Intraoperative hypotension prediction method and related products based on spatiotemporal information fusion
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-11-04
- Publication Date
- 2026-08-14
AI Technical Summary
[0005](1)正负例样本不均衡
[0027]相较于现有技术的有益效果：本发明使用术中监测的生理指标，利用基于时空融合的深度学习模型对术中低血压进行提前预测，从而让医生提前采取措施，降低患者的术后不良风险和死亡风险，分别通过FCN层和GRU层从所述输入矩阵中提取生理指标的空间特征
和时间特征
充分考虑了生理指标的局部空间特征和时间特征，先使用FCN使用三个叠加的卷积模块提取生理指标的局部空间特征和捕获变量之间的关联，再使用三个叠加的GRU层充分学习数据的时间依赖关系，有利于提高术中低血压的预测准确性；另外，本发明通过对所述原始样本数据X中的错误数据进行保留，并使用常值-1的方式填补缺失数据(由于医学数据中的错误数据通常为负值)，扩展样本数据，再通过滑动窗口的方法生成一个新的样本数据，并在间隙窗口中根据术中低血压的定义为观察间隙中的数据打上相应的标签，得到模型的输入矩阵
数据，解决了由于医学数据普遍存在着数据缺失和错误的问题，通过样本上采样解决了正常数据普遍比发病数据多而导致的正负例样本不均衡的问题。
Smart Images
Figure CN115719641B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of medical clinical technology, and in particular to a method for predicting intraoperative hypotension based on spatiotemporal information fusion and related products. Background Technology
[0002] Intraoperative hypotension can cause a range of adverse postoperative symptoms in patients. Early prediction of intraoperative hypotension can help doctors take preventative measures, thereby reducing the frequency of patients experiencing it. Intraoperative hypotension prediction uses historical intraoperative monitoring data to predict whether intraoperative hypotension will occur within a future period. Since intraoperative monitoring data is typically a time series, intraoperative hypotension prediction is usually considered a binary classification task within the time series.
[0003] Currently, some intraoperative hypotension prediction methods are based on arterial pressure waveform data. This involves manually designing feature extraction algorithms to extract features from the raw arterial pressure waveform data, and then using machine learning methods such as Logistic Regression (LR), Naive Bayes (NB), and Support Vector Machine (SVM) to classify the extracted features. These methods require manual feature design, while deep learning can automatically learn and recognize features and is often used for intraoperative hypotension prediction. Based on the type of data used, it can be divided into two categories: one is downsampling data from arterial waveform data at a certain frequency (e.g., 100Hz) and then using deep learning methods for classification. However, blood pressure changes are closely related to other physiological indicators, and some researchers have begun to explore a second approach, combining blood pressure indicators with other physiological indicators for intraoperative hypotension prediction. Among deep learning methods, LSTM is widely used for intraoperative hypotension prediction because it can solve the problem of long-term dependencies in time series. CNN, because its convolutional operations can extract local spatial features of multivariate time series and capture the correlations between variables, is also being explored by some researchers for intraoperative hypotension prediction.
[0004] However, in the process of implementing the inventive technical solutions in the embodiments of this application, the inventors of this application discovered that the above-mentioned technical solutions have at least the following technical problems:
[0005] (1) Imbalance between positive and negative samples. In medical data, there are always more normal data than disease data, which is the problem of imbalance between positive and negative samples. This problem will seriously affect the classification performance of the model.
[0006] (2) It is impossible to simultaneously take into account the temporal and spatial features of multivariate time series. Existing methods for predicting intraoperative hypotension based on multiple physiological indicators, which use superimposed LSTM or superimposed CNN, cannot simultaneously extract the temporal and spatial features of multivariate time series. Summary of the Invention
[0007] To address the aforementioned technical issues, this invention utilizes intraoperative physiological indicators and employs a spatiotemporal fusion-based deep learning model to predict intraoperative hypotension in advance, thereby enabling doctors to take preventative measures and reduce postoperative adverse risks and mortality risks for patients.
