Method, device and equipment for predicting hypotension probability in perioperative period and medium

Abstract

The invention provides a perioperative period hypotension probability prediction method which can be applied to the technical field of perioperative periods. The method comprises the following steps: collecting a multi-modal physiological signal and static demographic information of a patient in a perioperative period; segmenting the multi-modal physiological signal into signal segments with fixed lengths to serve as analysis units; performing weighted summation on time proximity, hemodynamic severity and blood pressure tendency to obtain a continuous perioperative period hypotension probability soft label of each signal fragment; discrete medical priori features are extracted from the arterial blood pressure signals in the signal segments; based on the signal segments and the medical prior features, frequency domain features and time domain features are extracted by using a double-current neural network; splicing the frequency domain feature, the time domain feature and the medical prior feature into a fusion feature vector; training a double-flow neural network by using the fusion feature vector and an adaptive optimization algorithm to obtain a trained double-flow neural network; and predicting the hypotension probability in the perioperative period by using a double-flow neural network.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of perioperative period, more particularly to a perioperative hypotension probability prediction method, device, equipment and medium. BACKGROUND

[0002] Perioperative hypotension is a common complication during anesthesia and surgery, which is closely related to patient's underlying disease, anesthetic drug effect, surgical operation stimulation and body fluid imbalance. Hypotension can lead to insufficient perfusion of vital organs, significantly increase the risk of postoperative complications such as myocardial injury, acute kidney injury, cognitive dysfunction, and even threaten the life of the patient. Therefore, accurately and real-time predicting the risk of perioperative hypotension and taking preventive measures in advance is the key to improving the prognosis of patients.

[0003] However, the prevention and monitoring of hypotension in current clinical practice still faces multiple challenges. First, anesthesiologists mainly rely on real-time blood pressure monitoring combined with personal experience for risk assessment, but this method is highly subjective, different doctors have different interpretations of blood pressure fluctuation trends, and is easily affected by factors such as fatigue and distraction; at the same time, blood pressure monitoring is a "post-feedback", which can only trigger intervention when hypotension has occurred, while organ damage may have occurred within minutes after blood pressure drops, limiting the effectiveness of prevention. Second, some studies attempt to predict risk by combining preoperative patient information with traditional statistical models, but such methods have fuzzy output and lack real-time capability, making it difficult to dynamically capture the dynamic changes of intraoperative physiological signals and adapt to the rapid fluctuations in patient status during surgery. In addition, in recent years, machine learning technology has been introduced into this field, which realizes real-time prediction by analyzing arterial pressure waveform, electrocardiogram, pulse oxygen saturation and other signals, but still has significant limitations: model training data are mostly derived from specific hospitals or surgical types, with insufficient generalization ability, leading to performance decline in different patient groups or surgical scenarios; at the same time, the heterogeneity of patient physiological characteristics is not fully considered, and the prediction accuracy for high-risk patients is low; in addition, feature engineering relies on manual design, requiring manual extraction of time / frequency domain features, which may miss key information and the feature selection process lacks theoretical basis.

[0004] In summary, the existing technology in perioperative hypotension prediction generally has problems such as insufficient timeliness, imbalance between accuracy and individualization, and lack of explainability. SUMMARY

[0005] (I) Technical problems to be solved

[0006] To solve the core technical problems of low accuracy, inability to effectively handle individual differences, and difficulty in achieving precise real-time warning in the prediction of perioperative hypotension, the present application provides a perioperative hypotension probability prediction method, device, equipment and medium, through the innovative dual-flow neural network architecture, using DWT (discrete wavelet transform) -Transformer frequency domain flow and block-convolution time domain flow, the time domain and frequency domain features of multi-modal physiological signals can be synchronously and deeply processed, which is the core of high-precision prediction; at the same time, the unique perioperative hypotension probability soft label strategy converts the traditional binary label into a continuous probability curve that integrates time proximity, hemodynamic severity and blood pressure trend, effectively improving the continuity and fineness of the prediction; in addition, by integrating 81-dimensional medical prior discrete features and deep features extracted by the dual-flow network, the effective fusion of medical prior knowledge and deep learning features is achieved, enhancing the model's interpretability, robustness and generalization ability.

[0007] (Two) Technical solutions

[0008] To solve the above technical problems, the embodiments of the present application provide a perioperative hypotension probability prediction method, device, equipment and medium.

