Early warning method for infectious shock and storage medium

Abstract

The invention relates to an early warning method for infectious shock and a storage medium. Firstly, the limitation of a single mode is broken through by obtaining multi-mode physiological feature parameters of a user; then, constructing a graph neural network according to the multi-modal physiological feature parameters to obtain spatial correlation features, and obtaining time evolution features based on short-term time sequence features and long-term time sequence features; the method comprises the following steps of: obtaining a time-space feature vector by fusing a space correlation feature and a time evolution feature, outputting a risk probability and a time sensitivity index of the infectious shock of a user through a two-channel mechanism, and predicting the risk of the infectious shock from two dimensions of time and space, so that the risk level of the user can be more accurately predicted, and the risk prediction accuracy is improved. And the timeliness of risk early warning is improved, and timely intervention on the infectious shock of the user is realized, so that the treatment effect is improved.

Description

Technical Field

[0001] The present application relates to the field of medical information processing technology, and in particular to an early warning method and storage medium for septic shock. Background Art

[0002] Septic shock, also known as septic shock, is a severe stage of sepsis. It refers to a clinical syndrome characterized by an uncontrolled systemic inflammatory response triggered by infection, leading to circulatory decline and organ dysfunction. Without timely intervention, it can easily lead to multi-organ failure and be life-threatening. Currently, the diagnosis and early warning of septic shock mainly rely on traditional static scoring systems (such as SOFA and qSOFA scores). However, these scoring systems have the following limitations. Traditional static scoring methods (such as MAP <65 mmHg) have difficulty reflecting the dynamic changes of physiological parameters in real time and cannot fully consider the complex nonlinear relationships between multiple monitoring indicators. Furthermore, most prediction models rely on single-modality data and lack comprehensive data analysis and fusion. Therefore, traditional early warning methods are not accurate enough in predicting septic shock, which can easily lead to missed opportunities for clinical intervention and affect treatment efficacy. Summary of the Invention

[0003] The purpose of this application is to provide an early warning method for septic shock to solve the problem of insufficient timeliness of traditional early warning of septic shock.

[0004] In order to achieve the above objectives, the present application provides, in a first aspect, an early warning method for septic shock, comprising: Acquiring multimodal physiological characteristic parameters of the user, wherein the multimodal physiological characteristic parameters include hemodynamic monitoring parameters, microcirculation monitoring parameters, and laboratory parameters, and the physiological characteristic parameters of each modality include at least one indicator parameter; Constructing a graph neural network based on the multimodal physiological characteristic parameters to obtain spatial correlation features corresponding to the multimodal physiological characteristic parameters; Extracting short-term time series features and long-term time series features of the multimodal physiological characteristic parameters based on the graph neural network, and obtaining time evolution features corresponding to the multimodal physiological characteristic parameters based on the short-term time series features and the long-term time series features; Fusing the spatial correlation feature and the temporal evolution feature to obtain a spatiotemporal feature vector corresponding to the physiological feature parameter; Based on the spatiotemporal feature vector, a risk probability and a time sensitivity index of septic shock of the user are obtained, and based on the risk probability and the time sensitivity index, a warning level of septic shock of the user is determined.

[0005] Optionally, each physiological characteristic parameter includes a physiological characteristic index and an index parameter value corresponding to the physiological characteristic index, and obtaining the multimodal physiological characteristic parameters of the user includes: According to the sampling method set for each modality, the indicator parameter values corresponding to the physiological characteristic indicators of each modality are collected respectively to obtain the initial indicator parameter values of the multimodality; Time-aligning the initial indicator parameter values of the multimodal states to obtain a time-synchronized multimodal physiological feature sequence; The multimodal physiological feature sequence is sliced according to a set time interval to obtain index parameter values of multiple time slices, each of which corresponds to a feature index.

[0006] Optionally, constructing a graph neural network based on the multimodal physiological characteristic parameters to obtain spatial correlation features corresponding to the multimodal physiological characteristic parameters includes: Determine the characteristic indicators in each time slice as a node, and determine the indicator parameter values corresponding to the plurality of characteristic indicators as node attributes corresponding to the node; Determining associations between different feature indicators as edges, and determining edge weights corresponding to the edges based on an adjacency matrix of the edges; A graph neural network constructed based on the nodes, the node attributes, the edges, and the edge weights; Message passing is performed through the graph neural network, and edge weights are used to perform weighted aggregation on the node attributes of each node based on the node attributes of its neighboring nodes to obtain a feature vector for each node, wherein the feature vector of each node includes spatial correlation features between the node and its neighboring nodes.

[0007] Optionally, determining the indicator parameter values corresponding to the plurality of characteristic indicators as the node attributes corresponding to the node includes: For each time slice, obtaining characteristic indicators in the current time slice and the previous time slice and indicator parameter values corresponding to the characteristic indicators; determining a rate of change of the characteristic indicator according to the current time slice and the previous time slice; Determining a standard deviation of the characteristic indicator according to the indicator parameter value in the current time slice, wherein the standard deviation reflects the volatility of the characteristic indicator in the current time slice; The node attribute of the current time slice is obtained according to the indicator parameter value of the current time slice, the change rate of the characteristic indicator, and the standard deviation of the characteristic indicator.

[0008] Optionally, determining the association relationship between different characteristic indicators as an edge, and determining the edge weight corresponding to the edge based on the adjacency matrix of the edge includes: For each time slice, obtain the information entropy of the two nodes connected by each edge and the mutual information entropy between the two nodes; Obtaining a length of a current time slice, and determining an attenuation factor based on the length of the current time slice; Based on the information entropy, the mutual information entropy, and the attenuation factor, an adjacency matrix of the edge is determined, and a value of the adjacency matrix is determined as an edge weight corresponding to the edge.

[0009] Optionally, extracting short-term time series features and long-term time series features of the multimodal physiological characteristic parameters based on the graph neural network includes: Obtaining a multimodal indicator time series based on physiological characteristic parameters in a current time slice and a set number of time slices before the current time slice; Inputting the indicator time series into the dilated causal convolutional network to calculate the short-term time series characteristics of the physiological characteristic parameters; The indicator time series is input into a converter model to calculate the long-term time series characteristics of the physiological characteristic parameters.

