A device for analyzing the causes of sepsis based on matrix-based causal disentanglement

Abstract

The present invention discloses a device for analyzing the causal factors of sepsis based on matrix-based causal disentanglement. The device relates to the field of critical care medicine data processing and analysis. The device comprises a data acquisition module, a causal graph construction module, a model training module, and a factor analysis module. The causal graph construction module groups historical electronic health record time series data and constructs a structural causal matrix based on a hidden Markov chain. The model training module trains a serialized causal decoupling model based on the structural causal matrix. The serialized causal decoupling model includes a serialized variational autoencoder, a priori network, a posterior network, and a generative network. The factor analysis module groups current electronic health record time series data and determines the strength of association between each data group and sepsis based on the serialized causal decoupling model to analyze the causal factors of sepsis. This invention improves the accuracy of analyzing the causes of sepsis.

Description

Technical Field

[0001] The present invention relates to the field of critical care medicine data processing and analysis, and in particular to a sepsis causal factor analysis device based on matrix-type causal disentanglement. Background Art

[0002] Sepsis, a global health priority, remains a highly complex, diverse, and fatal disease. According to the World Health Organization, millions of people die from sepsis each year, and this number continues to rise. Treatment of sepsis is highly time-dependent, with evidence demonstrating that every hour of delay in antibiotic administration significantly increases the mortality rate. Early prediction of sepsis before onset provides clinicians with valuable time and effectively reduces mortality.

[0003] The causes of sepsis involve a complex combination of factors, including the infectious agent, host response, and a variety of environmental and biological variables. Sepsis is often associated with severe infection, with the most common sources including pulmonary, abdominal, urinary tract, and skin and soft tissue infections. Infection triggers a host immune response, leading to the massive release of inflammatory mediators. This excessive inflammatory response can cause extensive tissue damage and organ dysfunction.

[0004] In hospitals, the Sepsis-related Organ Failure Assessment (SOFA) is an effective method for assessing the severity of sepsis in patients. The SOFA score is based on several key indicators, including respiratory, cardiovascular, liver, coagulation, kidney, and central nervous system function. When the worst score within a day is greater than or equal to 2, the patient is generally considered to have sepsis. Although the SOFA score can diagnose sepsis with relative accuracy, it uses fewer indicators and is mainly used for diagnosis rather than early prediction.

[0005] In recent years, with the rapid development of machine learning and deep learning, these technologies have made progress in analyzing and understanding the causes of sepsis. For example, Horng et al. used models such as support vector machines, generalized linear models, naive Bayes, and expectation maximization to conduct a retrospective analysis of sepsis patients at a tertiary academic teaching hospital, revealing the relationship between different etiologies and the development of sepsis. Zhang et al. used Local Interpretable Model-Agnostic Explanations (LIME) to develop a model for real-time analysis of sepsis causes, effectively supporting clinical staff in understanding the impact of different factors on sepsis. Liu et al. used text mining techniques to analyze clinical text to improve understanding of the causes of sepsis. Gho et al. expanded text mining algorithms for sepsis cause analysis and advocated for modeling common topics to improve the accuracy of sepsis causes.

[0006] However, deep learning methods face a number of challenges in medicine, particularly in analyzing the causes of sepsis. First, due to their complexity, deep learning models lack interpretability. While they can capture correlations between data, they often struggle to parse true causal relationships, potentially leading to the learning of spurious correlations and insufficient generalization capabilities.

[0007] Several research efforts have attempted to address these issues. For example, Li et al. proposed a causal Markov model based on a variational auto-encoder (VAE). By distinguishing between hidden variables that are causally related to the disease and those that are not directly related to the disease, this model avoids learning spurious correlations and provides a robust approach for analyzing the causes of sepsis. Nie et al. employed instrumental variable learning to combine chest X-ray (CXR) images with electronic health record (EHR) data, imbuing the instrumental variables with medical semantics, effectively eliminating spurious correlations and improving the ability to analyze the causes of sepsis. Nie et al. further enhanced the interpretability of chest X-ray images and understanding of the causes of sepsis through a structured causal model and backdoor adjustment methods.

[0008] To address the problem of insufficient medical data, Shang et al. used graph neural networks to understand the relationships between medical codes and pre-trained the model on patients' initial visit data. By simulating analysis tasks in real-world scenarios, they improved the model's understanding of the causes of sepsis and fine-tuned the model in subsequent patient visits to improve parsing accuracy.

[0009] In general, existing research has provided some ideas and methods for analyzing the causes of sepsis, but some challenges still exist, including how to make full use of EHR data to effectively solve the problem of data sparsity, how to construct a reasonable structural causal graph to analyze the correct causal relationship, and how to improve the interpretability of the model to clarify the relevant indicators of sepsis. Summary of the Invention

[0010] The purpose of the present invention is to provide a sepsis causal factor analysis device based on matrix causal disentanglement, which can achieve accurate analysis and judgment of the cause of sepsis based on effective data decoupling analysis.