[0008] The above-mentioned objectives are mainly achieved through the following technical solutions:
[0009] The first aspect is a method for predicting intraoperative hypotension based on spatiotemporal information fusion, including:
[0010] Monitor intraoperative physiological data;
[0011] The monitored physiological data are input into a spatiotemporal fusion-based deep learning model, which outputs the prediction results of intraoperative hypotension.
[0012] The spatiotemporal fusion-based deep learning model includes a preprocessing layer, an FCN layer, a GRU layer, and an output layer. The training method for the model includes:
[0013] Step 1: The preprocessing layer uses the intraoperative monitored physiological index data as the raw sample data X, and processes it to obtain the input matrix that the model can accept. Erroneous data in the original sample data X are retained, and missing data are filled with -1. A new sample data is then generated using a sliding window method. In the gap window, the data in the observation gap are labeled according to the definition of intraoperative hypotension. The labels include 0 and 1, where label 1 indicates that the corresponding data is hypotension data and label 0 indicates that the corresponding data is normal blood pressure data.
[0014] Step 2, the FCN layer uses three stacked convolutional modules from the input matrix. Spatial characteristics of extracting physiological indicators
[0015] Step 3, spatial features The input consists of time features of physiological indicators extracted from three superimposed GRUs.
[0016] Step 4: The output layer calculates and outputs the probability of intraoperative hypotension occurring or not occurring in the future period based on the extracted spatiotemporal fusion features, wherein the spatiotemporal fusion features include the spatial features. and time characteristics
[0017] Secondly, an intraoperative hypotension prediction device based on spatiotemporal information fusion includes:
[0018] The physiological indicators module is used to monitor intraoperative physiological indicators.
[0019] The intraoperative hypotension prediction module is used to input the monitored physiological index data into a spatiotemporal fusion-based deep learning model and output the intraoperative hypotension prediction result.
[0020] The spatiotemporal fusion-based deep learning model includes:
[0021] The preprocessing module is used to take the intraoperative monitored physiological index data as the raw sample data X and process it to obtain the input matrix that the model can receive. Erroneous data in the original sample data X are retained, and missing data are filled with -1. A new sample data is then generated using a sliding window method. In the gap window, the data in the observation gap are labeled according to the definition of intraoperative hypotension. The labels include 0 and 1, where label 1 indicates that the corresponding data is hypotension data and label 0 indicates that the corresponding data is normal blood pressure data.
[0022] The FCN module is used to extract data from the input matrix using three stacked convolutional modules. Spatial characteristics of extracting physiological indicators
[0023] The GRU module is used to store spatial features. The input consists of time features of physiological indicators extracted from three superimposed GRUs.
[0024] The output module is used to calculate and output the probability of intraoperative hypotension occurring or not occurring in the future within a certain period of time based on the extracted spatiotemporal fusion features, wherein the spatiotemporal fusion features include the spatial features. and time characteristics
[0025] Thirdly, an electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements some or all of the steps of the above-described method for predicting intraoperative hypotension based on spatiotemporal information fusion.
[0026] Fourthly, a computer-readable storage medium storing a computer program that, when executed by a processor, implements some or all of the steps of the above-described method for predicting intraoperative hypotension based on spatiotemporal information fusion.