[0009] According to a first aspect of the present application, a perioperative hypotension probability prediction method is provided, comprising: collecting multi-modal physiological signals and static demographic information of a patient during the perioperative period; identifying perioperative hypotension event segments and normal event segments in the multi-modal physiological signals based on a predetermined standard, and cutting the continuous multi-modal physiological signals into signal segments of a fixed length as analysis units; obtaining a continuous perioperative hypotension probability soft label for each signal segment by weighted sum of time proximity, hemodynamic severity and blood pressure trend; extracting discrete medical prior features from the arterial blood pressure signal in the signal segment; based on the signal segment and the medical prior features, using a dual-flow neural network to extract global frequency domain features through a frequency domain flow and local and global time domain features through a time domain flow; concatenating the frequency domain features, time domain features and medical prior features into a fusion feature vector; designing a loss function based on the distribution difference of the fusion feature vector and the soft label, using the fusion feature vector to train the dual-flow neural network using an adaptive optimization algorithm to obtain a trained dual-flow neural network; and using the dual-flow neural network to predict the perioperative hypotension probability to obtain a hypotension probability prediction curve.

[0010] In some exemplary embodiments, the multi-modal physiological signals include at least one of an arterial blood pressure signal, an electrocardiogram signal or a photoplethysmography pulse wave signal; and the static demographic information includes at least one of age, gender, height or weight.

[0011] In some example embodiments, the time proximity is calculated based on a time difference between the current signal segment and the occurrence of the perioperative hypotension event, the smaller the time difference, the higher the time proximity weight; the hemodynamic severity is calculated based on the mean arterial pressure value of the current signal segment, the lower the mean arterial pressure, the higher the hemodynamic severity weight; the blood pressure trend is calculated based on the slope of the mean arterial pressure change in a short period before the current signal segment, the faster the blood pressure decreases, the higher the blood pressure trend weight.

[0012] In some example embodiments, the medical prior features include morphological features, statistical features, hemodynamic parameters, and time-frequency domain features, wherein the morphological features include pulse transit time; the statistical features include mean arterial pressure and standard deviation of arterial pressure; the hemodynamic parameters include cardiac output, stroke volume, and peripheral vascular resistance; and the time-frequency domain features include the ratio of low-frequency component to high-frequency component of heart rate variability.

[0013] In some example embodiments, the global frequency domain features are extracted through a frequency domain stream, including: independently performing discrete wavelet transform on each physiological signal channel, decomposing the signal into wavelet coefficients of different frequency bands through a filter bank, and then converting the wavelet coefficients into a spectrum block; assigning a position code to each spectrum block to obtain a spectrum block with a position code; inputting the spectrum block with the position code into a Transformer encoder to capture long-range dependencies between frequency domain features through a self-attention mechanism, and outputting global frequency domain features.

[0014] In some example embodiments, the local and global time domain features are extracted through a time domain stream, including: dividing the time domain signal into overlapping small segments, each segment containing a fixed number of sampling points; taking the arterial blood pressure signal segment as a query, the electrocardiogram and photoplethysmography pulse wave signal segments as keys and values, and enhancing the feature representation of each small segment through a cross-modal attention mechanism to obtain enhanced features; inputting the enhanced features into a multi-layer one-dimensional convolutional network to extract local and global time domain features.

[0015] In some example embodiments, the fusion feature vector is used to train the dual-stream neural network using an adaptive optimization algorithm, including: inputting the fusion feature vector into the fully connected layer of the dual-stream neural network, mapping the high-dimensional features to the target prediction space through the fully connected layer; applying an activation function to the output of the fully connected layer to generate a continuous probability estimate value of the occurrence of the perioperative hypotension event in the future preset time period; calculating the relative entropy between the probability estimate value and the pre-generated soft label probability distribution, and taking the relative entropy as a loss function; and iteratively updating the weight parameters of the dual-stream neural network using a stochastic gradient descent optimization algorithm to minimize the loss function.

[0016] According to a second aspect of the present application, a perioperative hypotension probability prediction device is provided, comprising: a data acquisition module configured to acquire multi-modal physiological signals and static demographic information of a patient during a perioperative period; a preprocessing module configured to identify perioperative hypotension event segments and normal event segments in the multi-modal physiological signals based on preset standards, and cut the continuous multi-modal physiological signals into signal segments of a fixed length as analysis units; a first acquisition module configured to obtain a continuous perioperative hypotension probability soft label of each signal segment by weighted summation of time proximity, hemodynamic severity and blood pressure trend; a second acquisition module configured to extract discrete medical prior features from an arterial blood pressure signal in the signal segment; a third acquisition module configured to extract global frequency domain features through a frequency domain stream and local and global time domain features through a time domain stream based on the signal segment and the medical prior features by using a dual-flow neural network; a feature fusion module configured to splice the frequency domain features, the time domain features and the medical prior features into a fusion feature vector; a network training module configured to design a loss function based on a distribution difference between the fusion feature vector and the soft label, train the dual-flow neural network by using the fusion feature vector and an adaptive optimization algorithm, and obtain a trained dual-flow neural network; and a probability prediction module configured to predict a perioperative hypotension probability by using the dual-flow neural network, and obtain a hypotension probability prediction curve.

[0017] According to a third aspect of the present application, an electronic device is provided, comprising: one or more processors; a memory configured to store one or more computer programs, wherein the one or more processors execute the one or more computer programs to implement the steps of the method.