[0011] To achieve the above objectives, the present invention provides a device for analyzing the causes of sepsis based on matrix-based causal disentanglement, comprising:

[0012] A data acquisition module is configured to acquire historical electronic health record time series data, including laboratory test indicators, personal attributes, clinical text, and disease tags at multiple times within a set historical time period; the disease tag is either 0 or 1, where a disease tag of 1 indicates sepsis, and a disease tag of 0 indicates no sepsis;

[0013] a causal graph construction module, connected to the data acquisition module, for grouping the historical electronic health record time series data to obtain multiple groups of historical factor data, and constructing a structural causal graph based on a hidden Markov chain based on each group of historical factor data to obtain a structural causal matrix; the structural causal graph is used to describe the historical factor data as a high-dimensional latent space;

[0014] A model training module is connected to the causal graph construction module and is used to train and optimize the serialized causal decoupling model based on the structural causal matrix according to each set of historical factor data to obtain a trained serialized causal decoupling model; the serialized causal decoupling model includes a serialized variational autoencoder, a priori network, a posterior network and a generative network; the serialized variational autoencoder is used to encode the historical factor data; the priori network and the posterior network are used to parse the causal relationship in the structural causal graph based on the encoded historical factor data, and decouple to obtain multiple latent variables; the multiple latent variables are: latent variables related to sepsis, latent variables related to other diseases, and latent variables unrelated to the disease; the generative network is used to analyze the correlation strength between the latent variables related to sepsis and sepsis;

[0015] The factor analysis module is connected to the model training module and is used to group the current electronic health record time series data to obtain multiple groups of current factor data, and determine the correlation strength between each group of current factor data and sepsis based on the trained serialized causal decoupling model to analyze the causal factors of sepsis.

[0016] According to the specific embodiments provided by the present invention, the present invention discloses the following technical effects: The present invention provides a device for analyzing the causes of sepsis based on matrix causal disentanglement, constructs a structural causal graph based on a hidden Markov chain, which can accurately describe the causal relationship of each variable over time, and proposes a framework based on a serialized variational autoencoder to effectively learn the causal relationship described by the structural causal graph. Through this framework, the latent variables related to sepsis are successfully decoupled from the confounding factors. The serialized causal decoupling model is trained and optimized using the time series data in the historical EHR, and the clinical data and physiological parameters of the patient are integrated, making full use of the temporal and spatial information in the EHR data to decouple complex causal relationships, which not only improves the interpretability of the cause analysis of sepsis, but also provides important support for a deep understanding of the pathogenesis of sepsis, so that clinical researchers can understand the pathogenesis of sepsis more clearly. BRIEF DESCRIPTION OF THE DRAWINGS

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

[0018] Figure 1 This is a module schematic diagram of the sepsis causal factor analysis device based on matrix causal disentanglement provided by the present invention.

[0019] Figure 2 This is a flow chart of the method for analyzing the causal factors of sepsis based on matrix-type causal disentanglement provided by the present invention.

[0020] Figure 3 It is a structural causal graph based on hidden Markov chain.

[0021] Figure 4 Schematic diagram of a sequential variational autoencoder.

[0022] Figure 5 Schematic diagram of the network structure of the serialized causal decoupling model.

[0023] Figure 6 This is a schematic diagram of an implementation process of the method for analyzing the causes of sepsis based on matrix-based causal disentanglement provided by the present invention.

[0024] Figure 7 This is a schematic diagram of another implementation process of the method for analyzing the causes of sepsis based on matrix-based causal disentanglement provided by the present invention.

[0025] Figure 8 Schematic diagram of the structure of group matrix decoupling analysis.

[0026] Figure 9 Receiver operating characteristic curves of various models for early prediction of sepsis.

[0027] Figure 10 This is a diagram of the internal structure of a computer device. DETAILED DESCRIPTION

[0028] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0029] The purpose of the present invention is to provide a device for analyzing the causes of sepsis based on matrix causal disentanglement. By combining causal disentanglement, hidden Markov chain, and serialized variational autoencoder technologies, a serialized causal decoupling model is established to fully utilize the time and distribution information contained in the data, effectively avoid the influence of confounding factors, avoid false correlations, and improve the accuracy of sepsis cause analysis.

[0030] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0031] Example 1: Figure 1 As shown, this embodiment provides a device for analyzing causal factors of sepsis based on matrix causal disentanglement, including a data acquisition module 101, a causal graph construction module 102, a model training module 103 and a factor analysis module 104.

[0032] The data acquisition module 101 is used to acquire historical electronic health record time series data.

[0033] The causal graph construction module 102 is connected to the data acquisition module 101. The causal graph construction module 102 is used to group the historical electronic health record time series data to obtain multiple groups of historical factor data, and construct a structural causal graph based on a hidden Markov chain based on each group of historical factor data to obtain a structural causal matrix.

[0034] The model training module 103 is connected to the causal graph construction module. The model training module 103 is used to train and optimize the serialized causal decoupling model based on each set of historical factor data and the structural causal matrix to obtain a trained serialized causal decoupling model.

[0035] The factor analysis module 104 is connected to the model training module 103. The factor analysis module 104 is used to group the current electronic health record time series data to obtain multiple groups of current factor data, and based on the trained serialized causal decoupling model, determine the correlation strength between each group of current factor data and sepsis to analyze the causal factors of sepsis.

[0036] Furthermore, the model training module 104 includes: an exogenous variable determination submodule, a coding submodule, a priori / posteriori submodule, an association strength determination submodule and a training submodule.

[0037] The exogenous variable determination submodule is used to extract the spatial distribution characteristics of laboratory test indicators and personal attributes at any time during the historical setting period to determine t exogenous variables at the time; according to t -2 time exogenous variables, t Exogenous variables at time -1 and t The exogenous variables at the moment are encoded using the serialized variational autoencoder to obtain t The time distribution information of the exogenous variables at the moment, where t For any time within the historical setting period, 1< t < T , T Set the final moment of the period for history.

[0038] The encoding submodule is used to adopt the sequential variational autoencoder to t -2 laboratory test indicators, t -1 moment laboratory test indicators and t The laboratory test indicators at the time are coded to obtain t The time distribution information of laboratory test indicators at the moment. t -2 moments of personal attributes, t -1 moment of personal attributes and t Encode the personal attributes of the moment to obtain t Time distribution information of personal attributes at each moment.

