Intraoperative invasive blood pressure continuous prediction device and system and storage medium

Abstract

The invention relates to the technical field of clinical medical treatment, and discloses an intraoperative invasive blood pressure continuous prediction device and system and a storage medium, and the device comprises an embedded layer which is configured to encode positions, timestamps and feature dimension information in an input sequence to obtain an input vector, a statistical feature extractor, and an output vector; the system comprises an intra-operative blood pressure sequence calculation module configured to calculate statistical characteristics of an intra-operative blood pressure sequence, a multi-layer perceptron module configured to learn a distribution adaptation factor in response to the input intra-operative blood pressure sequence and the statistical characteristics thereof, and a distribution adaptation attention layer configured to calculate distribution adaptation attention based on the distribution adaptation factor; the encoder and the decoder achieve precise and efficient blood pressure sequence continuous prediction through the synergistic effect. The anesthetist can be assisted in managing the blood pressure of the patient in the operation period and coping with risks possibly occurring in the operation process and related to abnormal blood pressure changes.

Description

Technical Field

[0001] The present application relates to the field of clinical medical technology, and in particular to a device and system for continuous prediction of intraoperative invasive blood pressure and a storage medium. Background Art

[0002] Intraoperative hypotension (IOH) is closely associated with 30-day postoperative mortality and postoperative complications. Postoperative complications rank as the third leading cause of death worldwide and are therefore not to be ignored. Therefore, accurately predicting intraoperative blood pressure trends can help clinical anesthesiologists rationally plan clinical resources and develop emergency plans, thereby effectively reducing the risk of postoperative complications and mortality, which has significant positive implications.

[0003] Currently, intraoperative hypotension is primarily predicted by leveraging preoperative medical records and real-time intraoperative vital sign data. However, these methods generally fail to provide real-time predictions. To achieve real-time tracking and prediction of intraoperative hypotension events, some research has turned to analyzing arterial waveform signals. By analyzing continuous monitoring data, traditional machine learning models are constructed to predict the risk of intraoperative hypotension in real time. These methods primarily focus on predicting intraoperative hypotension events. Based on retrospective intraoperative multivariate vital sign monitoring data, prediction is achieved by analyzing the association between hypotension risk events at a given moment and historical multivariate vital sign monitoring data within a limited window of 5, 10, and 15 minutes prior. This problem is defined as a binary classification problem. In reality, it is unreasonable to crudely define the occurrence or non-occurrence of intraoperative hypotension as an instantaneous event. Firstly, since there is currently no consistent clinical definition of intraoperative hypotension, these methods, based on the universal definition of positive and negative samples, can only predict specific intraoperative hypotension events, resulting in insufficient generalizability. Secondly, relying solely on classification results ignores the clinical significance of the hypotension development process, increasing the burden on anesthesiologists during intraoperative decision-making.

[0004] In addition, with the rise of deep learning methods, such as Transformer and its variants, they have also been gradually introduced into the prediction of intraoperative hypotension, such as Figure 1 As shown in the figure, the traditional intraoperative blood pressure prediction model defines the prediction of intraoperative blood pressure as a time series prediction task, which can realize real-time and continuous prediction of the blood pressure change sequence within 5 minutes, 10 minutes and 15 minutes in the future starting from the prediction point during the operation. It is the most advanced method at present.

[0005] However, from a clinical perspective, when clinical anesthesiologists predict the risk of intraoperative hypotension for patients during surgery, they usually combine the patient's preoperative basic examination data, the doctor's diagnostic records, and real-time vital sign monitoring data during the operation, and use their own clinical experience to judge whether to actively administer fluids, vasopressors, or inotropic drugs to deal with the possible risk of intraoperative hypotension. The main drawbacks of this prediction method are the possible lag in prediction and the inability to guarantee the accuracy of prediction due to the uneven clinical experience level of anesthesiologists. This means that during surgery, the patient may suffer from the adverse effects or even injury of intraoperative hypotension before receiving appropriate intervention.