[0010] Optionally, obtaining the time evolution characteristics corresponding to the multimodal physiological characteristic parameters based on the short-term time series characteristics and the long-term time series characteristics includes: Performing feature splicing on the short-term time series features and the long-term time series features, and performing linear layer transformation to obtain the time evolution features corresponding to the physiological feature parameters; or Inputting the short-term time series features and the long-term time series features into a first target gating network, and obtaining weights of the short-term time series features and the long-term time series features through dynamic calculation; Based on the weights of the short-term time series features and the long-term time series features, the short-term time series features and the long-term time series features are weightedly fused to obtain the time evolution features corresponding to the physiological characteristic parameters.

[0011] Optionally, fusing the spatial correlation feature and the temporal evolution feature to obtain the spatiotemporal feature vector corresponding to the physiological feature parameter includes: Inputting the spatial correlation feature and the temporal evolution feature into a second target gating network, and obtaining weights of the spatial correlation feature and the temporal evolution feature through dynamic calculation; Based on the weights of the spatial correlation feature and the temporal evolution feature, the spatial correlation feature and the temporal evolution feature are weightedly fused to obtain a spatiotemporal feature vector corresponding to the physiological feature parameter.

[0012] Optionally, obtaining a risk probability and a time sensitivity index of septic shock of the user based on the spatiotemporal feature vector, and determining a warning level of septic shock of the user based on the risk probability and the time sensitivity index, includes: Inputting the spatiotemporal feature vector into a logistic regression model to obtain the risk probability of septic shock of the user; Inputting the spatiotemporal feature vector into a lightweight long short-term memory network to obtain a time sensitivity index of septic shock of the user; Normalizing the risk probability and the time sensitivity index, and obtaining a set time sensitivity compression coefficient; Obtaining a comprehensive risk score according to the normalized risk probability, the time sensitivity index, and the sensitivity compression coefficient; A warning level of septic shock of the user is determined based on the comprehensive risk score.

[0013] A second aspect of the present application provides a computer-readable storage medium, wherein the computer-readable storage medium stores a program, and the program can be loaded by a processor to execute the above-mentioned early warning method for septic shock.

[0014] The beneficial effects of this application are: This application first obtains the user's multimodal physiological characteristic parameters, integrates hemodynamic monitoring data, microcirculation monitoring parameters and laboratory parameters, breaks through the limitations of a single modality, and refines the granularity of the indicator parameters of each modality to comprehensively capture the user's physiological characteristic parameters. Then, based on the graph neural network, the spatial correlation characteristics of the multimodal physiological characteristic parameters are constructed, the spatial dependence of the multimodal physiological characteristic parameters is quantified, and the time evolution characteristics corresponding to the multimodal physiological characteristic parameters are obtained by combining the short-term time series characteristics and the long-term time series characteristics, thereby reducing the risk lag judgment caused by the single time dimension analysis. Then, by fusing the spatial correlation characteristics and the time evolution characteristics into a spatiotemporal feature vector, and outputting the risk probability and time sensitivity index of the user's infectious shock through a dual-channel mechanism, the risk prediction of infectious shock is performed from the dual dimensions of time and space, which can more accurately predict the user's risk level, improve the timeliness of the risk warning, and achieve timely intervention in the user's infectious shock, thereby improving the treatment effect.

[0015] Other features and advantages of the present application will be described in detail in the subsequent detailed description. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 This is a schematic diagram of an application scenario of a septic shock early warning method provided in an embodiment of the present application; Figure 2This is a flow chart of a method for early warning of septic shock provided in an embodiment of the present application; Figure 3 A schematic diagram of a graph neural network provided in an embodiment of the present application. DETAILED DESCRIPTION

[0017] The following will be combined with the drawings in the embodiments of this application to clearly and completely describe the technical solutions in the embodiments of this application. Obviously, the embodiments described are only part of the embodiments of this application, not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without making creative efforts are within the scope of protection of this application.

[0018] In the description of this application, it should be understood that the terms "first" and "second" are used for descriptive purposes only and are not to be construed as indicating or implying relative importance or implicitly specifying the number of the technical features indicated. Therefore, features specified as "first" or "second" may explicitly or implicitly include one or more of the described features. In the description of this application, "plurality" means two or more, unless otherwise specifically qualified. In this application, the word "exemplary" is used to mean "serving as an example, illustration, or illustration." Any embodiment described in this application as "exemplary" is not necessarily to be construed as preferred or advantageous over other embodiments. The following description is provided to enable anyone skilled in the art to implement and use the present application. In the following description, details are listed for illustrative purposes. It should be understood that one of ordinary skill in the art will recognize that the present application can be implemented without these specific details. In other instances, well-known structures and processes are not described in detail to avoid obscuring the description of this application with unnecessary detail. Therefore, this application is not intended to be limited to the embodiments shown, but is to be accorded the widest scope consistent with the principles and features disclosed herein.

[0019] In an embodiment of the present application, a warning method for septic shock is applied to electronic equipment. Figure 1 As shown, Figure 1 This is a schematic diagram of an application scenario of the early warning method for septic shock in an embodiment of the present application. The application scenario of the early warning method for septic shock in an embodiment of the present application includes an electronic device 110 for the early warning method for septic shock. The electronic device 110 is integrated with a computer-readable storage medium corresponding to the early warning method for septic shock to execute the steps of the early warning method for septic shock.

[0020] It is understandable that Figure 1The electronic devices in the application scenario of the early warning method for septic shock shown, or the devices contained in the electronic devices, do not constitute a limitation on the embodiments of the present application. That is, the number and types of devices contained in the application scenario of the early warning method for septic shock, or the number and types of devices contained in each device, do not affect the overall implementation of the technical solutions in the embodiments of the present application, and can all be regarded as equivalent replacements or derivatives of the technical solutions claimed in the embodiments of the present application.

[0021] In the embodiments of the present application, the electronic device 110 may be a standalone device or a device network or device cluster. For example, the electronic device 110 described in the embodiments of the present application includes, but is not limited to, a computer, a network host, a single network device, a collection of multiple network devices, or a cloud device composed of multiple devices. A cloud device is composed of a large number of computers or network devices based on cloud computing.

[0022] Those skilled in the art will understand that Figure 1 The application scenario shown in the figure is only one application scenario corresponding to the technical solution of the present application and does not constitute a limitation on the application scenario of the technical solution of the present application. Other application scenarios may also include Figure 1 More or fewer electronic devices shown in, or electronic device network connection relationships, such as Figure 1 Only one electronic device is shown. It is understandable that the scenario of the early warning method for septic shock may also include one or more other electronic devices, which are not specifically limited here. The electronic device 110 may also include a memory and a processor, and the memory is used to store information related to the early warning method for septic shock.