[0027] Compared to existing technologies, this invention offers several advantages: It utilizes intraoperative physiological indicators and a spatiotemporal fusion-based deep learning model to predict intraoperative hypotension in advance, allowing doctors to take preventative measures and reduce postoperative adverse risks and mortality for patients. This is achieved through FCN and GRU layers, respectively, from the input matrix. Spatial characteristics of extracting physiological indicators and time characteristics This invention fully considers the local spatial and temporal features of physiological indicators. First, it uses FCN with three stacked convolutional modules to extract the local spatial features of physiological indicators and the correlation between captured variables. Then, it uses three stacked GRU layers to fully learn the temporal dependencies of the data, which helps improve the prediction accuracy of intraoperative hypotension. Furthermore, this invention expands the sample data by preserving erroneous data in the original sample data X and filling in missing data with a constant value of -1 (since erroneous data in medical data is usually negative). A new sample data is then generated using a sliding window method, and the data in the observation gap are labeled according to the definition of intraoperative hypotension within the gap window to obtain the model's input matrix. The data addresses the problem of missing and erroneous data in medical data, and solves the problem of imbalance between positive and negative samples caused by the fact that normal data are generally more numerous than disease data through sample upsampling. Attached Figure Description
[0028] Figure 1 This is a schematic diagram of the overall structure of the deep learning model based on spatiotemporal fusion provided in Embodiment 1 of the present invention;
[0029] Figure 2 This is a flowchart illustrating the deep learning model training method based on spatiotemporal fusion provided in Embodiment 1 of the present invention.
[0030] Figure 3 This is a schematic diagram of filling missing data using a sliding window, provided in Embodiment 1 of the present invention.
[0031] Figure 4 This is a schematic diagram of the three-layer stacked GRU structure provided in Embodiment 1 of the present invention;
[0032] Figure 5 This is a schematic diagram of the structure of the deep learning model training device based on spatiotemporal fusion provided in Embodiment 2 of the present invention;
[0033] Figure 6 This is a flowchart illustrating the intraoperative hypotension prediction method based on spatiotemporal information fusion provided in Embodiment 3 of the present invention.
[0034] Figure 7 This is a schematic diagram of the intraoperative hypotension prediction device based on spatiotemporal information fusion provided in Embodiment 4 of the present invention. Detailed Implementation
[0035] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.
[0036] Example 1
[0037] A model training method is used to train a spatiotemporal fusion-based deep learning model, which is used for intraoperative hypotension prediction, such as... Figure 1 As shown, the model includes a preprocessing layer, an FCN layer, a GRU layer, a Dropout layer, and an output layer, as follows: Figure 2 As shown, the training method includes the following steps:
[0038] Step 1: The preprocessing layer uses the intraoperative monitored physiological index data as the raw sample data X, and processes it to obtain the input matrix that the model can accept. Specifically, such as Figure 3 As shown, erroneous data in the original sample data X are retained, and missing data is filled with -1. A new sample data is then generated using a sliding window method. In the gap window, the data in the observation gap are labeled according to the definition of intraoperative hypotension. The labels include 0 and 1, where label 1 indicates that the corresponding data is hypotension data and label 0 indicates that the corresponding data is normal blood pressure data. In this embodiment of the invention, intraoperative hypotension is defined as "mean arterial pressure lasting less than 65 mmHg for one minute".
[0039] Step 2, the FCN layer uses three stacked convolutional modules from the input matrix. Spatial characteristics of extracting physiological indicators
[0040] Step 3, spatial features The input consists of time features of physiological indicators extracted from three superimposed GRUs.
[0041] Step 4: The output layer calculates and outputs the probability of intraoperative hypotension occurring or not occurring in the future period based on the extracted spatiotemporal fusion features, wherein the spatiotemporal fusion features include the spatial features. and time characteristics The probabilities of the two classes are calculated using the Softmax function, where the probability of class i is... C is 2.
[0042] Compared to existing technologies, this invention utilizes intraoperative physiological indicators and a spatiotemporal fusion-based deep learning model to predict intraoperative hypotension in advance. This allows doctors to take preventative measures, reducing postoperative adverse risks and mortality for patients. The prediction is achieved through FCN and GRU layers, respectively, from the input matrix. Spatial characteristics of extracting physiological indicators and time characteristics This invention fully considers the local spatial and temporal features of physiological indicators. First, it uses FCN with three stacked convolutional modules to extract the local spatial features of physiological indicators and the correlation between captured variables. Then, it uses three stacked GRU layers to fully learn the temporal dependencies of the data, which helps improve the prediction accuracy of intraoperative hypotension. Furthermore, this invention retains erroneous data in the original sample data X and fills in missing data with a constant value of -1 (since erroneous data in medical data is usually negative). Then, it generates new sample data using a sliding window method, and labels the data in the observation gap according to the definition of intraoperative hypotension in the gap window, thus obtaining the model's input matrix. The data addresses the problem of missing and erroneous data in medical data, and solves the problem of imbalance between positive and negative samples caused by the fact that normal data are generally more numerous than disease data through sample upsampling.