[0039] The a priori and posterior submodule is used to adopt the a priori network, according to t The time distribution information of the exogenous variables at the moment and t -1 time the posterior mean of each latent variable, calculate t The prior mean and prior variance of each latent variable at the moment, and define the prior distribution; using the posterior network, according to t Time distribution information of exogenous variables at each moment, t Time distribution information of laboratory test indicators at each moment,t Time distribution information of personal attributes at each moment and t -1 time, the posterior mean of each latent variable is determined t The posterior mean and posterior variance of the latent variable at time t, and define the posterior distribution.

[0040] The association strength determination submodule is used to adopt the generation network, according to t The posterior mean and posterior variance of each latent variable at the moment are determined t The latent variables after reparameterization at each moment, and according to t The latent variables after moment reparameterization are reconstructed t Laboratory test indicators at all times and t The personal attributes of the moment, according to T The latent variables related to sepsis and other diseases after the reparameterization of time are reconstructed. T Clinical text of the moment, according to T The latent variables related to sepsis after time reparameterization were used to analyze the correlation strength between the latent variables related to sepsis and sepsis.

[0041] The training submodule is used to reconstruct the laboratory test indicators at each moment within the historical set period, the personal attributes reconstructed at each moment, T Constantly reconstructed clinical texts, T Predicted disease labels at each moment, association strength at each moment, personal attributes at each moment, T Clinical texts at all times, T The diseased label at the moment, the prior distribution and the posterior distribution are optimized with the lower bound of evidence as the optimization target, so as to obtain a trained serialized causal decoupling model.

[0042] Furthermore, if Figure 2 As shown, the present invention is based on the above-mentioned sepsis causal factor analysis device based on matrix causal disentanglement, and further provides a sepsis causal factor analysis method based on matrix causal disentanglement, including steps 201 to 204.

[0043] Step 201: Obtain historical electronic health record time series data.

[0044] The historical electronic health record time series data includes laboratory test indicators at multiple times within a historical set period A , personal attributes B , clinical text Z (including antibiotic use) and disease labels Y .

[0045] Specifically, laboratory test indicators include platelet count, white blood cell count, and blood pressure. Personal attributes include height, weight, age, and race. Clinical text includes medical orders and surgical records. In this example, EHR data from intensive care unit (ICU) patients in the MIMIC-IV database were analyzed.

[0046] The present invention has a great impact on laboratory testing indicators. A , personal attributes B , clinical text Z Conduct group training, and keep clinical text in each group Z The data of antibiotics in each group were collected and then analyzed separately. The following steps are the analysis process of each group.

[0047] Step 202: Group the historical electronic health record time series data to obtain multiple groups of historical factor data, and construct a structural causal graph based on a hidden Markov chain based on each group of historical factor data to obtain a structural causal matrix. The structural causal graph is used to describe the historical electronic health record time series data as a high-dimensional latent space.

[0048] The structural causal graph describes the known input data as an unknown high-dimensional latent space, then disentangles the high-dimensional latent space into different latent variables, reconstructs the input data based on causal relationships, and finally uses only latent variables related to sepsis to analyze the causal factors of sepsis.

[0049] like Figure 3 As shown, the structure of the structural causal diagram is: and Jointly generate exogenous variables , the exogenous variable The latent space decoupling described is described as three latent variables: latent variables related to sepsis , hidden variables related to other diseases and latent variables unrelated to the disease . In each time step, by Refactoring ,pass and Refactoring ,pass 、 and Refactoring . In the last time step, by The causative factors of sepsis were analyzed.

[0050] In a structural causal graph, a child node is determined only by the parent node pointing to it and an independent exogenous noise, as shown in the following formula.

[0051] .

[0052] .

[0053] .

[0054] .

[0055] in, Decide The distribution of for t Momentary Impact The exogenous noise, , Represents a moment, represent The previous moment of the moment.

[0056] Right now t Exogenous variables at time Depend on t Laboratory test indicators at all times 、 t Personal attributes of the moment and independent exogenous noise Decide; t Hidden variables associated with sepsis at all times Depend on t Hidden variables associated with sepsis at time -1 、 t Exogenous variables at time and independent exogenous noise Decide; t Hidden variables associated with other diseases at all times Depend on t -1 Hidden variables related to other diseases 、 t Exogenous variables at time and independent exogenous noise Decide; t Hidden variables that are not related to the disease at any time Depend on t- Hidden variables that are not related to the disease at time 1 、 t Exogenous variables at time and independent exogenous noise Decide; t Laboratory test indicators at all times Depend on t Hidden variables associated with sepsis at all times and independent exogenous noise Decide; t Personal attributes of the moment Depend on t Hidden variables associated with sepsis at all times 、 t Hidden variables associated with other diseases at all times 、 t Hidden variables that are not related to the disease at any time and independent exogenous noise Decide; t Illness label at all times Depend on t Hidden variables associated with sepsis at all times and independent exogenous noise Decide; t Clinical text at the moment Depend on t Hidden variables associated with sepsis at all times 、 t Hidden variables associated with other diseases at all times and independent exogenous noise Decide.

[0057] In the present invention, t Time refers to the continuous time in units of one hour. t hours.

[0058] Step 203: According to each set of historical factor data and based on the structural causal matrix, the serialized causal decoupling model is trained and optimized to obtain a trained serialized causal decoupling model.

[0059] The serialized causal decoupling model includes a serialized variational autoencoder, a priori network, a posterior network, and a generative network. The serialized variational autoencoder is used to encode historical factor data. The priori network and the posterior network are used to analyze the causal relationship in the structural causal graph based on the encoded historical factor data and decouple multiple latent variables. The multiple latent variables are: latent variables related to sepsis, latent variables related to other diseases, and latent variables unrelated to the disease. The generative network is used to analyze the strength of the association between the latent variables related to sepsis and sepsis.