[0023] Furthermore, in the application scenario of the septic shock early warning method in the embodiments of the present application, the electronic device 110 may be provided with a display device, or the electronic device 110 may not be provided with a display device and may be in communication with an external display device 120. The display device 120 is used to output the results of the execution of the septic shock early warning method in the electronic device. The electronic device 110 may access a backend database 130. The backend database 130 may be a local memory of the electronic device 110 or a cloud database located in the cloud. The backend database 130 stores information related to the septic shock early warning method.

[0024] It should be noted that Figure 1 The application scenario of the early warning method for septic shock shown is merely an example. The application scenario of the early warning method for septic shock described in the embodiment of the present application is intended to more clearly illustrate the technical solution of the embodiment of the present application and does not constitute a limitation on the technical solution provided in the embodiment of the present application.

[0025] Based on the application scenario of the above-mentioned early warning method for septic shock, an embodiment of the early warning method for septic shock is proposed, which is described in detail below with reference to the accompanying drawings.

[0026] Figure 2 Schematic diagram of a method for early warning of septic shock provided in the embodiment of the present application. Figure 2 As shown, the early warning method can execute steps 201-205 and other steps through the processor in the above-mentioned electronic device 110, which will be described in detail below.

[0027] Step 201: Acquire multimodal physiological characteristic parameters of the user.

[0028] Multimodal physiological characteristic parameters refer to a set of parameters that reflect the physiological state and function of the human body, obtained from multiple different physiological information or measurement methods. These multimodal physiological characteristic parameters may include hemodynamic monitoring parameters, microcirculatory monitoring parameters, and laboratory parameters. Each modality's physiological characteristic parameters include at least one indicator parameter. The indicator parameter is the specific characteristic information used to measure and evaluate the human physiological state within each modality.

[0029] Hemodynamic monitoring parameters refer to a series of indicators used to reflect cardiac and vascular function and blood circulation status. These are macroscopic monitoring parameters that assess cardiac pumping function, vascular resistance, and blood volume. For example, these parameters may include, but are not limited to, mean arterial pressure (MAP), heart rate (HR), cardiac output (CO), cardiac index (CI), partial pressure of carbon dioxide (Venous-Arterial) (Pv-aCO2gap), systemic vascular resistance index (SVRI), central venous oxygen saturation (ScvO2), and the proportion of functional microvessels (PPV). In one example, these parameters can be collected using noninvasive monitoring devices. For example, hemodynamic parameters can be automatically collected hourly for emergency or critically ill patients.

[0030] Microcirculation monitoring parameters are used to assess the blood circulation status of microvessels, such as arterioles, capillaries, and venules. These are microscopic monitoring parameters that can directly reflect tissue and cell perfusion and are of great significance for early tissue hypoxia and tracheal dysfunction. For example, microcirculation monitoring parameters may include, but are not limited to, capillary flow index (MFI), tissue oxygenation index (StO2), tissue partial pressure of carbon dioxide (PtCO2), and capillary refill time (CRT). In one example, microcirculation monitoring parameters can be collected by a microcirculation monitoring instrument, for example, automatically collecting relevant indicators of a user's microcirculation every six hours.

[0031] Laboratory parameters refer to data obtained from laboratory tests on samples such as blood, urine, and cerebrospinal fluid, which can reflect the body's metabolic, immune, and coagulation functions. Laboratory parameters can include various reported data, such as lactate (Lac) and oxygenation indices such as arterial partial pressure of oxygen / fraction of inspired oxygen (PaO2 / FiO2). In one example, laboratory test reports can be uploaded to a database, from which relevant indicator parameters of laboratory parameters can be retrieved.

[0032] Step 202: Build a Graph Neural Network (GNN) based on the multimodal physiological characteristic parameters to obtain spatial correlation features corresponding to the multimodal physiological characteristic parameters.

[0033] GNNs are neural network models used to process graph-structured data. GNNs are composed of nodes and edges, where nodes represent entities and edges represent relationships between entities. In the embodiments of this application, the indicator parameters included in the multimodal physiological characteristic parameters are nodes in the GNN, and the relationships between the indicator parameters are edges in the GNN.

[0034] The GNN in the embodiment of the present application is a dynamic GNN, which can process the changes of each node, edge and node feature in the GNN over time or other dynamic conditions according to the dynamically changing physiological characteristic parameters. As an example, the GNN updates the feature representation of the node through information transmission and aggregation between nodes, and obtains the spatial correlation features between nodes by aggregating the information of the node and its neighboring nodes, using graph convolution or attention mechanism, etc. Spatial correlation features refer to the feature representation formed by the mutual relationship and dependency characteristics between the nodes mined in the spatial dimension in the graph structure data. In other words, by constructing a GNN based on the indicator parameters contained in the multimodal physiological characteristic parameters, the spatial correlation features between the indicator parameters can be obtained.

[0035] Step 203: Based on the graph neural network, extract the short-term time series features and long-term time series features of the multimodal physiological characteristic parameters, and obtain the time evolution features corresponding to the multimodal physiological characteristic parameters based on the short-term time series features and the long-term time series features.

[0036] In order to comprehensively capture the dynamic changing trends of physiological characteristic parameters of patients with septic shock over time, on the basis of dynamically establishing a graph neural network, a cross-scale time series modeling mechanism is introduced to extract the short-term and long-term time series features of the physiological characteristic parameters respectively, and fuse them to obtain the time evolution characteristics corresponding to the multimodal physiological characteristic parameters, so as to obtain the perception ability of pathological evolution patterns at different time scales.

[0037] Short-term time series features describe the changing patterns and patterns of multimodal physiological parameters over a relatively short timeframe, focusing primarily on fluctuations and trends within these parameters within a short time window. Extracting short-term time series features can capture rapid changes in physiological parameters, helping to promptly detect sudden changes or abnormalities in physiological status. Long-term time series features describe the changing trends and patterns of multimodal physiological parameters over a longer timeframe, reflecting chronic changes in the human physiological state, the progression of disease, and the body's long-term response to treatment. Temporal evolution features are a feature representation derived from the integration of short-term and long-term time series features. They not only include information on changes in physiological parameters at different timescales, but also consider the interrelationships and dynamic evolution between short-term and long-term time series features. This allows for more accurate calculation of the warning level for septic shock.

[0038] Step 204: Fuse the spatial correlation features and the temporal evolution features to obtain a spatiotemporal feature vector corresponding to the physiological feature parameters.