[0018] According to a fourth aspect of the present application, a computer-readable storage medium is provided, which stores a computer program or instructions, wherein the computer program or instructions are executed by a processor to implement the steps of the method.

[0019] (Three) beneficial effects

[0020] As can be seen from the above technical solutions, the perioperative hypotension probability prediction method, device, equipment and medium provided by the embodiments of the present application have at least the following beneficial effects:

[0021] (1) The dual-flow network is used to comprehensively extract time-frequency features of multi-modal signals, and the medical prior is fused, which overcomes the limitations of traditional single signal source and insufficient feature extraction, and improves the prediction accuracy and robustness.

[0022] (2) The innovative probability soft label strategy converts the hypotension risk into a continuous probability curve containing time, severity and trend information, provides a dynamic risk view superior to traditional binary warning, is beneficial to early and accurate intervention, and realizes dynamic, continuous and fine risk assessment.

[0023] (3) Deep learning model can more sensitively identify early signs through fine mining of multi-dimensional features and continuous probability optimization, and reduce unnecessary false positives due to the continuity of risk assessment, enhance early warning capability, and reduce false positives.

[0024] (4) By integrating multi-modal real-time data, demographic information and hemodynamic parameters calculated therefrom as prior features, the prediction is more physiologically meaningful, providing a basis for individualized management and improving clinical practicability and individualized monitoring potential. BRIEF DESCRIPTION OF DRAWINGS

[0025] The above and other objects, features and advantages of the present application will become more apparent from the following description when taken in conjunction with the accompanying drawings, in which:

[0026] Figure 1 A structural diagram of a perioperative hypotension probability prediction system according to an embodiment of the present application is schematically shown.

[0027] Figure 2 A flowchart of a perioperative hypotension probability prediction method according to an embodiment of the present application is schematically shown.

[0028] Figure 3 An optimal transmission perioperative hypotension probability prediction result according to an embodiment of the present application is schematically shown.

[0029] Figure 4 A dual-flow neural network extraction and fusion flowchart according to an embodiment of the present application is schematically shown.

[0030] Figure 5 A perioperative hypotension probability prediction device according to an embodiment of the present application is schematically shown.

[0031] Figure 6 A block diagram of an electronic device of a perioperative hypotension probability prediction method according to an embodiment of the present application is schematically shown. DETAILED DESCRIPTION

[0032] Hereinafter, embodiments of the present application will be described with reference to the accompanying drawings. It is to be understood, however, that these descriptions are merely exemplary and are not intended to limit the scope of the present application. In the following detailed description, numerous specific details are set forth in order to provide a thorough understanding of the embodiments of the present application. However, it will be apparent to one skilled in the art that one or more embodiments can be practiced without these specific details. In other instances, well-known structures and techniques have been omitted in order to avoid obscuring the concepts of the present application.

[0033] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of the application. As used herein, the terms "comprises", "comprising", "includes", "including" and the like are specifically intended to be open-ended and to mean that other features, steps, operations, and / or components can be added.

[0034] All terms used herein including technical and scientific terms have the meanings commonly understood by one of ordinary skill in the art unless otherwise defined herein. It should be noted that the terms "comprise", "comprising", "comprises" and the like can have the meaning ascribed to it in U.S. Patent law; it can mean "includes", "including", and the like; and it can further mean "consists of".

[0035] In situations where similar terminology is used for similar purposes, it is intended that variations on the terminology be understood as referring to similar concepts. For example, where a component is referred to as being "connected" to another component, it is intended that "connected" can include being directly connected to, or being indirectly connected to via another component.

[0036] Figure 1 A structure diagram of a perioperative hypotension probability prediction system according to an embodiment of the present application is schematically shown.

[0037] As shown in Figure 1 , a perioperative hypotension probability prediction system according to an embodiment of the present application.

[0038] As shown in Figure 1 , a perioperative hypotension probability prediction system according to an embodiment of the present application comprises a physiological signal acquisition module, a patient information input module, a data preprocessing module, a soft label probability generation module, a probability prediction core processing unit, and a data display and early warning module, wherein all the modules are integrated in a multi-modal physiological signal acquisition device. The probability prediction core processing unit runs a double-flow neural network model, which includes a medical prior feature extraction submodule and a double-flow deep feature extraction submodule, and finally predicts and outputs the continuous risk probability of perioperative hypotension and the optional classification result through a feature fusion and probability output submodule, and provides real-time risk information to clinicians through the data display and early warning module.

[0039] Based on Figure 1 , a perioperative hypotension probability prediction system, the present application further provides a perioperative hypotension probability prediction method, which combines Figures 2-4 The method is described in detail.

[0040] Figure 2 A flowchart of a perioperative hypotension probability prediction method according to an embodiment of the present application is schematically shown.

[0041] As Figure 2 shown, a perioperative hypotension probability prediction method according to an embodiment of the present application includes steps S110-S180.