[0006] In addition, from a technical perspective, due to factors such as physiological differences between individual patients, diversity of surgical conditions, and inconsistency of monitoring equipment, intraoperative blood pressure prediction tasks often face severe challenges of data distribution deviation. Traditional machine learning models and existing time series prediction models based on the "encoding-decoding" framework cannot effectively avoid the impact of these problems on prediction performance when applied to intraoperative blood pressure prediction tasks. This may lead to a decline in model performance when faced with inconsistently distributed samples, limiting its universality and reliability in actual clinical environments. Summary of the Invention

[0007] To address the above-mentioned issues, the present application provides a device and system for continuous intraoperative invasive blood pressure prediction, as well as a storage medium. Based on a distribution-adaptive attention strategy, a continuous intraoperative invasive blood pressure prediction scheme based on distribution-adaptive attention is established. This application aims to mitigate distribution differences by incorporating statistical information into the model's internal design. By learning the statistical feature information of the input window, the model's attention weight is automatically adjusted, enhancing the ability to learn the intrinsic distribution patterns of non-stationary sequences, thereby reducing performance fluctuations caused by distribution differences.

[0008] On the one hand, from a clinical perspective, this application uses a time series prediction model to formally simulate the comprehensive judgment process of clinical anesthesiologists on the risk of intraoperative hypotension. Through the continuous prediction of blood pressure sequences, the development process of intraoperative hypotension is demonstrated, and through the distribution adaptation attention mechanism, the medical relationship between the various parts of the intraoperative blood pressure sequence is modeled in the model, providing sufficient evidence to assist anesthesiologists in making decisions. On the other hand, from a technical perspective, this application uses a time series prediction model to intuitively present the changing trend of the patient's intraoperative blood pressure sequence over time, and through the distribution adaptation attention mechanism, it greatly reduces the impact of the distribution differences of the intraoperative blood pressure sequence on the model performance, while improving the robustness of the model to dynamically changing environments.

[0009] In order to achieve the above objectives, the technical solutions adopted in this application are as follows: In a first aspect, the present application provides a device for continuous prediction of intraoperative invasive blood pressure, the device comprising: an embedding layer configured to encode position, timestamp, and feature dimension information in an input sequence to obtain an input vector; wherein the input sequence includes a first input sequence and a second input sequence, and the input vector includes a first input vector and a second input vector, wherein the first input vector and the second input vector are obtained by encoding the first input sequence and the second input sequence, respectively; A statistical feature extractor is configured to calculate the statistical features of the intraoperative blood pressure sequence; wherein the statistical features of the blood pressure sequence include The mean of the input time window and standard deviation ; a multi-layer perceptron module, wherein the signal output terminal of the statistical feature extractor is connected to the signal input terminal of the multi-layer perceptron module, and the multi-layer perceptron module is configured to learn a distribution adaptation factor in response to an input intraoperative blood pressure sequence and its statistical features; A distribution adaptive attention layer, wherein the signal output end of the multilayer perceptron module is connected to the signal input end of the distribution adaptive attention layer, and the distribution adaptive attention layer is configured to guide the calculation of the distribution adaptive attention based on the distribution adaptation factor to obtain the distribution adaptive attention; an encoder, wherein the signal output ends of the distribution adaptive attention layer and the embedding layer are connected to the signal input end of the encoder, and the encoder is configured to use the first input vector as an input sequence of the encoder, capture long-term dependencies in the input sequence based on the distribution adaptive attention, and adopt multi-head attention to learn multiple feature expressions in parallel as the output of the encoder; A decoder, wherein the signal output ends of the encoder and the embedding layer are connected to the signal input end of the decoder, and the decoder is configured to rely on an autoregressive mechanism when generating predictions of future time steps, using the previous prediction as input at each step to construct a continuous output sequence. At the same time, a multi-head attention mechanism is used to not only process the second input vector, but also fuse the output of the encoder with the information of the current step, and output the prediction result after a series of feedforward networks and linear transformations.

[0010] In a second aspect, the present application provides a system for continuous prediction of intraoperative invasive blood pressure, the system comprising: a memory for storing a computer program; and a processor for executing the computer program to implement the following method: Encoding the position, timestamp, and feature dimension information in the input sequence to obtain an input vector; wherein the input sequence includes a first input sequence and a second input sequence, and the input vector includes a first input vector and a second input vector, and the first input vector and the second input vector are obtained by encoding the first input sequence and the second input sequence, respectively; Calculate the statistical characteristics of the blood pressure sequence during the operation; wherein the statistical characteristics of the blood pressure sequence include The mean of the input time window and standard deviation ; Based on the input intraoperative blood pressure sequence and its statistical characteristics, a distribution adaptation factor is learned; Guiding calculation of distribution adaptive attention based on the distribution adaptation factor to obtain distribution adaptive attention; constructing an encoder and a decoder, wherein the encoder is configured to take the first input vector as an input sequence of the encoder, capture long-term dependencies in the input sequence based on the distribution adaptation attention, and adopt multi-head attention to learn multiple feature expressions in parallel as the output of the encoder; The decoder is configured to rely on an autoregressive mechanism when generating predictions for future time steps. Each step uses the previous prediction as input to construct a continuous output sequence. At the same time, a multi-head attention mechanism is used to not only process the second input vector, but also fuse the output of the encoder with the information of the current step, and output the prediction result after a series of feedforward networks and linear transformations.