[0060] This paper uses a causal disentanglement method based on a sequenced variational autoencoder to study the causal relationships described by hidden Markov chains and decouple latent variables. By designing a specialized encoder, the temporal and distributional features of the input data are deeply explored and mapped into a high-dimensional latent space. The structure of this latent space is first defined and trained using a priori networks to constrain the latent space construction, making it more reasonable and interpretable. This process not only helps to reveal the complex relationships between latent variables but also identifies key factors associated with sepsis.

[0061] Specifically, step 203 includes steps 301 to 308 .

[0062] Step 301: For any moment in the historical setting period, through multiple consecutive fully connected layers Extract features from laboratory test indicators and personal attributes at the time , split it into two parts evenly on the last dimension of the tensor, the first part Directly defined as the mean , the latter part Through the activation function Later defined as variance , and finally reparameterize to obtain an exogenous variable that conforms to the Gaussian distribution at the moment .

[0063] Step 302: According to t -2 time exogenous variables, t Exogenous variables at time -1 and t The exogenous variables at the moment are encoded using the serialized variational autoencoder to obtain t The time distribution information of the exogenous variables at the moment. Among them, t For any moment in the historical setting period, 1< t <T,T Set the last moment in the history period.

[0064] In this embodiment, a point convolution based encoder is first used to t -2 time exogenous variables, t Exogenous variables at time -1 and t The exogenous variables at the moment are encoded to obtain t Then, multiple preliminary feature maps at the moment t The multiple preliminary feature maps at the moment are spliced together, and the encoder based on depthwise convolution is used to encode the spliced preliminary feature maps to obtain t The final feature map at the moment. Finally, according to t The final feature map at the moment is calculated using the formula Sure tThe time distribution information of the exogenous variables at the moment. Among them, for t Time distribution information of exogenous variables at each moment for t The final feature map at time t, for t exogenous variables at time, represents a full connection operation, Represents the shape of the reshaped tensor, Represents the addition of corresponding points in the tensor.

[0065] Step 303: According to t -2 laboratory test indicators at the moment, t -1 moment laboratory test indicators and t The laboratory test indicators at the moment are encoded using the serialized variational autoencoder to obtain t Time distribution information of laboratory test indicators at each moment.

[0066] Step 304: According to t -2 moments of personal attributes, t -1 moment of personal attributes and t The personal attributes at the moment are encoded using the serialized variational autoencoder to obtain t Time distribution information of personal attributes at each moment.

[0067] like Figure 4 As shown, the sequential variational autoencoder will include three consecutive hours of data including the data at the moment to be encoded Parallel in the channel direction, first do a indivual Point convolution of convolution kernel: ;in, is the preliminary feature map, It is an encoder based on point convolution, where N is equal to the feature dimension of the encoded data, and the feature dimension refers to the number of indicators of laboratory test indicators.

[0068] Point convolution can combine the input feature maps in the direction of the channel. Several convolution kernels will output several feature maps. Point convolution only changes the number of feature maps, not the size of the feature maps. preliminary feature maps, and concatenate them together to form X The feature map of Depthwise convolution of convolution kernel: ;in, is the first preliminary feature map, is the Nth preliminary feature map, Indicates the concatenation of tensors. This is an encoder based on depthwise convolution. Each channel is convolved with only one convolution kernel. Depthwise convolution does not change the number of feature maps, only their size. Finally, the feature map is reshaped and, together with the data to be encoded at that moment, it is resized and added together through a fully connected layer to output the encoded result, forming a residual sum structure: ;in, is the time distribution information of the input data, is a fully connected layer, Reshape a tensor.

[0069] Step 305: Using the prior network, according to t Time distribution information of exogenous variables at each moment, t -1 time, the posterior mean of each latent variable is determined t The prior mean and prior variance of each latent variable at time t, and define the prior distribution.

[0070] like Figure 5 In the prior network shown in the figure, spatial distribution features are first extracted from laboratory test indicators and personal attributes to jointly generate exogenous variables that can indicate sample distribution. The temporal distribution information is extracted through the serialized variational autoencoder and encoded as . Respectively t The hidden variables at time -1 are concatenated, and then the prior means and prior variances of the three hidden variables are parameterized through a layer of gated recurrent unit (GRU) and two different fully connected layers.

[0071] In this embodiment, the prior network is determined using the following formula t The prior mean and prior variance of each latent variable at time t.

[0072] .

[0073] .

[0074] .

[0075] in, for t The prior mean of the latent variable associated with sepsis at time t, for t The prior variance of the latent variable associated with sepsis at time t, for tThe prior mean of the latent variables related to other diseases at each moment, for t Prior variance of latent variables related to other diseases at each moment, for t The prior mean of the latent variable that is not related to the disease at any time, for t The prior variance of the latent variable that is not related to the disease at any given moment, for t The posterior mean of the latent variable associated with sepsis at time -1, for t The posterior mean of the latent variables related to other diseases at time -1, for t The posterior mean of the latent variable that is not related to the disease at time -1, for t Time distribution information of exogenous variables at the moment, represents a gated recurrent unit, 、 、 、 、 、 Represents fully connected layers that are trained independently, and their superscripts indicate that the fully connected layers act on the generated variables { The prior mean of or prior variance .

[0076] Taking the latent variables related to sepsis as an example, and The size after splicing is 128, parameterized The GRU input size is 128, the output size is 256, the input size of the two fully connected layers is 256, and the output size is 64. This encoding method makes full use of the temporal and spatial information in the data, avoids learning false correlations in the data, and improves robustness.

[0077] The prior distribution is: .

[0078] in, is the prior distribution, are the parameters of the prior network, for t The set of hidden parameters at time , ={ , }, for t The hidden parameter set at time -1, for texogenous variables at time, for t Hidden variables associated with sepsis at all times, for t Hidden variables related to sepsis at time -1, for t Hidden variables that are constantly associated with other diseases, for t -1 time is associated with other diseases, for t Hidden variables that are not related to the disease at any time, for t -1 time is a hidden variable that is not related to the disease.