[0042] In step S110, the consent or authorization of the patient is obtained, and after the consent or authorization of the patient is obtained, the multi-modal physiological signals and static demographic information of the patient during the perioperative period are collected.

[0043] In some exemplary embodiments, the multi-modal physiological signals include arterial blood pressure signals (ABP), electrocardiogram signals (ECG), and photoplethysmogram signals (PPG), and the static demographic information includes age, gender, height, and weight. The multi-modal physiological signals can comprehensively reflect the physiological state of the patient during the perioperative period, and the static demographic information can be used as auxiliary information to provide a rich data basis for subsequent accurate prediction of perioperative hypotension probability.

[0044] In step S120, the perioperative hypotension event segments and normal event segments in the multi-modal physiological signals are identified based on a preset standard, and the continuous multi-modal physiological signals are cut into fixed-length signal segments as analysis units.

[0045] Before step S120, data cleaning (including denoising, outlier processing, and missing value interpolation) of the original multi-modal physiological signals can also be included to obtain preprocessed data.

[0046] Identifying event segments and cutting signal segments helps to convert complex continuous signals into analyzable units for subsequent processing and analysis. Data cleaning can improve data quality and reduce the interference of noise and outliers on the prediction results.

[0047] In some exemplary embodiments, the preset standard can adopt a clinical standard, such as an average arterial pressure less than 65 mmHg for more than 1 minute.

[0048] In step S130, a continuous perioperative hypotension probability soft label for each signal segment is obtained by weighted summation of time proximity, hemodynamic severity, and blood pressure trend.

[0049] In an embodiment of the present application, a continuous perioperative hypotension probability soft label P_prob is generated for each signal segment, replacing the traditional binary label. The soft label fuses the following three-dimensional information: time proximity (P_time): based on the time difference between the current segment and the occurrence of perioperative hypotension events Calculation, The smaller, the higher P_time; hemodynamic severity (P_spatial): calculated based on the mean arterial pressure (MAP) value of the current segment, the lower the MAP, the higher P_spatial; blood pressure trend (P_trend): calculated based on the change slope of MAP in the past short period, the more obvious the downward trend, the higher P_trend. The data of the above three dimensions are normalized and the probability is calculated using the same function, and the final soft label is the weighted sum:

[0050] (1)

[0051] wherein , represents the weight of the time proximity dimension, represents the weight of the hemodynamic severity dimension, and γ represents the weight of the blood pressure trend dimension. The combination of the last three probabilities becomes the soft label of each segment. The soft label is used to replace the original binary label, which contains more physiological information, and quantifies the risk of hypotension events, improves the continuity and accuracy of the prediction.

[0052] In step S140, discrete medical prior features are extracted from the arterial blood pressure signal in the signal segment.

[0053] In some exemplary embodiments, the medical prior features include morphological features, statistical features, hemodynamic parameters, and time-frequency domain features, wherein the morphological features include pulse transit time; the statistical features include arterial pressure mean and arterial pressure standard deviation; the hemodynamic parameters include cardiac output, stroke volume, peripheral vascular resistance; and the time-frequency domain features include the ratio of low-frequency component and high-frequency component of heart rate variability.

[0054] Figure 3 The pre-processing of multiple bioelectric signals and the extraction of medical features are schematically shown.

[0055] As shown in Figure 3 , physiological-related discrete features are extracted from each ABP signal segment. First, ABP signal fine processing, such as low-pass filtering, detecting diastolic pressure, systolic pressure, and other key points. Second, hemodynamic parameter calculation: based on Windkessel model, combined with demographic information, to estimate cardiac output (CO), stroke volume (SV), peripheral vascular resistance (SVR), etc. For example:

[0056] (2)

[0057] wherein is the area under the pulse wave pressure curve, is the input impedance, Z0 is the characteristic impedance of the aorta:

[0058] (3)

[0059] wherein, is the density of blood, is a function of the cross-sectional area of the blood vessel related to the pressure P, and is calculated as:

[0060] (4)

[0061] wherein, A max represents the maximum cross-sectional area of the blood vessel, and The linear regression model is associated with the age and gender of the patient, and the specific parameters are fitted by clinical data. In the calculation of Z0, the height and weight of the patient are used.

[0062] (5)

[0063] wherein, represents the body surface area, Height represents the height of the patient, and Weight represents the weight of the patient.

[0064] Based on the above method, a total of 81-dimensional medical prior features are obtained.

[0065] The medical prior features integrate clinical medical knowledge and physiological signal features, which can enhance the interpretability, robustness and generalization ability of the model, and make the prediction results more consistent with the actual clinical situation.

[0066] In step S150, based on the signal segment and the medical prior feature, a dual-flow neural network is used to extract global frequency domain features through a frequency domain flow and to extract local and global time domain features through a time domain flow.