[0011] In a third aspect, the present application provides a non-transitory computer-readable storage medium storing instructions. When the instructions are executed by a processor, the following steps are performed: Encoding the position, timestamp, and feature dimension information in the input sequence to obtain an input vector; wherein the input sequence includes a first input sequence and a second input sequence, and the input vector includes a first input vector and a second input vector, and the first input vector and the second input vector are obtained by encoding the first input sequence and the second input sequence, respectively; Calculate the statistical characteristics of the blood pressure sequence during the operation; wherein the statistical characteristics of the blood pressure sequence include The mean of the input time window and standard deviation ; Based on the input intraoperative blood pressure sequence and its statistical characteristics, a distribution adaptation factor is learned; Guiding calculation of distribution adaptive attention based on the distribution adaptive factor to obtain distribution adaptive attention; constructing an encoder and a decoder, wherein the encoder is configured to take the first input vector as an input sequence of the encoder, capture long-term dependencies in the input sequence based on the distribution adaptation attention, and adopt multi-head attention to learn multiple feature expressions in parallel as the output of the encoder; The decoder is configured to rely on an autoregressive mechanism when generating predictions for future time steps. Each step uses the previous prediction as input to construct a continuous output sequence. At the same time, a multi-head attention mechanism is used to not only process the second input vector, but also fuse the output of the encoder with the information of the current step, and output the prediction result after a series of feedforward networks and linear transformations.

[0012] This application has at least the following beneficial effects: (1) In response to the problem that traditional intraoperative blood pressure prediction models cannot achieve real-time continuous blood pressure prediction, this application defines the prediction of intraoperative blood pressure as a time series prediction task, so that the model can predict the value and change trend of the blood pressure series within 5 minutes, 10 minutes and 15 minutes from the prediction point, 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.

[0013] (2) During surgery, rapid changes in blood pressure are often a key indicator of changes in the patient's physiological state. Performing only global normalization will blur the characteristics of these key events, resulting in an overly uniform distribution of attention weights in the attention layer, reducing the model's sensitivity to key events, and ultimately causing the model to produce overly smooth outputs that lack event responsiveness. To address the above issues, this application designs a blood pressure continuous prediction device, system, and storage medium based on distributed adaptive attention. While utilizing the stability of globally normalized data, it can also adapt to the potential non-stationary characteristics within each input window, thereby capturing the unique dynamic characteristics of blood pressure data and improving the model's predictive ability. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 The figure shows a structural diagram of a traditional intraoperative blood pressure prediction model according to the prior art.

[0015] Figure 2 A structural diagram of a device for continuous prediction of intraoperative invasive blood pressure according to an embodiment of the present application is shown.

[0016] Figure 3 A data processing flow chart of the embedding layer according to an embodiment of the present application is shown.

[0017] Figure 4 A data processing flow chart of a statistical feature extractor according to an embodiment of the present application is shown. DETAILED DESCRIPTION

[0018] The following describes the embodiments of the present application through specific examples. Those skilled in the art can easily understand the other advantages and effects of the present application from the content disclosed in this specification. The present application can also be implemented or applied through other different specific embodiments. The details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present application. It should be noted that the following embodiments and features in the embodiments can be combined with each other unless they conflict.

[0019] The specific implementation of the present application is further described in detail below with reference to the accompanying drawings and examples.

[0020] This embodiment of the present application proposes a device for continuous intraoperative invasive blood pressure prediction. This device continuously predicts a patient's systolic arterial pressure (SAP), diastolic arterial pressure (DAP), and mean arterial pressure (MAP) during surgery, assisting anesthesiologists in managing blood pressure during surgery and addressing risks associated with abnormal blood pressure changes, most commonly intraoperative hypotension.