[0039] In order to accurately identify the early dynamic risks of septic shock, the embodiment of the present application also introduces a spatiotemporal fusion mechanism to fuse the spatial correlation features and temporal evolution features of physiological characteristic parameters to obtain a two-dimensional comprehensive representation including time and space, namely the spatiotemporal feature vector, thereby providing multi-dimensional feature input for the subsequent calculation of the warning level.

[0040] Step 205: Based on the spatiotemporal feature vector, obtain the risk probability and time sensitivity index of the user's septic shock, and determine the warning level of the user's septic shock based on the risk probability and time sensitivity index.

[0041] In the embodiment of the present application, the risk probability of septic shock refers to the size of the possibility of the user developing septic shock. The time sensitivity index refers to an indicator used to measure the time urgency of the user developing septic shock, reflecting the relative length of time remaining from the current moment to the time when the user's septic shock may deteriorate significantly or enter an irreversible stage. The higher the time sensitivity index, the shorter the remaining effective intervention time. Conversely, the lower the time sensitivity index, the more time there is for operations such as observing the condition and specifying a plan. A comprehensive score can be calculated based on the risk probability and time sensitivity, and then the warning level of the user's septic shock can be obtained based on the correspondence between the comprehensive score and the warning level. For example, the higher the warning level, the earlier and faster the user needs to be intervened and treated. The embodiment of the present application quantifies the possibility and urgency of the risk through a dual-channel output design, breaks through the limitations of a single indicator, and improves the assessment integrity of the risk warning.

[0042] The embodiment of the present application breaks through the limitation of a single modality, and refines the granularity of the indicator parameters of each modality to fully capture the physiological characteristic parameters of the user. At the same time, the spatial correlation characteristics of the multimodal physiological characteristic parameters are constructed based on the graph neural network, the spatial dependence of the multimodal physiological characteristic parameters is quantified, and the time evolution characteristics corresponding to the multimodal physiological characteristic parameters are obtained by combining the short-term time series characteristics and the long-term time series characteristics, thereby reducing the risk lag judgment caused by the single time dimension analysis. Then, by fusing the spatial correlation characteristics and the time evolution characteristics into a spatiotemporal feature vector, and outputting the risk probability and time sensitivity index of the user's infectious shock through a dual-channel mechanism, the risk prediction of infectious shock is carried out from the dual dimensions of time and space, which can more accurately predict the user's risk level, improve the timeliness of the risk warning, and achieve timely intervention in the user's infectious shock, thereby improving the treatment effect.

[0043] Since the GNN in the embodiment of the present application is a dynamic GNN, the computational complexity of the GNN can be reduced and the efficiency of the GNN data processing can be improved by slicing the dynamically changing time. In the embodiment of the present application, each physiological characteristic parameter includes a physiological characteristic index and an indicator parameter value corresponding to the physiological characteristic index. A physiological characteristic index refers to a quantitative description used to measure and reflect the physiological state of the human body, and each physiological characteristic index represents a specific aspect of the physiological function of the human body. Each physiological characteristic index has a corresponding parameter value to quantitatively represent the state of the physiological characteristic index, which can be collected by a device that collects parameter values.

[0044] In step 201, the index parameter values corresponding to the physiological characteristic indicators of each modality can be collected respectively according to the sampling method set for each modality to obtain the initial index parameter values of the multimodality. Then, the initial index parameter values of the multimodality are time-aligned to obtain a time-synchronized multimodal physiological characteristic sequence. Finally, the multimodal physiological characteristic sequence is sliced according to the set time interval to obtain the index parameter values of multiple time slices, and each index parameter value corresponds to a characteristic indicator. Each time slice can include data obtained at multiple moments. Therefore, in each time slice, the median or average can be taken as the index parameter value in the time slice.

[0045] In the embodiments of the present application, the initial indicator parameter value refers to the original parameter value for sampling the physiological characteristic parameters of each modality. The sampling method set for each modality may be different. For example, the hemodynamic monitoring parameters can be set to be collected once every 1 hour, and the microcirculation monitoring parameters can be set to be collected once every 6 hours. Therefore, it is necessary to time-align the indicator parameter values of the physiological characteristic parameters with different sampling frequencies. For example, using methods such as linear interpolation or strip interpolation, the multi-modal physiological characteristic parameters are unified on the time axis to obtain a time-synchronized multi-modal physiological characteristic sequence. Then, according to the set time interval, the physiological characteristic sequence is segmented, wherein the set time interval can be pre-set as needed, for example, set to 2 hours as a time slice, etc. In this way, the indicator parameter values of multiple time slices can be obtained, and each indicator parameter value corresponds to a characteristic indicator. Through standardized sampling, precise alignment and structured segmentation, the multi-modal physiological characteristic parameters with different acquisition frequencies can be converted into time series features in a unified format.

[0046] For each time slice, a GNN within the time slice can be constructed to obtain the spatial correlation characteristics of each feature indicator within the time slice. Specifically, in step 202, the feature indicators within each time slice can be determined as nodes, and the indicator parameter values corresponding to the multiple feature indicators can be determined as node attributes corresponding to the nodes. Then, the association relationship between different feature indicators is determined as an edge, and the edge weight corresponding to the edge is determined based on the adjacency matrix of the edge. The association relationship refers to the dynamic dependency relationship between different feature indicators. Next, a graph neural network is constructed based on nodes, node attributes, edges and edge weights. Finally, message passing is performed through the graph neural network, and the edge weights are used to perform weighted aggregation on the node attributes of each node based on the node attributes of the neighboring nodes of each node to obtain the feature vector of each node. Among them, the feature vector of each node includes the spatial correlation characteristics of the node and the neighboring nodes.

[0047] Figure 3 This is a schematic diagram of a graph neural network provided in the embodiment of this application. Figure 3 As shown, the expression of GNN can be: ; in, for GNN expression of time slice, Represents a node set, that is, multiple feature indicators in A collection of time slices, such as to , is an edge set, that is, the edges between different feature indicators The association relationship of time slices. is a set of edge weights, that is, the relationship between multiple feature indicators is The correlation strength of the time slice.