[0067] In some exemplary embodiments, the global frequency domain features are extracted through the frequency domain flow, including: performing discrete wavelet transform on each physiological signal channel independently, decomposing the signal into wavelet coefficients of different frequency bands through a filter bank, and then converting the wavelet coefficients into a spectrum block; assigning a position code to each spectrum block to obtain a spectrum block with a position code; inputting the spectrum block with the position code into a Transformer encoder, capturing long-range dependency relationships between frequency domain features through a self-attention mechanism, and outputting global frequency domain features.

[0068] ​In some exemplary embodiments, the local and global time domain features are extracted by time domain stream extraction, including: dividing the time domain signal into overlapping small segments, each segment containing a fixed number of sampling points; taking the arterial blood pressure signal segment as the query, the electrocardiogram and the photoplethysmography signal segment as the key and the value, enhancing the feature representation of each small segment through the cross-modal attention mechanism to obtain the enhanced features; inputting the enhanced features into a multi-layer one-dimensional convolutional network to extract the local and global time domain features.

[0069] The frequency domain stream can capture the global frequency domain features of the signal, reflecting the energy distribution of the signal in different frequency bands; the time domain stream can extract the local and global time domain features of the signal, preserving the time sequence and local change information of the signal. The dual-stream neural network combines the advantages of both, and can more comprehensively extract signal features.

[0070] In step S160, the frequency domain features, the time domain features and the medical prior features are spliced into a fusion feature vector.

[0071] Fusing features from different sources can comprehensively utilize various information, improve the richness and expression ability of the features, and help improve the prediction performance of the model.

[0072] For example, the preprocessed multi-modal signal segments and the medical prior features are input into the dual-stream neural network. (1) DWT-Former frequency domain stream: multi-layer discrete wavelet transform (DWT) is performed on each physiological signal channel (ABP, ECG, PPG), and a suitable wavelet basis (such as db4, db2) is selected. The formula of the wavelet transform is as follows:

[0073] (6)

[0074] (7)

[0075] (8)

[0076] wherein, represents the wavelet coefficient of the discrete wavelet transform of the c-th physiological signal channel at the l-th layer, DWT represents the discrete wavelet transform operation, represents the preprocessed signal of the c-th physiological signal channel, represents the wavelet basis selected for the c-th physiological signal channel, L represents the number of layers of the discrete wavelet transform, and represents the wavelet coefficient of the c-th physiological signal channel at the l-th layer. db4 represents the db4 wavelet basis function, and h k represents the filter coefficient of the db4 wavelet basis. represents the scale function, k represents the summation index, and represents the wavelet coefficient of the c-th physiological signal channel at the l-th layer. db2 represents the db2 wavelet basis function, and g k represents the filter coefficient of the db2 wavelet basis.

[0077] Convert wavelet coefficients into spectral tokens and add positional encoding:

[0078] (9)

[0079] Among them, E (p) This represents the position code of the p-th spectral block. This represents the element in the wavelet coefficient matrix from position pS to position (p+1)S, where i and j are the row and column indices of the matrix.

[0080] Input a Transformer encoder to extract global frequency domain features. The Patch-Conv temporal stream divides each temporal signal into patches. A cross-modal attention mechanism (e.g., using ABP as the query and ECG / PPG as the key / value) is employed to enhance feature representation. These enhanced features are then input into a multi-scale residual one-dimensional convolutional network to extract local and global temporal morphological features. :

[0081] (10)

[0082] Among them, H l Represents the output features of the l-th layer of the one-dimensional convolutional network, ReLU represents the activation function, and Conv1D represents the one-dimensional convolution operation. Indicates the first The output features of a one-dimensional convolutional network, where k represents the kernel size and d represents the dilation rate of the convolution.

[0083] (11)

[0084] Among them, f tm This represents the temporal morphological features after the max pooling operation, where t represents the index, ranging from 1 to T′, and maxpool represents the max pooling operation. This represents the feature of the 16th layer one-dimensional convolutional network at position t.

[0085] (3) Feature fusion: Integrating prior medical features with and The features are concatenated to form a fused feature vector. :

[0086] (12)

[0087] Among them, F concat F represents the fused feature vector. time F represents the local and global temporal features extracted from the temporal stream. freqF represents the global frequency domain feature of the frequency domain stream extraction MLP represents the feature after the medical prior feature is processed by a multi-layer perception (MLP).

[0088] The specific double-flow neural network extraction and fusion process is described in Figure 4 .

[0089] In step S170, a loss function is designed based on the distribution difference of the fusion feature vector and the soft label, the double-flow neural network is trained using the fusion feature vector and an adaptive optimization algorithm, and a trained double-flow neural network is obtained.