[0021] See also Figure 2 , which is a structural diagram of the device for continuous intraoperative invasive blood pressure prediction provided by an embodiment of the present application. The device includes the following components: an embedding layer 100, a statistical feature extractor 200, a multi-layer perceptron module 300, a distribution-adaptive attention layer 400, an encoder 500, and a decoder 600.

[0022] Figure 2 middle, and represent the input sequence of the encoder and the input sequence of the decoder respectively, and Represents the input sequence The statistical characteristics of and It means that the multi-layer perceptron module is The distribution adaptation factor learned in is a positive scalar factor is the displacement vector. 、 、 , representing the input query, key, and value respectively. Figure 2The arrows in the figure show the direction of signal flow. The input sequence of the encoder and the input sequence of the decoder are vital sign monitoring data, such as intraoperative blood pressure sequence, which are collected by existing medical equipment and can be transmitted to the embedding layer 100 and the statistical feature extractor 200 respectively. The output of the embedding layer 100, that is, the first input vector and the second input vector are input to the encoder 500 and the decoder 600 respectively. The output of the statistical feature extractor 200 is used as the input of the multi-layer perceptron module 300, and the output of the multi-layer perceptron module 300 is used as the input of the distribution adaptation attention layer 400. The output of the distribution adaptation attention layer 400 is used as the input of the encoder 500. The input of the decoder 600 includes the output of the encoder 500 and the second input vector output by the embedding layer 100. Finally, the output of the decoder 600 is the predicted blood pressure sequence. .

[0023] The functions of each component and their specific definitions are described in detail below.

[0024] During surgery, the patient's vital signs monitoring data is usually time series data. Therefore, the embedding layer 100 in the intraoperative invasive blood pressure continuous prediction device has the function of encoding the position, timestamp, and feature dimension information in the input sequence. Figure 3 , which is a data processing flow chart of the embedding layer provided in the embodiment of the present application. First, a fixed position encoding is used to retain the relative position information in the sequence to obtain the position vector , then each global timestamp is represented by an embedding with a finite vocabulary size , helps the model understand the position relationship of the input data in the entire time series. In order to adjust the feature dimension, a one-dimensional convolution filter is used to convert the input sequence Projection to Vector of dimensions Finally, add the above three parts to get the input vector of time coding After the input sequence is encoded by the embedding layer, the model can not only utilize local sequence information, but also understand and deal with long-term temporal dependencies.

[0025] See also Figure 4 , which is a data processing flow chart of the statistical feature extractor provided in the embodiment of the present application. For the entire intraoperative invasive blood pressure continuous prediction device, compared with the original non-stationary attention architecture, the distribution adaptation architecture removes the input standardization and output denormalization modules, and directly uses the globally standardized input, so that it only calculates and adjusts the mean and standard deviation based on the data in each time series segment. The role of the statistical feature extractor is to calculate its statistical features. and , both of which represent The mean and standard deviation of the input time window at each moment provide essential parameter support for the calculation of distribution-adaptive attention.

[0026] Since each patient has a different physiological condition, their physiological sequence also has the problem of inconsistent sample distribution. At the same time, since the patient's physiological state will also change significantly during the operation, these changes may be caused by a variety of factors, such as the use of drugs, differences in surgical operations, and the patient's own physiological reactions. These factors make the blood pressure sequence non-stationary. These problems make it difficult for the blood pressure prediction model to capture and adapt to the individual variability and dynamic changes of non-stationarity in the blood pressure sequence, thereby affecting the generalization ability and accuracy of the model. Through the multi-layer perceptron module 300, distribution adaptation factors can be learned from the input time series. They can be used to control the standard deviation of local data to adapt to the volatility of different time periods and calibrate the changes in local mean to eliminate data drift. The specific calculation process is shown in formula (1): ; (1) Where, represents a positive scalar factor, represents the intraoperative blood pressure sequence, Represents a displacement vector.