[0079] Step 306: Using the posterior network, according to t Time distribution information of exogenous variables at each moment, t Time distribution information of laboratory test indicators at each moment, t Personal attribute time distribution information at each moment and t -1 time the posterior mean of each latent variable, calculate t The posterior mean and posterior variance of the latent variable at time t, and define the posterior distribution.

[0080] This paper proposes a method that uses a posterior distribution that is easy to learn to approximate the prior distribution. The posterior network further optimizes the representation of the latent space, analyzes causal relationships, and decouples low-dimensional latent variables. A sequential variational autoencoder decodes the latent variables into new samples in the original data space. Finally, by analyzing the latent variables associated with sepsis, the causal factors of sepsis are analyzed.

[0081] like Figure 5 In the posterior network shown in the figure, spatial distribution features are also extracted from laboratory test indicators and personal attributes to jointly generate exogenous variables that can indicate sample distribution. The laboratory test indicators, personal attributes and exogenous variables are encoded into 、 and .Will 、 and Respectively t The hidden variables at time -1 are concatenated, and the means and variances of the three hidden variables are parameterized through a long short-term memory network (LSTM) and two different fully connected layers.

[0082] In this embodiment, the posterior network is determined using the following formula t The posterior mean and posterior variance of each latent variable at time t.

[0083] .

[0084] .

[0085] .

[0086] in, for t The posterior mean of the latent variable associated with sepsis at time t, for t The posterior variance of the latent variable associated with sepsis at time t, for t The posterior mean of the latent variables related to other diseases at each moment, for t The posterior variance of the latent variables related to other diseases at each moment, for t The posterior mean of the latent variable that is not related to the disease at any time, for t The posterior variance of the latent variable that is not related to the disease at any given moment, for t The posterior mean of the latent variable associated with sepsis at time -1, for t The posterior mean of the latent variables related to other diseases at time -1, for t The posterior mean of the latent variable that is not related to the disease at time -1, for t Time distribution information of exogenous variables at the moment, for t Time distribution information of laboratory test indicators at each moment, for t Time distribution information of personal attributes at each moment, represents the long short-term memory network, 、 、 、 、 、 Represents different fully connected layers.

[0087] Taking the latent variables related to sepsis as an example, 、 、 and The size after splicing is 320, parameterized The LSTM layer has an input size of 320 and an output size of 128, and the two fully connected layers have an input size of 128 and an output size of 64.

[0088] The posterior distribution is: .

[0089] in, is the posterior distribution, are the parameters of the posterior network, for t The first intermediate parameter set at time, , is the second intermediate parameter set, , for t Laboratory test indicators at all times, for t The personal attributes of the moment, for T Clinical texts of the moment, for T The disease label at all times, for T Hidden variables associated with sepsis at all times, for T Hidden variables that are constantly associated with other diseases, Represents a continuous multiplication operation.

[0090] Step 307: Using the generated network, according to t The posterior mean and posterior variance of each latent variable at the moment are determined t The latent variables after reparameterization at each moment, and according to t The latent variables after moment reparameterization are reconstructed t Laboratory test indicators and t Personal attributes of the moment. At the same time, according to T The latent variables after constant reparameterization were used to reconstruct clinical texts, and based on the latent variables related to sepsis, the strength of the association between each set of historical factor data and sepsis was confirmed.

[0091] like Figure 5 In the generative network shown in Figure 1, after obtaining the posterior mean and posterior variance, the latent variables following the Gaussian distribution can be reparameterized based on the posterior mean and posterior variance.

[0092] In this embodiment, the generation network is determined by the following formula t The latent variables after reparameterization at each moment.

[0093] .

[0094] .

[0095] .

[0096] in, for t The latent variables related to sepsis after time reparameterization, for t Hidden variables related to other diseases after constant reparameterization, for t The latent variables that are not related to the disease after constant reparameterization, for t The posterior mean of the latent variable associated with sepsis at time t, for t The posterior variance of the latent variable associated with sepsis at time t, for t The posterior mean of the latent variables related to other diseases at each moment, for t The posterior variance of the latent variables related to other diseases at each moment, for t The posterior mean of the latent variable that is not related to the disease at any time, for t The posterior variance of the latent variable that is not related to the disease at any given moment, For one with Random matrices of the same size that follow a normal distribution with mean 0 and variance 1, For one with Random matrices of the same size that follow a normal distribution with mean 0 and variance 1, For one with Random matrices of the same size that follow a normal distribution with mean 0 and variance 1, Represents the multiplication of corresponding points of two matrices.

[0097] The generative network is reconstructed using the following formula t Laboratory test indicators at all times, t Personal attributes of the moment and T Clinical text of the moment.

[0098] .

[0099] .

[0100] .

[0101] in, for t Laboratory test indicators that are constantly being reconstructed, for t Personal attributes that are constantly reconstructed, forT Clinical texts that are constantly being reconstructed, for t The latent variables related to sepsis after time reparameterization, for t Hidden variables related to other diseases after constant reparameterization, for t The latent variables that are not related to the disease after constant reparameterization, Represents a generative model consisting of multiple consecutive fully connected layers.

[0102] At each moment, the generative model is a Gaussian distribution parameterized by a fully connected layer. For example, in the generation of personal attributes, the three re-parameterized latent variables are concatenated and passed through multiple consecutive fully connected layers as a generative network with an input size of 128 and an output size of 7.

[0103] In this embodiment, laboratory test indicators and personal attributes are reconstructed at each moment, and clinical text is generated only at the last moment. At the same time, after fully learning and extracting features from multiple consecutive hours of data, at the last moment, the decoupled latent variables related to sepsis are used for classification to determine T Predicted disease label at the moment: ,in, It is a classification model. Since the latent variables have already been extracted, a complex classification model is not required. Instead, multiple consecutive fully connected layers are used, with an input size of 64 and an output size of 2. Finally, a sigmoid layer is used to output the classification result. The classification result is represented by a disease label score. A higher score indicates a stronger association with sepsis in the historical factor data.