[0090] In some exemplary embodiments, step S170 can include: inputting the fusion feature vector into the fully connected layer of the double-flow neural network, mapping the high-dimensional feature to the target prediction space through the fully connected layer; applying an activation function to the output of the fully connected layer to generate a continuous probability estimate value of the occurrence of the perioperative hypotension event in the future preset time period; calculating the relative entropy between the probability estimate value and the pre-generated soft label probability distribution, and taking the relative entropy as the loss function; and using a stochastic gradient descent optimization algorithm to iteratively update the weight parameters of the double-flow neural network to minimize the loss function.

[0091] For example, after the double-flow neural network successfully extracts and fuses the time-frequency domain features of the multi-modal physiological signals and the medical prior discrete features, it enters the final model training and output stage. This stage aims to convert the learned deep feature representation into an accurate and continuous probability estimate of the future perioperative hypotension event risk, and optimize the model performance through a carefully designed loss function and training strategy. The input to the fully connected layer. The fully connected layer serves as the output head of the model, responsible for mapping high-dimensional features to the target prediction space. Subsequently, through the Sigmoid activation function, the continuous probability value of the occurrence of perioperative hypotension in the future minutes is output . This output value is the model's prediction of the specific time window in the future. (1) Loss function: To effectively supervise the model to learn this continuous perioperative hypotension probability and make it approach the target soft label generated in S102, the Kullback-Leibler (KL) divergence is used to measure the difference between the predicted probability distribution and the target soft label distribution : :

[0092] (13)

[0093] (2) Model training: the core goal of model training is to minimize the loss function through iterative optimization. In the training process, efficient stochastic gradient descent optimization algorithms such as Adam (Adaptive Moment Estimation) are used to update the weight parameters of the dual-flow neural network. The training process is carried out on a large and high-quality perioperative physiological signal dataset. To ensure the robustness and good generalization ability of the model performance and avoid overfitting, a k-fold cross-validation strategy is adopted.

[0094] In step S180, the dual-flow neural network is used to predict the probability of perioperative hypotension, and a hypotension probability prediction curve is obtained.

[0095] After model training, the structure and weight values of the model are saved. And stored in the self-developed multi-modal physiological signal acquisition device. After inputting the individual information of the patient, the device starts effective multi-bioelectric signal acquisition, and the probability prediction unit automatically outputs the continuous perioperative hypotension probability, providing accurate and reliable prediction curves for physicians.

[0096] The loss function can effectively supervise the model to learn the continuous perioperative hypotension probability, making it approach the target soft label. The adaptive optimization algorithm can accelerate model convergence and improve training efficiency. The k-fold cross-validation strategy can ensure the robustness and good generalization ability of the model performance, and avoid overfitting.

[0097] Figure 5 A perioperative hypotension probability prediction device according to an embodiment of the application is schematically shown.

[0098] As shown in Figure 5 The perioperative hypotension probability prediction device 800 of this embodiment includes a data acquisition module 810, a preprocessing module 820, a first acquisition module 830, a second acquisition module 840, a third acquisition module 850, a feature fusion module 860, a network training module 870, and a probability prediction module 880.

[0099] The data acquisition module 810 is configured to acquire multi-modal physiological signals and static demographic information of a patient during a perioperative period.

[0100] The preprocessing module 820 is configured to identify perioperative hypotension event segments and normal event segments in the multi-modal physiological signals based on a preset standard, and cut the continuous multi-modal physiological signals into fixed-length signal segments as analysis units.

[0101] The first acquisition module 830 is configured to obtain a continuous perioperative hypotension probability soft label for each signal segment by weighted summation of time proximity, hemodynamic severity, and blood pressure trend.

[0102] The second acquisition module 840 is configured to extract discrete medical prior features from the arterial blood pressure signal in the signal segment.

[0103] The third acquisition module 850 is configured to extract global frequency domain features through a frequency domain stream and extract local and global time domain features through a time domain stream based on the signal segment and the medical prior features by using a dual-stream neural network.

[0104] The feature fusion module 860 is configured to splice the frequency domain features, the time domain features and the medical prior features into a fusion feature vector.

[0105] The network training module 870 is configured to design a loss function based on a distribution difference between the fusion feature vector and the soft label, train the dual-stream neural network by using the fusion feature vector and using an adaptive optimization algorithm to obtain a trained dual-stream neural network.

[0106] The probability prediction module 880 is configured to predict a probability of perioperative hypotension by using the dual-stream neural network to obtain a hypotension probability prediction curve.