[0048] In an embodiment of the present application, the calculation of node attributes can be obtained in the following manner. For each time slice, first, obtain the characteristic index in the current time slice and the previous time slice and the index parameter value corresponding to the characteristic index. Then, determine the rate of change of the characteristic index based on the current time slice and the previous time slice. Determine the standard deviation of the characteristic index based on the index parameter value in the current time slice. The standard deviation reflects the volatility of the characteristic index in the current time slice. Finally, based on the index parameter value of the current time slice, the rate of change of the characteristic index and the standard deviation of the characteristic index, obtain the node attributes of the current time slice. The node attributes are the values of each node in the The state vector of the time slice.

[0049] As an example, node attributes can be calculated using the following formula: ; in, for Node attributes of the time slice, for Time slice The index parameter value of the node is Time slice The rate of change of the node, , It is the indicator parameter value of the previous time slice. for Time slice The standard deviation of the node's characteristic metric.

[0050] In an embodiment of the present application, the calculation of edge weights can quantify the strength of the relationship through time-decayed mutual information entropy. Specifically, for each time slice, the information entropy of the two nodes connected by each edge and the mutual information entropy between the two nodes are first obtained. Information entropy is a concept used to measure the amount of information. The greater the uncertainty, the higher the information entropy. Mutual information entropy is an indicator to measure the degree of mutual dependence between two nodes. The larger the mutual information entropy, the stronger the correlation between the two nodes. Conversely, the smaller the mutual information entropy, the smaller the correlation between the two nodes. Then, the length of the current time slice is obtained, and the attenuation factor is determined based on the length of the current time slice. The attenuation factor can represent the degree of attenuation of the edge weight over time. Finally, based on the information entropy, mutual information entropy and attenuation factor, the adjacency matrix of the edge is determined. The adjacency matrix is used to represent the connection relationship between nodes, and the value of the adjacency matrix can be further expanded to the edge weight corresponding to the edge.

[0051] As an example, the edge weight can be calculated using the following formula: ; in, is the edge weight, Time slice The mutual information entropy within can be used to measure the parameters and Dynamic association relationships, for example, It can be 3 hours. and For nodes and The information entropy can reflect the uncertainty of the corresponding indicator parameter value of the node. is the time decay coefficient, which can control the weight decay rate of the historical indicator parameter value, for example, It can be 4 hours.

[0052] In one example, assuming the collected MAP is 58 mmHg and lactate is 4.8 mmol / L, the mutual information entropy within 3 hours is , time decay term , the edge weight can be calculated The calculation results show that the correlation strength between MAP and lactate increased by 3 times, suggesting the synergistic worsening of circulatory failure and tissue hypoxia.

[0053] MAP and lactate were chosen as examples for edge construction because they represent the core indicators of macrocirculatory and metabolic functions, respectively. Their dynamic coupling changes can serve as a key basis for judging the progression of septic shock and intervention response, and have strong temporal causal characteristics and high clinical credibility. It should be noted that in addition to the edge construction example between MAP and lactate, the present invention also supports the construction of dynamic edges between other indicator pairs in graph neural networks, such as MAP–HR, CO–SVRI, MFI–ScvO2, CRT–Lac, etc. The selection of the above edges is based on clinical pathological connections and temporal mutual information calculations, which can comprehensively cover the multiple stages of septic shock from macrohemodynamic disorders to microcirculatory disorders, and then to the deterioration of metabolic function, thereby enhancing the expressiveness and risk identification capabilities of the dynamic graph structure.

[0054] In step 203, a hybrid architecture of a dilated causal convolutional network (TCN) and a transformer model can be used, combined with multi-scale convolution and position encoding, to extract short-term temporal features and long-term temporal features of physiological characteristic parameters. Specifically, a multimodal indicator time series can be obtained based on the physiological characteristic parameters in the current time slice and a set number of time slices before the current time slice. The indicator time series refers to an indicator sequence with temporal characteristics constructed based on the multimodal physiological characteristic parameters of the current time slice and a period of time in the past. For example, assuming that the multimodal physiological characteristic parameters within the previous 6 hours are extracted and the sampling interval is set to 10 minutes, there are a total of 36 time slices. Assuming that each time slice contains 12 physiological indicator dimensions (such as MAP, HR, Lac, ScvO2, MFI, etc.), the indicator time series is finally formed, that is, the input tensor : .

[0055] Then, the indicator time series is input into the TCN to calculate the short-term time series features of the physiological characteristic parameters. The indicator time series is input into the Transformer model to calculate the long-term time series features of the physiological characteristic parameters.

[0056] TCNs are a method for extracting data features. However, unlike the fully connected operations in traditional convolutional neural networks (CNNs), TCNs employ causal convolutions, meaning that the computation at each time step depends only on the current and previous time steps, preserving the causal relationships of time series data. Dilated convolutions use different dilation rates (such as 1, 2, 4, and 8) to capture features at different time scales. For example, d=1 captures short-term changes (sharp fluctuations within a 10-minute period), such as a sudden increase in heart rate; d=2 captures trends over slightly longer periods (20 minutes); and d=4 captures physiological adjustments over longer periods (40 minutes). This allows TCNs to respond in real time to rapid changes in a patient's physiological state, such as a sharp increase in heart rate or a sudden drop in MAP. For example, suppose patient Zhang San's heart rate (HR) increased from 110 bpm to 130 bpm over the past 30 minutes, while his MAP dropped sharply from 70 mmHg to 58 mmHg during the same period. TCN captures such sudden fluctuations through dilated convolution and strengthens the weights of these fluctuations, allowing the system to respond to these sudden changes in a timely manner.

[0057] The Transformer model is primarily designed to capture long-term trends—physiological changes and potential risks—over extended periods of time (e.g., more than two hours). The Transformer model can "look" at the entire input sequence, calculate the relative importance of each time point, and weight them accordingly. The Transformer model extracts long-term temporal features through a self-attention mechanism and a multi-head attention matrix. The self-attention matrix calculates the similarity between each time step and other time steps (i.e., "attention weights") and then uses these weights to weightedly average the data to obtain global information. The multi-head attention mechanism uses multiple independent "attention heads" to process different features in parallel, helping the model learn a variety of complex temporal relationships. For example, if a person's lactate (Lac) level has gradually increased over the past six hours, from 3.5 mmol / L to 4.8 mmol / L, the Transformer model can identify this slowly increasing trend and capture this gradually accumulating long-term risk.

[0058] After the TCN and Transformer models, two outputs can be obtained: as well as . The short-term time series features extracted from TCN capture rapidly changing patterns, for example, reflecting acute fluctuations in the past 30 minutes to 1 hour. The long-term time series features extracted from the Transformer model capture potential trend changes, for example, reflecting the slowly accumulated risk trend in the past 6 hours.

