensional convolution operation; As the activation function, nonlinear transformation is introduced to further optimize the representation of features. Finally, one-dimensional convolution and maximum pooling operations are performed on the input feature map to focus the feature map, downsample the feature map to half of its original size to reduce the computational complexity and retain key information, and represent the encoder from the sequence. Layer passed to The whole focusing operation process is repeated, and the number of self-attention focusing layers is gradually reduced by building multiple copies of the main stack. Finally, the final hidden representation of the sequence representation encoder is obtained by connecting the outputs of all stacks.

Images