[0107] According to an embodiment of the present application, any multiple modules of the data acquisition module 810, the preprocessing module 820, the first acquisition module 830, the second acquisition module 840, the third acquisition module 850, the feature fusion module 860, the network training module 870 and the probability prediction module 880 can be combined in one module for implementation, or any one of the modules can be split into multiple modules. Alternatively, at least part of the functions of one or more of the modules can be combined with at least part of the functions of other modules, and implemented in one module. According to an embodiment of the present application, at least one of the data acquisition module 810, the preprocessing module 820, the first acquisition module 830, the second acquisition module 840, the third acquisition module 850, the feature fusion module 860, the network training module 870 and the probability prediction module 880 can be at least partially implemented as a hardware circuit, such as a field programmable gate array (FPGA), a programmable logic array (PLA), a system on chip, a system on board, a system on package, an application specific integrated circuit (ASIC), or any other reasonable manner of integrating or packaging a circuit, etc. hardware or firmware, or implemented in any one of software, hardware and firmware or in a proper combination of any of them. Alternatively, at least one of the data acquisition module 810, the preprocessing module 820, the first acquisition module 830, the second acquisition module 840, the third acquisition module 850, the feature fusion module 860, the network training module 870 and the probability prediction module 880 can be at least partially implemented as a computer program module which can perform corresponding functions when the computer program module is run.

[0108] Figure 6 A block diagram of an electronic device according to an embodiment of the method of perioperative hypotension probability prediction of the present application is schematically shown.

[0109] As shown in Figure 6 The electronic device 900 according to an embodiment of the present application includes a processor 901 which can perform various appropriate actions and processes in accordance with a program stored in a read only memory (ROM) 902 or a program loaded from a storage section 908 into a random access memory (RAM) 903. The processor 901 can include, for example, a general purpose microprocessor (e.g., a CPU), an instruction set processor, and / or a related chipset and / or a dedicated microprocessor (e.g., an application specific integrated circuit (ASIC)), and the like. The processor 901 can also include an on-board memory for cache use. The processor 901 can include a single processing unit or multiple processing units to perform the various actions of the method processes according to embodiments of the present application.

[0110] In the RAM 903, various programs and data required for the operation of the electronic device 900 are stored. The processor 901, the ROM 902, and the RAM 903 are connected to each other via a bus 904. The processor 901 performs various operations of the method processes according to embodiments of the present application by executing the programs in the ROM 902 and / or the RAM 903. Note that the programs can also be stored in one or more memories other than the ROM 902 and the RAM 903. The processor 901 can also perform various operations of the method processes according to embodiments of the present application by executing the programs stored in the one or more memories.

[0111] According to an embodiment of the present application, the electronic device 900 can also include an input / output (I / O) interface 905 which is also connected to the bus 904. The electronic device 900 can also include one or more of the following components connected to the input / output (I / O) interface 905: an input section 906 including a keyboard, a mouse, etc.; an output section 907 including a display such as a cathode ray tube (CRT), a liquid crystal display (LCD), etc., and a speaker, etc.; a storage section 908 including a hard disk, etc.; and a communication section 909 including a network interface card such as a LAN card, a modem, etc. The communication section 909 performs communication processing via a network such as the Internet. A drive 910 is also connected to the input / output (I / O) interface 905 as necessary. A removable recording medium 911 such as a magnetic disk, a magneto-optical disk, a semiconductor memory, etc. is attached to the drive 910 as necessary so that a computer program read therefrom is installed into the storage section 908 as necessary.

[0112] The application further provides a computer readable storage medium, which can be included in the device / apparatus / system described in the above embodiments, or can exist separately without being assembled into the device / apparatus / system. The computer readable storage medium carries one or more programs, which, when executed, implement the method according to the embodiments of the application.

[0113] According to the embodiments of the application, the computer readable storage medium can be a non-volatile computer readable storage medium, which can include, but is not limited to, a portable computer diskette, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), a portable compact disc read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing. In the present application, the computer readable storage medium can be any tangible medium that contains or stores a program, which can be used by or in connection with an instruction execution system, apparatus, or device. For example, according to the embodiments of the application, the computer readable storage medium can include one or more memories of the ROM 902 and / or the RAM 903 described above and / or one or more memories other than the ROM 902 and the RAM 903.

[0114] Those skilled in the art can understand that the features described in the various embodiments of the application can be combined and / or integrated in various combinations, even if such combinations or integrations are not explicitly described in the application. In particular, the features described in the various embodiments of the application can be combined and / or integrated in various combinations without departing from the spirit and teachings of the application. All these combinations and / or integrations fall within the scope of the application.

Claims

1. A perioperative hypotension probability prediction method, characterized by, The method comprises the following steps: Collecting multi-modal physiological signals and static demographic information of a patient during a perioperative period; Identifying perioperative hypotension event segments and normal event segments in the multi-modal physiological signals based on preset standards, and cutting continuous multi-modal physiological signals into fixed-length signal segments as analysis units; Obtaining a continuous perioperative hypotension probability soft label of each signal segment by weighted summation of time proximity, hemodynamic severity and blood pressure trend; Extracting discrete medical prior features from arterial blood pressure signals in the signal segments; Based on the signal segments and the medical prior features, extracting global frequency domain features through a frequency domain stream and extracting local and global time domain features through a time domain stream by using a dual-flow neural network; Concatenating the frequency domain features, the time domain features and the medical prior features into a fusion feature vector; Designing a loss function based on the distribution difference between the fusion feature vector and the soft label, training the dual-flow neural network by using the fusion feature vector through an adaptive optimization algorithm to obtain a trained dual-flow neural network; and Using the dual-flow neural network to predict a perioperative hypotension probability to obtain a hypotension probability prediction curve.