[0043] In a preferred embodiment, step 2 specifically includes:
[0044] Batch normalization (BN) is used on the input matrix in the FCN layer. Standardize the sample data in the sample;
[0045] The standardized sample data is then input into three stacked convolutional modules, each containing a one-dimensional convolution. Convolution operations are performed through these one-dimensional convolutions to extract the spatial features of the time series. The ReLU function is used as the activation function to increase the nonlinearity of the neural network.
[0046] Because multiple physiological indicators can influence each other when used for hypotension prediction, and their shapes can vary across different time periods, this invention employs three stacked convolutional modules to extract local spatial features of the physiological indicators. Each convolutional module contains a one-dimensional convolution, Batch Normalization (BN), and a ReLU activation function.
[0047] One-dimensional convolution can extract local spatial features of a time series using convolution operations. Taking 11 physiological indicators as an example, the size of the convolution kernel is defined as 3×11. This is because the sample distributions in a batch are not similar, and a single sample... The values in each dimension of the dataset differ significantly, necessitating the use of batch normalization (BN) to standardize each batch of samples. This makes the input distribution of each batch more similar, allowing the network to focus on learning the differences between categories. The ReLU activation function is used to increase the non-linearity of the neural network; its equation is: y = max{0, x}. The ReLU function possesses good sparsity and non-linearity properties, enabling it to better uncover relevant features and fit the training data. Input matrix The shape is 1200×11, and after the convolution operation of FCN, the spatial feature has a shape of 1198×128.
[0048] In a preferred embodiment, a convolution operation is performed using one-dimensional convolution to extract spatial features of the time series. Specifically, this includes: using a sliding window in the input matrix The data in the window slides from top to bottom, and the data in the window is multiplied by the data in the convolution kernel one by one and then summed. The sliding step size is 1, and the size of the sliding window and the convolution kernel is 3*n, where n is the number of physiological indicators.
[0049] Taking the above example of using 11 physiological indicators, the convolution operation uses a sliding window. The slider moves from top to bottom, multiplying the data in the window with the data in the convolution kernel one by one and then summing them. The slider stride is 1, and the size of the sliding window and the convolution kernel is 3×11, with a total of 128.
[0050] Using three stacked convolutional modules can extract spatial features of physiological indicators and capture the correlation between indicators, which is beneficial to improving the prediction accuracy of intraoperative hypotension.
[0051] In a preferred embodiment, step 3 specifically includes:
[0052] A three-layer stacked GRU was used to extract the temporal features of physiological indicators. Each layer of GRU includes a reset gate. i and update gate z i , where r i Responsible for processing new input information s i With memory h i-1 Combined, z i Responsible for determining memory h i-1 Which ones need to be memorized? i Retained in the middle;
[0053] r i=sigmoid(W r s i +U r h i-1 )+b r ,
[0054] z i =sigmoid(W Z s i +U z h i-1 )+b z ,
[0055]
[0056]
[0057] Among them, s i Let L represent the input vector at time i, where i ≤ L, and L is... Length, h represents the current memory content. i h represents the final memory content at the current time. L The last memory content output by the last GRU.