[0104] The distribution of laboratory test indicators is , the distribution of personal attributes is , the distribution of clinical text and disease labels is: .

[0105] The goal of optimizing the network structure is the Evidence Lower Bound (ELBO): .

[0106] According to the causal hidden Markov condition, the joint distribution probability at each moment is: .

[0107] Substituting the joint distribution probability into ELBO, we obtain the optimization objective of the serialized causal decoupling model as described in the following formula.

[0108] .

[0109] .

[0110] .

[0111] .

[0112] in, is the optimization target value of the serialized causal decoupling model, for t Time distribution information of exogenous variables at the moment, for t -1 time distribution information of exogenous variables, Express expectations, is the first intermediate variable, is the second intermediate variable, describes the divergence between the prior and posterior distributions, is the third intermediate variable, exist did not exist before the moment The posterior distribution at the moment is approximately equal to the prior distribution, which is equal to 0, so Can be omitted.

[0113] Step 308: Reconstruct the laboratory test indicators, personal attributes, and T Clinical texts and predicted disease labels reconstructed at each moment, combined with laboratory test indicators, personal attributes, T The clinical text, actual disease label, prior distribution and posterior distribution at each moment are optimized by maximizing ELBO to optimize the parameters of the prior network and the posterior network, thereby obtaining a trained serialized causal decoupling model.

[0114] like Figure 6 The figure shows a schematic diagram of an implementation process of the method for analyzing the causes of sepsis based on matrix causal disentanglement provided by the present invention. First, initial data is obtained, then the initial data is preprocessed, and then the causes of sepsis are analyzed based on the prior network, the posterior network and the generative network.

[0115] Step 204: Group the current electronic health record time series data to obtain multiple groups of current factor data, and determine the correlation strength between each group of current factor data and sepsis based on the trained serialized causal decoupling model to analyze the causal factors of sepsis.

[0116] It should be noted that the execution entities of the above steps 201 to 204 are all computers.

[0117] like Figure 7 As shown, the present invention extracts laboratory test indicators, personal attributes and clinical texts of ICU patients to establish a data set. The samples in the data set are a complex and mutually entangled high-dimensional space. The present invention decouples the high-dimensional space into low-dimensional latent variables and classifies them by latent variables related to sepsis, effectively eliminating the influence of confounding factors on model analysis. A structural causal graph describing the relationship between latent variables and observed data is constructed using a hidden Markov chain. A serialized variational autoencoder is used to learn causal features and decouple latent variables. The samples in the data set are cleaned and normalized, and a training set, a test set and a validation set are established. The appropriate time step and feature dimension are set, and the serialized causal decoupling model is trained and optimized using the training set. After training, the trained model is evaluated based on the validation set, and the parameters are adjusted again and trained repeatedly to achieve the best results. Finally, the trained serialized causal decoupling model is applied to current electronic health record data. By constructing a causal structure matrix and analyzing the correlation strength between different data and sepsis, the causal factors of sepsis are analyzed. The present invention integrates the patient's clinical data and physiological parameters, fully utilizes the temporal and spatial information in ICU patient data, decouples complex causal relationships, improves the interpretability of the results, and achieves accurate prediction of sepsis.

[0118] like Figure 8 As shown, the matrix-based causal disentanglement-based sepsis causal factor analysis method provided by the present invention includes the following key steps. First, the initial data is preprocessed to extract key information such as laboratory test indicators, personal attributes, and clinical text to ensure data quality and consistency. Next, the preprocessed data set is grouped, such as into multiple groups of data such as {G1, G2} and {G1, G3}, to obtain grouped sets. Causal decoupling analysis is performed on each data set. A structural causal graph based on a hidden Markov chain is constructed and the data is encoded using a sequenced variational autoencoder to analyze causal relationships. The latent variables are decoupled into latent variables related to sepsis, latent variables related to other diseases, and latent variables unrelated to the disease. A sepsis prediction score is then obtained for each grouped set. The score is used to determine which group has factors more relevant to sepsis. Further grouping is then performed, and the above process is repeated to obtain the final factor judgment result. The present invention first constructs a grouping matrix to obtain multiple grouping sets at one time. After multiple group judgments, superposition judgments are performed based on the analyzed results to obtain the final cause analysis results.

[0119] In order to increase the analysis speed, the present invention adopts parallel computing, and at the same time, takes into account the importance of antibiotic use, and distributes it as a key factor in all groups to ensure that its relevance to sepsis is fully considered. Through causal decoupling analysis, the sepsis prediction score of each group is calculated, and the factors that are more relevant to sepsis in each group are judged according to the size of the prediction score. Through repeated screening and grouping, the analysis is gradually refined, and the key causal factors of sepsis are finally determined. After multiple iterative analyses, the analysis results of each group are combined by the method of superposition judgment to form a final causal analysis report. This process can not only effectively avoid the influence of confounding factors and avoid false correlations, but also significantly improve the accuracy of sepsis causal analysis, providing an important reference for clinical diagnosis and treatment. In particular, through parallel computing and comprehensive consideration of antibiotic factors, the present invention can quickly and accurately identify key factors related to sepsis, thereby improving the efficiency and reliability of causal analysis.

[0120] The feasibility of the sepsis causal factor analysis method based on matrix causal disentanglement provided by the present invention is verified below in combination with specific data sets and calculation formulas.