[0059] Short-term time series features and long-term time series features can be fused through splicing or gating mechanisms to obtain time evolution features.

[0060] In one example, the short-term time series features and the long-term time series features may be concatenated and transformed through a linear layer to obtain the time evolution features corresponding to the physiological feature parameters.

[0061] In another example, the short-term time series features and the long-term time series features can be first input into the first target gating network. The first target gating network is a gating network for fusing the short-term time series features and the long-term time series features. The weights of the short-term time series features and the long-term time series features can be obtained through dynamic calculation. Then, based on the weights of the short-term time series features and the long-term time series features, the short-term time series features and the long-term time series features are weightedly fused to obtain the time evolution characteristics corresponding to the physiological characteristic parameters.

[0062] The expression of the time evolution characteristics can be calculated by the following formula: ; in, is the time evolution characteristic, is the short-term time series characteristic, is the long-term time series characteristic, Generate gating weights for the Sigmoid function (range 0-1), is the first gating weight matrix, is the feature splicing operation, The product of the elements achieves weighted control. This formula can combine short-term and long-term time series features to provide more accurate early warning results.

[0063] In one example, suppose that patient Zhang San's heart rate (HR) rises rapidly from 110 bpm to 130 bpm over the past 30 minutes, and MAP drops sharply from 70 mmHg to 58 mmHg. At the same time, the pulse variability (PPV) increases significantly. The TCN identifies the sudden hypoperfusion manifestation and outputs a highly responsive short-term time series feature. At the same time, the Transformer model observed that Zhang San's lactate (Lac) gradually increased over the past 6 hours, from 3.5→3.8→4.2→4.8mmol / L, showing a clear "metabolic deterioration trend", thus generating a long-term time series feature representing the accumulation of chronic risk. The second gating network can adaptively generate weights based on the input, for example: .

[0064] The above expression indicates that the currently determined emergency state is more urgent, with short-term time series features accounting for 64% and long-term time series features accounting for 36%. The resulting fused time evolution features can be used for subsequent risk warning calculations.

[0065] In the embodiment of the present application, the spatial correlation features and the temporal evolution features can also be fused through a gating network to obtain the spatiotemporal feature vector corresponding to the physiological feature parameters.

[0066] Specifically, the spatial correlation features and temporal evolution features can be input into a second target gating network. The second target gating network is a gating network used to fuse the spatial correlation features and temporal evolution features. First, the weights of the spatial correlation features and temporal evolution features are obtained through dynamic calculation. Then, based on the weights of the spatial correlation features and temporal evolution features, the spatial correlation features and temporal evolution features are weightedly fused to obtain the spatiotemporal feature vector corresponding to the physiological characteristic parameters.

[0067] In the embodiment of the present application, the spatial correlation features output by GNN reflect the coupling status between various physiological parameters, and the time evolution features reflect the dynamic evolution of indicators over time. The spatial correlation features and the time evolution features are then fused to generate a unified risk expression vector.

[0068] As an example, the spatiotemporal eigenvector can be calculated using the following formula: ; in, is the space-time eigenvector, is the spatial correlation feature, is the time evolution characteristic, Generate gating weights for the Sigmoid function (range 0-1), is the second gating weight matrix.

[0069] It should be noted that this fusion mechanism allows the model to dynamically adjust the importance weights of spatial structure information and temporal trend information according to the current situation.

[0070] In an example, suppose patient Zhang San is calculated based on the node association in the past 3 hours: The mutual information entropy between MAP and Lac was 0.33, with an edge weight of 0.33 × e-1 / 4 ≈ 0.257, indicating a significant synergistic decline in perfusion and metabolism. The edge weight between MFI and ScvO2 was 0.29 × e-0.5 / 4 ≈ 0.255, suggesting a simultaneous deterioration of microcirculatory disturbances and insufficient oxygen supply.

[0071] The spatial correlation characteristics of the GNN output are: Zhang San's MAP-lactic acid edge weight increases rapidly, and MFI-ScvO2 decreases synergistically. GNN detects that the coupling between multiple indicator parameter values is enhanced, and the spatial risk is high.

[0072] TCN analysis showed that the MAP decrease slope was -2.1 mmHg / h and the HR increase slope was +4.3 bpm / 10 min in the past 30 minutes, indicating short-term instability.

[0073] The Transformer model identified that lactate (Lac) showed a step-by-step increase trend within 6 hours: 3.5→3.8→4.2→4.8mmol / L, and predicted that it would continue to rise in the next 2 hours.

[0074] The output of the time evolution features is TCN identifying the intensified fluctuations in heart rate and MAP, and Transformer identifying the continuous increase in lactate.

[0075] The second target gating network automatically learns the fusion weights based on the input and obtains a second gating weight of 0.52. This indicates that the current state has both acute fluctuations and trend accumulation. Further analysis shows that the fusion ratios of spatial correlation features and temporal evolution features are 52% and 48% respectively. The final fusion output spatiotemporal feature vector is: .

[0076] Then, the warning level of septic shock can be predicted through a dual-channel output mechanism. Specifically, the spatiotemporal feature vector can be input into a logistic regression model to obtain the user's risk probability of septic shock.

[0077] As an example, the risk probability can be obtained by the following formula: ; in, is the user's risk probability of septic shock (ranging between 0 and 1) as a spatiotemporal feature vector. is the space-time feature vector. and The weights for the trained logistic regression model are shown in Figure 2. Because the initial calculations contradict the expected high risk, the rationality of the weights needs to be checked. Therefore, the weights for the logistic regression model can be obtained through training.

[0078] Then, the spatiotemporal feature vector is input into a lightweight long short-term memory (LSTM) network to obtain the user's time sensitivity index of infectious shock, that is, to evaluate the urgency.

[0079] As an example, the time sensitivity index can be obtained by the following formula: ; in, is the time sensitivity index, is the space-time eigenvector, is a historical spatiotemporal feature vector, for example, the spatiotemporal feature vector of the past 4 hours. The output time sensitivity index is the estimated effective intervention time window, indicating the time within which intervention can effectively block the development of septic shock.

[0080] In one example, assuming that the spatiotemporal feature vectors in the past three hours are [-0.2, 1.1], [0.3, 1.8], [-0.6, 2.5], the output time sensitivity index is 5.2 hours, indicating that intervention is required within 5.2 hours.