[0027] Traditional blood pressure prediction models may find it difficult to capture and adapt to the dynamic changes of individual variability and non-stationarity in blood pressure sequences, thus affecting the generalization ability and accuracy of the model. A distribution adaptive attention mechanism is introduced in the continuous prediction of intraoperative invasive blood pressure. The core idea is to understand the uniqueness of each block by learning the local statistical features of each sequence block. These local statistical features are then used to guide attention allocation, so that the model can dynamically adjust the weights according to the actual distribution characteristics of each sequence block. If the standard deviation is large, it means that the data points in the sequence block are more dispersed. It is hoped that the model can pay more attention to those data points that are more different from the average situation, by increasing the positive scalar factor of the attention weight. When the mean changes, it means that the overall trend of the sequence has changed. By adjusting a displacement vector , so that the model takes this trend change into account when calculating the attention weight. This enables the model to not only utilize the stability of the overall standardized data, but also adapt to the potential non-stationary characteristics within each input window, enhancing the ability to learn the inherent distribution pattern of the non-stationary sequence, thereby improving the model's prediction ability for dynamically changing physiological sequence data. The distribution adaptation attention mechanism is implemented by the distribution adaptation attention layer 400, and its specific calculation process is shown in formula (2): (2) Where, Represents the input query, key, and value respectively; Represents the vector dimension.

[0028] The encoder 500 and decoder 600 work together to achieve accurate and efficient continuous prediction of blood pressure sequences. The encoder 500 is designed to capture long-term dependencies in the input sequence. It uses a built-in distribution-adaptive attention mechanism to capture the intrinsic distribution and temporal dependencies of non-stationary time series, and adopts multi-head attention to learn multiple feature expressions in parallel. The decoder 600 relies on an autoregressive mechanism when generating predictions for future time steps. Each step uses the previous prediction as input to construct a continuous output sequence. At the same time, the decoder 600 also uses a multi-head attention mechanism to not only process the sequence it generates itself, but also fuse the output of the encoder 500 with the information of the current step to enhance prediction accuracy. Finally, after a series of feedforward networks and linear transformations, the decoder 600 outputs the prediction result. This design enables the device to efficiently process long time series and demonstrate excellent performance in intraoperative blood pressure prediction tasks.

[0029] In summary, the intraoperative invasive blood pressure continuous prediction device can predict the patient's blood pressure change trend in the future, prompting the medical team to implement more proactive and timely intervention measures to maintain the patient's blood pressure within a safe range, such as timely adjustment of anesthesia dose, optimization of fluid infusion strategy and rational use of vasoactive drugs, so as to reduce the occurrence of death and postoperative complications caused by abnormal blood pressure changes during surgery. This will significantly reduce the patient's physical damage caused by surgery and medical expenses during the treatment process, and has important practical application value.

[0030] The embodiment of the present application conducted the following comparative experiments on the task of intraoperative blood pressure prediction. This embodiment selected Informer, Patch_TST and CACformer as benchmark models for comparison, and used the Mean Absolute Error (MAE), Root Mean Squared Error (RMSE) and Correlation Coefficient (CORR) commonly used in time series prediction as evaluation indicators to evaluate the prediction effect of the device proposed in the embodiment of the present application on the task of intraoperative blood pressure prediction. The average prediction error indicators of diastolic pressure, mean arterial pressure and systolic pressure were used as the final result. The following symbols were used in the comparative experiment to indicate different model configurations: - represents the original performance data of the benchmark model; -Z is used to refer to the method of directly using the original data for training without global normalization; +Re represents the use of Revin for adaptive normalization method in the input and output stages; +NS indicates that the model uses the original non-stationary attention architecture; +DA represents that the model integrates the distribution adaptive attention architecture constructed by the present invention. The results of the comparative experiment are shown in Table 1: Table 1 Comparative experimental results of distribution adaptation attention architecture

[0031] The experimental results in Table 1 demonstrate that the performance of Patch_TST, Informer, and CACformer, all significantly improved when incorporating the distribution adaptation (+DA) architecture. In particular, DA-CACformer achieved improvements across all evaluation metrics compared to the original baseline. Specifically, for prediction lengths of 5, 10, and 15 minutes, the MAE decreased by 1.53%, 3.17%, and 6.92%, respectively, and the RMSE decreased by 4.07%, 4.60%, and 4.47%, respectively. The significant reduction in RMSE, a metric more sensitive to outliers, compared to MAE demonstrates that the distribution adaptation architecture is capable of identifying and handling non-stationary events, particularly when dealing with data subject to sudden and unconventional fluctuations, demonstrating enhanced robustness and predictive power. Experimental results show that the performance of all three baseline models significantly degrades on non-normalized data (-Z), further highlighting the importance of data normalization for model performance in blood pressure prediction tasks. In the +NS group of experiments, the results clearly demonstrate that our distribution adaptation attention architecture outperforms the non-stationary Transformer architecture. In the +Re experiment, although the reversible adaptive normalization module Revin was used at the input and output, the effect was not as good as the above two, which further shows that adaptive normalization outside the model has limited effect on improving the model's ability to identify and process non-stationary information.