[0121] The dataset used to evaluate the early sepsis prediction method of the present invention is the Medical Information Market for Intensive Care (MIMIC-IV v2.2) database. This database, in EHR format, includes various data, including personal attributes, laboratory test results, vital signs, disease diagnoses, and medication usage. In clinical practice, to save time and costs, doctors often stop testing a patient's parameters when they believe they are within the normal range, resulting in a large number of missing values in the database. Furthermore, due to irregular recording procedures or unknown issues, some outliers may also exist in the database. The present invention cleans the data based on the normal value ranges provided by Harutyunyan et al., merging all observations in hourly units. If the same parameter is tested multiple times within an hour, the most recent measurement is used. The data are sorted by missing rate, discarding parameters with excessively high missing rates. For missing values, both up-filling and down-filling methods are used. If a patient's parameter is missing a value, it indicates that the doctor believes the patient's parameter is normal, so a value within the normal range is filled in.

[0122] After data cleaning and filling, the maximum and minimum normalization method is selected for normalization: .

[0123] in, is the normalized index, For an indicator, For an indicator in the sample The minimum value of For an indicator in the sample The normalized data set is more standardized, free from the influence of data anomalies, and thus improves the accuracy of prediction.

[0124] After data processing, we obtained 3,511 positive samples and 17,538 negative samples. We divided the data into training, test, and validation sets in a 3:1:1 ratio. We set the maximum number of training epochs to 200 and the batch size to 128. We used the Adam optimizer for training, with an initial learning rate of 1e-3 and reducing it to one-fifth of the original rate every 10 epochs.

[0125] The proposed method for analyzing the causes of sepsis based on matrix causal disentanglement was evaluated on MIMIC-IV. Figure 9 As shown in the figure, the performance is measured by the area under the receiver operator characteristic (ROC) curve (Area Under Curve, AUC) and the F1 score, as shown in the following formula.

[0126] .

[0127] .

[0128] .

[0129] .

[0130] in, is the area under the ROC curve, The ratio of samples predicted to be positive to samples that are actually positive reflects the accuracy of the model in predicting positive samples. is the proportion of all actual positive classes that are correctly predicted as positive, reflecting the ability of the model to capture positive classes. The F1 score represents the harmonic mean of precision and recall, and is used to measure the balance between precision and recall. , is the prediction score of the positive sample, is the prediction score of the negative sample, is the number of positive samples, is the number of negative samples, TP is the number of true positive samples in the prediction results, FP is the number of false positive samples in the prediction results, and FN is the number of false negative samples in the prediction results.

[0131] The present invention uses Grad-CAM to generate a grayscale graph to represent the results of the causal factor analysis. The horizontal axis of the grayscale graph represents the number of hours since the onset of sepsis, and the vertical axis represents the variable name. Lighter colors indicate a lower contribution to the model's prediction of sepsis at that time, while darker colors indicate a higher contribution. When a variable is black at all time points, it indicates that the variable is clinically significant in indicating sepsis.

[0132] Embodiment 2: A computer device comprises: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement a method for analyzing causal factors of sepsis based on matrix-based causal disentanglement.

[0133] Embodiment 3: A computer-readable storage medium stores a computer program, which, when executed by a processor, implements a method for analyzing causal factors of sepsis based on matrix-type causal disentanglement.

[0134] Embodiment 4: A computer program product, comprising a computer program, wherein when the computer program is executed by a processor, a method for analyzing causal factors of sepsis based on matrix-based causal disentanglement is implemented.

[0135] Example 5: A computer device, which may be a database, and its internal structure diagram may be as follows Figure 10 As shown. The computer device includes a processor, a memory, an input / output interface (I / O) and a communication interface. The processor, memory and input / output interface are connected via a system bus, and the communication interface is connected to the system bus via the input / output interface. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and computer program in the non-volatile storage medium. The database of the computer device is used to store pending transactions. The input / output interface of the computer device is used to exchange information between the processor and an external device. The communication interface of the computer device is used to communicate with an external terminal via a network connection. When the computer program is executed by the processor, a method for analyzing the causes of sepsis based on matrix causal disentanglement is implemented.

[0136] It should be noted that the object information (including but not limited to object device information, object personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in the present invention are all information and data authorized by the object or fully authorized by all parties, and the collection, use and processing of relevant data must comply with the relevant laws, regulations and standards of relevant countries and regions.

[0137] Those skilled in the art will appreciate that all or part of the processes in the above-described method embodiments can be implemented by instructing the relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the above-described method embodiments. Any reference to memory, database, or other media used in the embodiments provided herein may include at least one of non-volatile and volatile memory. Non-volatile memory may include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory may include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM). The databases involved in the various embodiments provided herein may include at least one of a relational database and a non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the various embodiments provided herein may be, but are not limited to, general-purpose processors, central processing units (CPUs), graphics processing units (GPUs), digital signal processors (DSPs), programmable logic devices (PLDs), quantum computing-based data processing logic devices, and the like.

[0138] The technical features of the above embodiments can be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0139] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The above examples are only intended to help understand the method and core concept of the present invention. At the same time, those skilled in the art will find that the specific implementation methods and application scopes may vary based on the concept of the present invention. In summary, the contents of this specification should not be construed as limiting the present invention.