[0081] The risk probability and time sensitivity index output from both channels are then normalized, converting them to a specific range for unified processing and comparison. A set time sensitivity compression factor is then obtained. The time sensitivity compression factor is a parameter used to compress the time sensitivity index. This compression allows for further adjustment of the normalized time sensitivity index to better suit subsequent calculation requirements.

[0082] Next, based on the normalized risk probability, time sensitivity index, and sensitivity compression coefficient, a time weight is introduced to obtain a comprehensive risk score. As an example, the comprehensive risk score can be calculated using the following formula: ; in, represents the comprehensive risk score, It can be determined as a high-urgency severe illness. This indicator is used to make decisions on hierarchical management and priority treatment. is the time sensitivity compression coefficient. The smaller the value, the higher the weight of the time factor. Since the clinical guidelines recommend that the golden rescue time is 6 hours, it is preferred that It can be 6. , which can compress time urgency into a reasonable range and reduce the interference of extreme values.

[0083] Finally, the warning level of the user's septic shock is determined based on the comprehensive risk score. In an embodiment of the present application, the comprehensive risk score can be mapped to the warning level, and each warning level corresponds to a comprehensive risk score in an interval. For example, the risk level is divided into green warning, yellow warning and red warning, the level of green warning is lower than that of yellow warning, and the level of yellow warning is lower than that of red warning. When the comprehensive risk score is less than 0.5, it is a green warning, and routine observation can be performed, such as assessing vital signs every 2 hours. A comprehensive risk score greater than or equal to 0.5 and less than 1 is a yellow warning, and lactate and microcirculation monitoring parameters can be reviewed every 30 minutes to prepare for intervention. A comprehensive score greater than or equal to 1 is a red warning, and immediate intervention is required, such as through vasoactive drugs and target-directed fluid resuscitation, and upgrading antibiotics within one hour.

[0084] In one example, suppose Zhang San's comprehensive risk score is 1.26, MAP and lactate levels deteriorate synergistically (weight of the G spatial correlation feature is 0.83), and the temporal evolution characteristic indicates that lactate levels will exceed 6.0 mmol / L within 6 hours (TCN confidence level is 92%). Therefore, Zhang San needs to issue a red alert. Recommended interventions include: 1. Immediate intravenous bolus of norepinephrine (0.1 μg / kg / min); 2. Complete broad-spectrum antibiotic escalation (covering Gram-negative bacteria and anaerobic bacteria) within 30 minutes; and 3. Monitor MAP every 15 minutes and recheck microcirculatory function (MFI / PPV) every 30 minutes.

[0085] The early warning method for septic shock in the embodiments of the present application has a continuous feedback and optimization mechanism, capable of self-learning and adjustment based on real-time monitoring data to optimize early warning performance. Specifically, first, it has a data feedback mechanism that continuously optimizes model parameters through clinical feedback. For example, if there is a deviation between the early warning model's prediction and the actual treatment effect, the system can automatically adjust the parameters of the spatiotemporal modeling to better adapt to the new data pattern. Second, it has a dynamic adjustment mechanism that can perform personalized tuning based on actual monitoring data in different hospitals or clinical environments. If the microcirculation monitoring equipment is different, the system will automatically adjust the data processing algorithm to improve its prediction accuracy. Furthermore, it can also have a real-time optimization mechanism. After each early warning, the system will adjust the comprehensive risk score based on the patient's actual response data (such as changes in physiological parameters after treatment), further improving the model's predictive ability and accuracy.

[0086] The early warning method for septic shock of the embodiment of the present application has the following beneficial effects: 1. Improve early warning accuracy: By fusing multimodal data and introducing dynamic graph neural networks and time series modeling, the dynamic correlation between hemodynamics and microcirculatory parameters can be effectively captured, thereby improving the accuracy of early warning.

[0087] 2. Reduce warning delay: Compared with traditional static scoring systems, the embodiments of the present application can capture changes in patients' physiological parameters in real time, provide early warning, and significantly reduce warning delay.

[0088] 3. Accurate intervention recommendations: It can output personalized intervention recommendations based on the patient's real-time data, provide a scientific basis for clinical decision-making, and help optimize treatment plans.

[0089] 4. Feedback and self-optimization: It has the ability of self-learning and optimization, and can adjust the model according to real-time clinical feedback to improve long-term warning performance.

[0090] In summary, the embodiments of the present application utilize a deep learning algorithm based on dynamic graph neural networks and cross-scale time series modeling, integrating hemodynamic monitoring parameters, microcirculation monitoring parameters, and laboratory data to perform early risk assessment and real-time warning of septic shock. By constructing a dynamic correlation graph between physiological characteristic parameters, capturing short-term and long-term time series characteristics, it is possible to output risk probability values and time sensitivity indexes in real time, providing a scientific basis for clinical intervention. Therefore, the system provided in the embodiments of the present application can reduce the warning delay in traditional methods, improve warning accuracy, and possess a self-optimization and feedback mechanism, further enhancing its application value.

[0091] An embodiment of the present application further provides a computer-readable storage medium, wherein the computer-readable storage medium stores a program, which can be loaded by a processor and execute any one of the early warning methods for septic shock in the embodiments of the present application.

[0092] Since the instructions stored in the computer-readable storage medium can execute the steps of any of the early warning methods for septic shock provided in the embodiments of the present application, the beneficial effects that can be achieved by any of the early warning methods for septic shock provided in the embodiments of the present application can be achieved. Please refer to the previous embodiments for details and will not be repeated here.

[0093] Those skilled in the art will appreciate that all or part of the functions of the various methods in the above embodiments can be implemented by hardware or by computer program. When all or part of the functions in the above embodiments are implemented by computer program, the program can be stored in a computer-readable storage medium, and the storage medium can include: read-only memory, random access memory, disk, optical disk, hard disk, etc., and the program is executed by a computer to implement the above functions. For example, the program is stored in the memory of the device, and when the program in the memory is executed by the processor, all or part of the above functions can be implemented. In addition, when all or part of the functions in the above embodiments are implemented by computer program, the program can also be stored in a storage medium such as a server, another computer, disk, optical disk, flash disk or mobile hard disk, and saved in the memory of the local device by downloading or copying, or the system of the local device is updated. When the program in the memory is executed by the processor, all or part of the functions in the above embodiments can be implemented.

[0094] The above specific examples are used to illustrate the present application, which is only used to help understand the present application and is not intended to limit the present application. For those skilled in the art of the present application, based on the concept of the present application, they can also make some simple deductions, modifications or substitutions.