[0058] The role of the GRU layer is to learn the temporal dependencies of physiological indicators. Physiological indicators are typical time series, and time series are arranged in chronological order, with data from one moment to the next closely related. Using GRU, the temporal dependencies of physiological indicators can be learned. Employing a three-layer unit allows for a more comprehensive learning of the temporal information of physiological indicators, and its structure is as follows: Figure 4 As shown, the first two GRUs will output all intermediate memory contents h1-h L Because the input of this layer The shape is 1198×128, so the size of L is 1198, h i The shape is 1×128. The final output shape of the first two GRUs is 1198×128. The third GRU will only output the last memory content h. L Its shape is 1×128, and it has
[0059] In a preferred embodiment, step 3 is followed by:
[0060] The dropout layer randomly discards neural units of the GRU layer network to reduce the possibility of overfitting.
[0061] To illustrate the advantages of the technical solution of this invention, a set of experimental data is used below:
[0062] Eleven common physiological indicators were used to predict intraoperative hypotension: systolic blood pressure (ARTSystolic), diastolic blood pressure (ARTDiastolic), mean arterial pressure (ARTMean), heart rate (HR), SpO2 1 pulse rate (SpO2 1 Pulse Rate), SpO2 1 saturation (SpO2 1 Saturation), SpO2 1 perfusion index (SpO2 1 Perfusion Index), CO2 respiration rate (CO2 Respiration Rate), end-tidal CO2, inspired oxygen (Inspired O2), and end-tidal O2.
[0063] Because medical data commonly contains missing and erroneous data (e.g., systolic blood pressure of -34), to enable the final trained model to handle erroneous and missing data, erroneous data is retained. Since erroneous data are mostly negative values, -1 is used to impute missing data. Batch standardization is then used to standardize each batch of samples, and the missing data imputed with -1 will be distinguished from other normal data after batch standardization.
[0064] After fully considering the local spatial and temporal features of the monitoring data, as well as the problem of missing data, a training method for a deep learning model based on spatiotemporal fusion is proposed for intraoperative hypotension prediction. First, FCN is used to extract the local spatial features of the data and the correlation between the captured variables. Then, three superimposed GRUs are used to fully learn the temporal dependencies of the data.
[0065] The training and test sets were divided using 5-fold cross-validation, and experiments were conducted with gap window sizes T of 5 minutes, 10 minutes, and 15 minutes, respectively. The experimental results are shown in Table 1. The evaluation results are formatted as the average of 5 results (minimum value - maximum value). As can be seen from Table 1, compared with models LSTM and GRU that only consider temporal features or models FCN that only consider spatial features, the spatiotemporal fusion model FCN_GRU proposed in this invention can more accurately predict intraoperative hypotension.
[0066] Table 1 Overall Experimental Results
[0067]
[0068]
[0069]
[0070] To verify that class imbalance affects model performance, negative samples in the training set were upsampled by 10 times within a 5-minute interval window. The experimental results showed that the recall, F-metric, and AUC of all models were improved, but the precision decreased.
[0071] Table 2 Experimental results after 10x upsampling
[0072]
[0073] Example 2
[0074] A model training device is used to train a spatiotemporal fusion-based deep learning model for intraoperative hypotension prediction. The model includes a preprocessing layer, an FCN layer, a GRU layer, and an output layer, as follows: Figure 5 As shown, the device includes:
[0075] The preprocessing module is used to take the intraoperative monitored physiological index data as the raw sample data X and process it to obtain the input matrix that the model can receive. Erroneous data in the original sample data X are retained, and missing data are filled with -1. A new sample data is then generated using a sliding window method. In the gap window, the data in the observation gap are labeled according to the definition of intraoperative hypotension. The labels include 0 and 1, where label 1 indicates that the corresponding data is hypotension data and label 0 indicates that the corresponding data is normal blood pressure data.
[0076] The FCN module is used to extract data from the input matrix using three stacked convolutional modules. Spatial characteristics of extracting physiological indicators
[0077] The GRU module is used to store spatial features. The input consists of time features of physiological indicators extracted from three superimposed GRUs.