[0032] To verify the adaptability factor and To investigate the impact on prediction results, this embodiment designed an ablation experiment. As shown in Table 2, the baseline model used in the ablation experiment is CACformer, where -- represents the original performance data of CACformer. Indicates that only integrated factor, that is, only the offset adjustment of the data is considered; Representatives only integrated factors, that is, only the scaling adjustment of the data is considered. Through such an experimental setting, the contribution of each factor to the overall prediction performance can be evaluated separately.

[0033] Further observation of Table 2 shows that the positive scalar factor and displacement vector The addition of can significantly improve the prediction performance of the model. It can be noted that only adding The model performance is slightly better than This may be because the change of local standard deviation may contain important clues about the change of physiological state. By adjusting the local standard deviation of the physiological series, the model is provided with key information about volatility, which may be particularly important for blood pressure prediction. When combined with the application, the model performance is improved in the short-term and medium-term forecasts with a forecast length of 5 and 10 minutes, showing the advantage of capturing local statistical characteristics. However, when the forecast time span is further extended to the long-term forecast of 15 minutes, The contribution of the combination to performance improvement seems to become marginal. In this case, Model configuration and standalone use This suggests that the dynamic characteristics and distribution of physiological sequences may undergo more complex changes in the longer prediction time range.

[0034] Table 2 Ablation experiment results

[0035] Finally, the results in Table 2 show that the fusion and The performance of the two factor-adapted models on the intraoperative blood pressure prediction task was improved, confirming the positive role of the distribution adaptation factor in improving the prediction performance.

[0036] The present application also provides a system for continuous prediction of intraoperative invasive blood pressure, the system comprising: Memory for storing computer programs; A processor, configured to execute the computer program to implement the following method: Encoding the position, timestamp, and feature dimension information in the input sequence to obtain an input vector; wherein the input sequence includes a first input sequence and a second input sequence, and the input vector includes a first input vector and a second input vector, and the first input vector and the second input vector are obtained by encoding the first input sequence and the second input sequence, respectively; Calculate the statistical characteristics of the blood pressure sequence during the operation; wherein the statistical characteristics of the blood pressure sequence include The mean of the input time window and standard deviation ; Based on the input intraoperative blood pressure sequence and its statistical characteristics, a distribution adaptation factor is learned; Guiding calculation of distribution adaptive attention based on the distribution adaptive factor to obtain distribution adaptive attention; constructing an encoder and a decoder, wherein the encoder is configured to take the first input vector as an input sequence of the encoder, capture long-term dependencies in the input sequence based on the distribution adaptation attention, and adopt multi-head attention to learn multiple feature expressions in parallel as the output of the encoder; The decoder is configured to rely on an autoregressive mechanism when generating predictions for future time steps. Each step uses the previous prediction as input to construct a continuous output sequence. At the same time, a multi-head attention mechanism is used to not only process the second input vector, but also fuse the output of the encoder with the information of the current step, and output the prediction result after a series of feedforward networks and linear transformations.

[0037] In some embodiments, the position, timestamp, and feature dimension information in the input sequence are encoded to obtain an input vector, including: Use fixed position encoding to retain the relative position information in the input sequence to obtain the position vector , Each global timestamp is represented by an embedding with a finite vocabulary size , to understand the positional relationship of the input data in the entire time series; Using a one-dimensional convolution filter, the input sequence is projected onto Vector of dimensions ; The position vector , embedded representation and vector Add to get the time-coded input vector .

[0038] In some embodiments, based on the input intraoperative blood pressure sequence and its statistical characteristics, the distribution adaptation factor is learned using the following formula: ; (1) Where, represents a positive scalar factor, represents the intraoperative blood pressure sequence, Represents a displacement vector.

[0039] In some embodiments, the attention calculation is guided by the distribution adaptation factor, and the distribution adaptation attention is calculated by the following formula: (2) Where, Represents the input query, key, and value respectively; Represents the vector dimension.