2. The method of claim 1, wherein, The multi-modal physiological signals comprise at least one of an arterial blood pressure signal, an electrocardiogram signal or a photoplethysmography signal; and The static demographic information comprises at least one of age, gender, height or weight.

3. The method of claim 1, wherein, The time proximity is calculated based on the time difference between a current signal segment and the occurrence of a perioperative hypotension event, and the smaller the time difference is, the higher the time proximity weight is; The hemodynamic severity is calculated based on the mean arterial pressure value of a current signal segment, and the lower the mean arterial pressure is, the higher the hemodynamic severity weight is; The blood pressure trend is calculated based on the mean arterial pressure change slope in a short period before a current signal segment, and the faster the blood pressure decreases, the higher the blood pressure trend weight is.

4. The method of claim 1, wherein, The medical prior features comprise morphological features, statistical features, hemodynamic parameters and time-frequency domain features, The morphological features comprise pulse transit time; The statistical features comprise mean arterial pressure and arterial pressure standard deviation; The hemodynamic parameters comprise cardiac output, stroke volume and peripheral vascular resistance; and The time-frequency domain features comprise the ratio of low-frequency components to high-frequency components of heart rate variability.

5. The method of claim 1, wherein, The global frequency domain features are extracted through a frequency domain stream, which comprises: Performing discrete wavelet transform on each physiological signal channel independently, decomposing the signal into wavelet coefficients of different frequency bands through a filter bank, and then converting the wavelet coefficients into a spectrum block; Assigning a position code to each spectrum block to obtain a spectrum block with a position code; Inputting the spectrum block with the position code into a Transformer encoder to capture long-range dependency between frequency domain features through a self-attention mechanism, and outputting global frequency domain features.

6. The method of claim 1, wherein, The local and global time domain features are extracted through a time domain stream, which comprises: Dividing the time domain signal into overlapping segments, each segment containing a fixed number of sampling points; An arterial blood pressure signal segment is taken as a query, an electrocardiogram and a photoplethysmographic pulse wave signal segment are taken as a key and a value, a feature representation of each small segment is enhanced through a cross-modal attention mechanism, and an enhanced feature is obtained; The enhanced feature is input into a multi-layer one-dimensional convolutional network to extract local and global time domain features.

7. The method of claim 1, wherein, The adaptive optimization algorithm is used to train the double-flow neural network by using the fusion feature vector, including: The fusion feature vector is input into a full connection layer of the double-flow neural network, and high-dimensional features are mapped to a target prediction space through the full connection layer; An activation function is applied to an output of the full connection layer to generate a continuous probability estimation value of a future pre-set time period of occurrence of a perioperative hypotension event; A relative entropy between the probability estimation value and a pre-generated soft label probability distribution is calculated, and the relative entropy is taken as a loss function; and A random gradient descent optimization algorithm is used to iteratively update weight parameters of the double-flow neural network to minimize the loss function.

8. A perioperative hypotension probability prediction apparatus, characterized by, The device comprises: A data acquisition module is configured to acquire multi-modal physiological signals and static demographic information of a patient during a perioperative period; A preprocessing module is configured to identify perioperative hypotension event segments and normal event segments in the multi-modal physiological signals based on a pre-set standard, and cut continuous multi-modal physiological signals into signal segments of a fixed length as analysis units; A first acquisition module is configured to obtain a continuous perioperative hypotension probability soft label of each signal segment by weighted summation of time proximity, hemodynamic severity, and blood pressure trend; A second acquisition module is configured to extract discrete medical prior features from arterial blood pressure signals in the signal segments; A third acquisition module is configured to use a double-flow neural network to extract global frequency domain features through a frequency domain flow and extract local and global time domain features through a time domain flow based on the signal segments and the medical prior features; A feature fusion module is configured to splice the frequency domain features, the time domain features, and the medical prior features into a fusion feature vector; A network training module is configured to design a loss function based on a distribution difference between the fusion feature vector and the soft label, train the double-flow neural network by using the fusion feature vector through an adaptive optimization algorithm, and obtain a trained double-flow neural network; and A probability prediction module is configured to predict a perioperative hypotension probability by using the double-flow neural network to obtain a hypotension probability prediction curve.

9. An electronic device, comprising: Comprise: One or more processors; Memory for storing one or more computer programs, The one or more processors execute the one or more computer programs to implement the steps of the method according to any one of claims 1-7.

10. A computer readable storage medium having stored thereon a computer program or instructions, characterized in that, The computer program or instructions are executed by the processor to implement the steps of the method according to any one of claims 1-7.

Images