[0078] The output module is used to calculate and output the probability of intraoperative hypotension occurring or not occurring in the future within a certain period of time based on the extracted spatiotemporal fusion features, wherein the spatiotemporal fusion features include the spatial features. and time characteristics
[0079] In a preferred embodiment, the FCN module is specifically used for:
[0080] Batch normalization (BN) is used on the input matrix in the FCN layer. Standardize the sample data in the sample;
[0081] The standardized sample data is then input into three stacked convolutional modules, each containing a one-dimensional convolution. Convolution operations are performed using these one-dimensional convolutions to extract the spatial features of the time series data. The ReLU function is used as the activation function to increase the nonlinearity of the neural network.
[0082] In a preferred embodiment, the GRU module is specifically used for:
[0083] A three-layer stacked GRU was used to extract the temporal features of physiological indicators. Each layer of GRU includes a reset gate. i and update gate z i , where r i Responsible for processing new input information s i With memory h i-1 Combined, z i Responsible for determining memory h i-1 Which ones need to be memorized? i Retained in the middle;
[0084] r i =sigmoid(W r s i +U r h i-1 )+b r ,
[0085] z i =sigmoid(W Z s i +U z h i-1 )+b z ,
[0086]
[0087]
[0088] Among them, s i Let L represent the input vector at time i, where i ≤ L, and L is... Length, h represents the current memory content. i h represents the final memory content at the current time. L The last memory content output by the last GRU.
[0089] In a preferred embodiment, the device further includes:
[0090] The Dropout module is used to randomly discard neurons in the GRU layer network to reduce the possibility of overfitting.
[0091] The model training apparatus of this invention corresponds to the model training method described in the above embodiments and implements the corresponding functions. Since the steps of the model training method have been described in detail in the above embodiments, they will not be repeated here.
[0092] Example 3
[0093] Intraoperative hypotension prediction method based on spatiotemporal information fusion, such as Figure 6 As shown, it includes the following steps:
[0094] Monitor intraoperative physiological data;
[0095] The monitored physiological data are input into a spatiotemporal fusion-based deep learning model, which outputs the prediction results of intraoperative hypotension.
[0096] The specific training method for the spatiotemporal fusion-based deep learning model can be found in Example 1.
[0097] Example 4
[0098] Intraoperative hypotension prediction device based on spatiotemporal information fusion, such as Figure 7 As shown, it includes:
[0099] The physiological indicators module is used to monitor intraoperative physiological indicators.
[0100] The intraoperative hypotension prediction module is used to input the monitored physiological index data into a spatiotemporal fusion-based deep learning model and output the intraoperative hypotension prediction result.
[0101] The specific training method for the spatiotemporal fusion-based deep learning model can be found in Example 1.
[0102] Example 5: A computer-readable storage medium storing a computer program that, when executed by a processor, implements some or all of the steps of the above-described method for predicting intraoperative hypotension based on spatiotemporal information fusion.
[0103] The embodiments of the present invention have been described in detail above. Specific examples have been used to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only for the purpose of helping to understand the method and core ideas of the present invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.
Claims
1. A method for predicting intraoperative hypotension based on spatiotemporal information fusion, characterized in that, include: Monitor intraoperative physiological data; The monitored physiological data are input into a spatiotemporal fusion-based deep learning model, which outputs the prediction results of intraoperative hypotension. The spatiotemporal fusion-based deep learning model includes a preprocessing layer, an FCN layer, a GRU layer, and an output layer. The training method for the model includes: Step 1: The preprocessing layer uses the intraoperative monitored physiological index data as the raw sample data X, and processes it to obtain the input matrix that the model can receive. Erroneous data in the original sample data X are retained, and missing data are filled with -1. A new sample data is then generated using a sliding window method. In the gap window, the data in the observation gap are labeled according to the definition of intraoperative hypotension. The labels include 0 and 1, where label 1 indicates that the corresponding data is hypotension data and label 0 indicates that the corresponding data is normal blood pressure data. Step 2, the FCN layer uses three stacked convolutional modules from the input matrix. Spatial characteristics of extracting physiological indicators Step 3, spatial features The input consists of time features of physiological indicators extracted from three superimposed GRUs. Step 4: The output layer calculates and outputs the probability of intraoperative hypotension occurring or not occurring in the future period based on the extracted spatiotemporal fusion features, wherein the spatiotemporal fusion features include the spatial features. and time characteristics 2. The intraoperative hypotension prediction method based on spatiotemporal information fusion as described in claim 1, characterized in that, Step 2 specifically includes: Batch normalization (BN) is used on the input matrix in the FCN layer. Standardize the sample data in the sample; The standardized sample data is input into three stacked convolutional modules, each containing a one-dimensional convolution. Convolution operations are performed through these one-dimensional convolutions to extract the spatial features of the time series data.