[0040] The present application also provides a non-transitory computer-readable storage medium storing instructions. When the instructions are executed by a processor, the following steps are performed: S1. Encode the position, timestamp, and feature dimension information in an input sequence to obtain an input vector; wherein the input sequence includes a first input sequence and a second input sequence, and the input vector includes a first input vector and a second input vector, and the first input vector and the second input vector are obtained by encoding the first input sequence and the second input sequence, respectively; S2. Calculate the statistical characteristics of the blood pressure sequence during the operation; wherein the statistical characteristics of the blood pressure sequence include The mean of the input time window and standard deviation ; S3. Based on the input intraoperative blood pressure sequence and its statistical characteristics, a distribution adaptation factor is learned; S4. Calculating distributed adaptive attention based on the distribution adaptation factor to obtain distributed adaptive attention; S5. Construct an encoder and a decoder, wherein the encoder is configured to use the first input vector as the input sequence of the encoder, capture the long-term dependencies in the input sequence based on the distribution adaptive attention, and use multi-head attention to learn multiple feature expressions in parallel as the output of the encoder; the decoder is configured to rely on an autoregressive mechanism when generating predictions for future time steps, and each step uses the previous prediction as input to construct a continuous output sequence. At the same time, the multi-head attention mechanism is used to not only process the second input vector, but also fuse the output of the encoder with the information of the current step, and output the prediction result after a series of feedforward networks and linear transformations.

[0041] It should be noted that the intraoperative invasive blood pressure continuous prediction system and the non-temporary computer-readable storage medium storing instructions proposed in the embodiments of the present application belong to the same technical concept as the previous intraoperative invasive blood pressure continuous prediction device, have the same technical principles and can achieve the same beneficial effects, which will not be repeated here.

[0042] The above implementation modes are only used to illustrate the present application and are not intended to limit the present application. Ordinary technicians in the relevant technical field may make various changes and modifications without departing from the spirit and scope of the present application. Therefore, all equivalent technical solutions also fall within the scope of the present application, and the scope of patent protection of the present application shall be defined by the claims.

Claims

1. A device for continuous prediction of invasive blood pressure during surgery, characterized in that: The device comprises: an embedding layer configured to encode position, timestamp, and feature dimension information in an input sequence to obtain an input vector; wherein the input sequence includes a first input sequence and a second input sequence, and the input vector includes a first input vector and a second input vector, wherein the first input vector and the second input vector are obtained by encoding the first input sequence and the second input sequence, respectively; A statistical feature extractor is configured to calculate the statistical features of the intraoperative blood pressure sequence; wherein the statistical features of the blood pressure sequence include The mean of the input time window and standard deviation ; a multi-layer perceptron module, wherein the signal output terminal of the statistical feature extractor is connected to the signal input terminal of the multi-layer perceptron module, and the multi-layer perceptron module is configured to learn a distribution adaptation factor in response to an input intraoperative blood pressure sequence and its statistical features; A distribution adaptive attention layer, wherein the signal output end of the multilayer perceptron module is connected to the signal input end of the distribution adaptive attention layer, and the distribution adaptive attention layer is configured to guide the calculation of the distribution adaptive attention based on the distribution adaptation factor to obtain the distribution adaptive attention; an encoder, wherein the signal output ends of the distribution adaptive attention layer and the embedding layer are connected to the signal input end of the encoder, and the encoder is configured to use the first input vector as an input sequence of the encoder, capture long-term dependencies in the input sequence based on the distribution adaptive attention, and adopt multi-head attention to learn multiple feature expressions in parallel as the output of the encoder; A decoder, wherein the signal output ends of the encoder and the embedding layer are connected to the signal input end of the decoder, and the decoder is configured to rely on an autoregressive mechanism when generating predictions of future time steps, using the previous prediction as input at each step to construct a continuous output sequence. At the same time, a multi-head attention mechanism is used to not only process the second input vector, but also fuse the output of the encoder with the information of the current step, and output the prediction result after a series of feedforward networks and linear transformations.

2. The device for continuous prediction of intraoperative invasive blood pressure according to claim 1, wherein: The embedding layer is further configured to: Use fixed position encoding to retain the relative position information in the input sequence to obtain the position vector , Each global timestamp is represented by an embedding with a finite vocabulary size , to understand the positional relationship of the input data in the entire time series; Using a one-dimensional convolution filter, the input sequence is projected onto Vector of dimensions ; The position vector , embedded representation and vector Add to get the time-coded input vector .

3. The device for continuous prediction of intraoperative invasive blood pressure according to claim 1, wherein: The multi-layer perceptron module is configured to respond to the input intraoperative blood pressure sequence and its statistical features and learn the distribution adaptation factor through the following formula: ； （1） Where, represents a positive scalar factor, represents the intraoperative blood pressure sequence, Represents a displacement vector.