Claims

1. A method for early warning of septic shock, characterized in that: include: Acquiring multimodal physiological characteristic parameters of the user, wherein the multimodal physiological characteristic parameters include hemodynamic monitoring parameters, microcirculation monitoring parameters, and laboratory parameters, and the physiological characteristic parameters of each modality include at least one indicator parameter; Constructing a graph neural network based on the multimodal physiological characteristic parameters to obtain spatial correlation features corresponding to the multimodal physiological characteristic parameters; Extracting short-term time series features and long-term time series features of the multimodal physiological characteristic parameters based on the graph neural network, and obtaining time evolution features corresponding to the multimodal physiological characteristic parameters based on the short-term time series features and the long-term time series features; Fusing the spatial correlation feature and the temporal evolution feature to obtain a spatiotemporal feature vector corresponding to the physiological feature parameter; Based on the spatiotemporal feature vector, a risk probability and a time sensitivity index of septic shock of the user are obtained, and based on the risk probability and the time sensitivity index, a warning level of septic shock of the user is determined.

2. The early warning method according to claim 1, characterized in that: Each physiological characteristic parameter includes a physiological characteristic index and an index parameter value corresponding to the physiological characteristic index. The method of obtaining the multimodal physiological characteristic parameters of the user includes: According to the sampling method set for each modality, the indicator parameter values corresponding to the physiological characteristic indicators of each modality are collected respectively to obtain the initial indicator parameter values of the multimodality; Time-aligning the initial indicator parameter values of the multimodal states to obtain a time-synchronized multimodal physiological feature sequence; The multimodal physiological feature sequence is sliced according to a set time interval to obtain index parameter values of multiple time slices, each of which corresponds to a feature index.

3. The early warning method according to claim 2, characterized in that: The constructing of a graph neural network according to the multimodal physiological characteristic parameters to obtain spatial correlation features corresponding to the multimodal physiological characteristic parameters includes: Determine the characteristic indicators in each time slice as a node, and determine the indicator parameter values corresponding to the plurality of characteristic indicators as node attributes corresponding to the node; Determining associations between different feature indicators as edges, and determining edge weights corresponding to the edges based on an adjacency matrix of the edges; A graph neural network constructed based on the nodes, the node attributes, the edges, and the edge weights; Message passing is performed through the graph neural network, and edge weights are used to perform weighted aggregation on the node attributes of each node based on the node attributes of its neighboring nodes to obtain a feature vector for each node, wherein the feature vector of each node includes spatial correlation features between the node and its neighboring nodes.

4. The early warning method according to claim 3, characterized in that: The step of determining the indicator parameter values corresponding to the plurality of characteristic indicators as the node attributes corresponding to the node includes: For each time slice, obtaining characteristic indicators in the current time slice and the previous time slice and indicator parameter values corresponding to the characteristic indicators; determining a rate of change of the characteristic indicator according to the current time slice and the previous time slice; Determining a standard deviation of the characteristic indicator according to the indicator parameter value in the current time slice, wherein the standard deviation reflects the volatility of the characteristic indicator in the current time slice; The node attribute of the current time slice is obtained according to the indicator parameter value of the current time slice, the change rate of the characteristic indicator, and the standard deviation of the characteristic indicator.

5. The early warning method according to claim 3, characterized in that: Determining the association relationship between different characteristic indicators as an edge, and determining the edge weight corresponding to the edge based on the adjacency matrix of the edge, including: For each time slice, obtain the information entropy of the two nodes connected by each edge and the mutual information entropy between the two nodes; Obtaining a length of a current time slice, and determining an attenuation factor based on the length of the current time slice; Based on the information entropy, the mutual information entropy, and the attenuation factor, an adjacency matrix of the edge is determined, and a value of the adjacency matrix is determined as an edge weight corresponding to the edge.

6. The early warning method according to claim 2, characterized in that: The extracting of short-term time series features and long-term time series features of the multimodal physiological characteristic parameters based on the graph neural network includes: Obtaining a multimodal indicator time series based on physiological characteristic parameters in a current time slice and a set number of time slices before the current time slice; Inputting the indicator time series into the dilated causal convolutional network to calculate the short-term time series characteristics of the physiological characteristic parameters; The indicator time series is input into a converter model to calculate the long-term time series characteristics of the physiological characteristic parameters.

7. The early warning method according to claim 6, characterized in that: The obtaining of the time evolution characteristics corresponding to the multimodal physiological characteristic parameters based on the short-term time series characteristics and the long-term time series characteristics includes: Performing feature splicing on the short-term time series features and the long-term time series features, and performing linear layer transformation to obtain the time evolution features corresponding to the physiological feature parameters; or Inputting the short-term time series features and the long-term time series features into a first target gating network, and obtaining weights of the short-term time series features and the long-term time series features through dynamic calculation; Based on the weights of the short-term time series features and the long-term time series features, the short-term time series features and the long-term time series features are weightedly fused to obtain the time evolution features corresponding to the physiological characteristic parameters.

8. The early warning method according to claim 1, characterized in that: The fusing of the spatial correlation feature and the temporal evolution feature to obtain the spatiotemporal feature vector corresponding to the physiological feature parameter includes: Inputting the spatial correlation feature and the temporal evolution feature into a second target gating network, and obtaining weights of the spatial correlation feature and the temporal evolution feature through dynamic calculation; Based on the weights of the spatial correlation feature and the temporal evolution feature, the spatial correlation feature and the temporal evolution feature are weightedly fused to obtain a spatiotemporal feature vector corresponding to the physiological feature parameter.

9. The early warning method according to claim 1, characterized in that: Obtaining the risk probability and time sensitivity index of septic shock of the user based on the spatiotemporal feature vector, and determining the warning level of septic shock of the user based on the risk probability and the time sensitivity index, includes: Inputting the spatiotemporal feature vector into a logistic regression model to obtain the risk probability of septic shock of the user; Inputting the spatiotemporal feature vector into a lightweight long short-term memory network to obtain a time sensitivity index of septic shock of the user; Normalizing the risk probability and the time sensitivity index, and obtaining a set time sensitivity compression coefficient; Obtaining a comprehensive risk score according to the normalized risk probability, the time sensitivity index, and the sensitivity compression coefficient; A warning level of septic shock of the user is determined based on the comprehensive risk score.

10. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a program, which can be loaded by a processor to execute the early warning method for septic shock according to any one of claims 1 to 9.

Images