3. The intraoperative hypotension prediction method based on spatiotemporal information fusion as described in claim 2, characterized in that, The method involves performing convolution operations using one-dimensional convolution to extract spatial features from the time series. Specifically, it includes: Using a sliding window on the input matrix The data in the window slides from top to bottom, and the data in the window is multiplied by the data in the convolution kernel one by one and then summed. The sliding step size is 1, and the size of the sliding window and the convolution kernel is 3*n, where n is the number of physiological indicators.
4. The intraoperative hypotension prediction method based on spatiotemporal information fusion as described in claim 2 or 3, characterized in that, Also includes: Using the ReLU function as the activation function increases the nonlinearity of the neural network.
5. The intraoperative hypotension prediction method based on spatiotemporal information fusion as described in claim 4, characterized in that, Step 3 specifically includes: A three-layer stacked GRU was used to extract the temporal features of physiological indicators. Each layer of GRU includes a reset gate. i and update gate z i , where r i Responsible for processing new input information s i With memory h i-1 Combined, z i Responsible for determining memory h i-1 Which ones need to be memorized? i Retained in the middle; r i =sigmoid(W r s i +U r h i-1 )+b r ; z i =sigmoid(W Z s i +U z h i-1 )+b z ; Among them, s i Let L represent the input vector at time i, where i ≤ L, and L is... Length, h represents the current memory content. i h represents the final memory content at the current time. L The last memory content output by the last GRU.
6. The method for predicting intraoperative hypotension based on spatiotemporal information fusion as described in any one of claims 1 to 5, characterized in that, Step 3 is followed by: The neural units of the GRU layer network are randomly dropped by the Dropout layer.
7. A device for predicting intraoperative hypotension based on spatiotemporal information fusion, characterized in that, include: The physiological indicators module is used to monitor intraoperative physiological indicators. The intraoperative hypotension prediction module is used to input the monitored physiological index data into a spatiotemporal fusion-based deep learning model and output the intraoperative hypotension prediction result. The spatiotemporal fusion-based deep learning model includes: The preprocessing module is used to take the intraoperative monitored physiological index data as the raw sample data X and process it to obtain the input matrix that the model can receive. Erroneous data in the original sample data X are retained, and missing data are filled with -1. A new sample data is then generated using a sliding window method. In the gap window, the data in the observation gap are labeled according to the definition of intraoperative hypotension. The labels include 0 and 1, where label 1 indicates that the corresponding data is hypotension data and label 0 indicates that the corresponding data is normal blood pressure data. The FCN module is used to extract data from the input matrix using three stacked convolutional modules. Spatial characteristics of extracting physiological indicators The GRU module is used to extract spatial features. The input consists of time features of physiological indicators extracted from three superimposed GRUs. The output module is used to calculate and output the probability of intraoperative hypotension occurring or not occurring in the future within a certain period of time based on the extracted spatiotemporal fusion features, wherein the spatiotemporal fusion features include the spatial features. and time characteristics 8. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the intraoperative hypotension prediction method based on spatiotemporal information fusion as described in any one of claims 1 to 6.
9. An electronic device, the device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the intraoperative hypotension prediction method based on spatiotemporal information fusion as described in any one of claims 1 to 6.