4. The device for continuous prediction of intraoperative invasive blood pressure according to claim 3, wherein: The distribution adaptive attention layer is configured to guide the calculation of the distribution adaptive attention based on the distribution adaptation factor, and the distribution adaptive attention is calculated by the following formula: （2） Where, Represents the input query, key, and value respectively; Represents the vector dimension.

5. A continuous prediction system for invasive blood pressure during surgery, characterized in that: The system comprises: Memory for storing computer programs; A processor, configured to execute the computer program to implement the following method: Encoding the position, timestamp, and feature dimension information in the input sequence to obtain an input vector; wherein the input sequence includes a first input sequence and a second input sequence, and the input vector includes a first input vector and a second input vector, and the first input vector and the second input vector are obtained by encoding the first input sequence and the second input sequence, respectively; Calculate the statistical characteristics of the blood pressure sequence during the operation; wherein the statistical characteristics of the blood pressure sequence include The mean of the input time window and standard deviation ; Based on the input intraoperative blood pressure sequence and its statistical characteristics, a distribution adaptation factor is learned; Guiding calculation of distribution adaptive attention based on the distribution adaptive factor to obtain distribution adaptive attention; constructing an encoder and a decoder, wherein the encoder is configured to take the first input vector as an input sequence of the encoder, capture long-term dependencies in the input sequence based on the distribution adaptation attention, and adopt multi-head attention to learn multiple feature expressions in parallel as the output of the encoder; The decoder is configured to rely on an autoregressive mechanism when generating predictions for future time steps. Each step uses the previous prediction as input to construct a continuous output sequence. At the same time, a multi-head attention mechanism is used to not only process the second input vector, but also fuse the output of the encoder with the information of the current step, and output the prediction result after a series of feedforward networks and linear transformations.

6. The system for continuous prediction of intraoperative invasive blood pressure according to claim 5, wherein: Encode the position, timestamp, and feature dimension information in the input sequence to obtain the input vector, including: Use fixed position encoding to retain the relative position information in the input sequence to obtain the position vector , Each global timestamp is represented by an embedding with a finite vocabulary size , to understand the positional relationship of the input data in the entire time series; Using a one-dimensional convolution filter, the input sequence is projected onto Vector of dimensions ; The position vector , embedded representation and vector Add to get the time-coded input vector .

7. The system for continuous prediction of intraoperative invasive blood pressure according to claim 5, wherein: According to the input intraoperative blood pressure sequence and its statistical characteristics, the distribution adaptation factor is learned by the following formula: ； （1） Where, represents a positive scalar factor, represents the intraoperative blood pressure sequence, Represents a displacement vector.

8. The system for continuous prediction of intraoperative invasive blood pressure according to claim 7, wherein: Based on the distribution adaptation factor, the distribution adaptation attention is calculated by the following formula: （2） Where, Represents the input query, key, and value respectively; Represents the vector dimension.

9. A non-transitory computer-readable storage medium storing instructions, characterized in that: When the instruction is executed by the processor, the following steps are performed: Encoding the position, timestamp, and feature dimension information in the input sequence to obtain an input vector; wherein the input sequence includes a first input sequence and a second input sequence, and the input vector includes a first input vector and a second input vector, and the first input vector and the second input vector are obtained by encoding the first input sequence and the second input sequence, respectively; Calculate the statistical characteristics of the blood pressure sequence during the operation; wherein the statistical characteristics of the blood pressure sequence include The mean of the input time window and standard deviation ; Based on the input intraoperative blood pressure sequence and its statistical characteristics, a distribution adaptation factor is learned; Guiding calculation of distribution adaptive attention based on the distribution adaptive factor to obtain distribution adaptive attention; Constructing an encoder and a decoder, wherein the encoder is configured to take the first input vector as an input sequence of the encoder, capture the input sequence based on the distribution adaptation attention, and adopt multi-head attention to parallelly learn multiple feature expressions as the output of the encoder; The decoder is configured to rely on an autoregressive mechanism when generating predictions for future time steps. Each step uses the previous prediction as input to construct a continuous output sequence. At the same time, a multi-head attention mechanism is used to not only process the second input vector, but also fuse the output of the encoder with the information of the current step, and output the prediction result after a series of feedforward networks and linear transformations.

Images