Claims

1. A device for analyzing the causes of sepsis based on matrix-based causal disentanglement, characterized in that: The sepsis causal factor analysis device based on matrix-type causal disentanglement includes: A data acquisition module is configured to acquire historical electronic health record time series data, including laboratory test indicators, personal attributes, clinical text, and disease tags at multiple times within a set historical time period; the disease tag is either 0 or 1, where a disease tag of 1 indicates sepsis, and a disease tag of 0 indicates no sepsis; a causal graph construction module, connected to the data acquisition module, for grouping the historical electronic health record time series data to obtain multiple groups of historical factor data, and constructing a structural causal graph based on a hidden Markov chain based on each group of historical factor data to obtain a structural causal matrix; the structural causal graph is used to describe the historical factor data as a high-dimensional latent space; A model training module is connected to the causal graph construction module and is used to train and optimize the serialized causal decoupling model based on the structural causal matrix according to each set of historical factor data to obtain a trained serialized causal decoupling model; the serialized causal decoupling model includes a serialized variational autoencoder, a priori network, a posterior network and a generative network; the serialized variational autoencoder is used to encode the historical factor data; the priori network and the posterior network are used to parse the causal relationship in the structural causal graph based on the encoded historical factor data, and decouple to obtain multiple latent variables; the multiple latent variables are: latent variables related to sepsis, latent variables related to other diseases, and latent variables unrelated to the disease; the generative network is used to analyze the correlation strength between the latent variables related to sepsis and sepsis; The generating network is determined by the following formula t The latent variables after reparameterization at each moment: ； ； ； in, for t The latent variables related to sepsis after time reparameterization, for t Hidden variables related to other diseases after constant reparameterization, for t The latent variables that are not related to the disease after constant reparameterization, for t The posterior mean of the latent variable associated with sepsis at time t, for t The posterior variance of the latent variable associated with sepsis at time t, for t The posterior mean of the latent variables related to other diseases at each moment, for t The posterior variance of the latent variables related to other diseases at each moment, for t The posterior mean of the latent variable that is not related to the disease at any time, for t The posterior variance of the latent variable that is not related to the disease at any given moment, For one with Random matrices of the same size that follow a normal distribution with mean 0 and variance 1, For one with Random matrices of the same size that follow a normal distribution with mean 0 and variance 1, For one with Random matrices of the same size that follow a normal distribution with mean 0 and variance 1, Represents the multiplication of corresponding points of two matrices, Represents the addition of corresponding points of tensors; The generative network is reconstructed using the following formula t Laboratory test indicators at all times, t Personal attributes of the moment and T Clinical text of the moment: ； ； ； in, for t Laboratory test indicators that are constantly being reconstructed, for t Personal attributes that are constantly reconstructed, for T Clinical texts that are constantly being reconstructed, for t The latent variables related to sepsis after time reparameterization, for t Hidden variables related to other diseases after constant reparameterization, for t The latent variables that are not related to the disease after constant reparameterization, Represents a generative model consisting of multiple consecutive fully connected layers; The factor analysis module is connected to the model training module and is used to group the current electronic health record time series data to obtain multiple groups of current factor data, and determine the correlation strength between each group of current factor data and sepsis based on the trained serialized causal decoupling model to analyze the causal factors of sepsis.

2. The device for analyzing the causes of sepsis based on matrix causal disentanglement according to claim 1, characterized in that: The model training module includes: The exogenous variable determination submodule is used to extract the spatial distribution characteristics of laboratory test indicators and personal attributes at any time during the historical setting period to determine t exogenous variables at the time; according to t -2 time exogenous variables, t Exogenous variables at time -1 and t The exogenous variables at the moment are encoded using the serialized variational autoencoder to obtain t The time distribution information of the exogenous variables at the moment, where t For any time within the historical setting period, 1< t < T , T The final moments of a period set for history; The encoding submodule is used to adopt the serialized variational autoencoder to t -2 laboratory test indicators, t -1 moment laboratory test indicators and t The laboratory test indicators at the time are coded to obtain t The time distribution information of laboratory test indicators at the moment; the serialized variational automatic encoder is used to t -2 moments of personal attributes, t -1 moment of personal attributes and t Encode the personal attributes of the moment to obtain t Time distribution information of personal attributes at each moment; A priori and posterior submodule, used to adopt the prior network, according to t The time distribution information of the exogenous variables at the moment and t -1 time the posterior mean of each latent variable, calculate t The prior mean and prior variance of each latent variable at the moment, and define the prior distribution; using the posterior network, according to t Time distribution information of exogenous variables at each moment, t Time distribution information of laboratory test indicators at each moment, t Time distribution information of personal attributes at each moment and t -1 time, the posterior mean of each latent variable is determined t The posterior mean and posterior variance of the hidden variables at the moment, and define the posterior distribution; The association strength determination submodule is used to adopt the generation network according to t The posterior mean and posterior variance of each latent variable at the moment are determined t The latent variables after reparameterization at each moment, and according to t The latent variables after moment reparameterization are reconstructed t Laboratory test indicators at all times and t The personal attributes of the moment, according to T The latent variables related to sepsis and other diseases after the reparameterization of time are reconstructed. T Clinical text of the moment, according to T The latent variables related to sepsis after momentary reparameterization were used to analyze the strength of association between the latent variables related to sepsis and sepsis; The training submodule is used to reconstruct the laboratory test indicators at each moment within the historical set period, the personal attributes reconstructed at each moment, T Constantly reconstructed clinical texts, T Predicted disease labels at each moment, association strength at each moment, personal attributes at each moment, T Clinical texts at all times, T The diseased label at the moment, the prior distribution and the posterior distribution are optimized with the lower bound of evidence as the optimization target, so as to obtain a trained serialized causal decoupling model.

3. The device for analyzing the causes of sepsis based on matrix causal disentanglement according to claim 2, characterized in that: The exogenous variable determination submodule is based on t -2 time exogenous variables, t Exogenous variables at time -1 and t The exogenous variables at the moment are encoded using the serialized variational autoencoder to obtain t The time distribution information of exogenous variables at the moment, including: Using a point convolution based encoder, t -2 time exogenous variables, t Exogenous variables at time -1 and t The exogenous variables at the moment are encoded to obtain t Multiple preliminary feature maps at the moment; Will t The multiple preliminary feature maps at the moment are spliced together, and the encoder based on depthwise convolution is used to encode the spliced preliminary feature maps to obtain t Final feature map of the moment; according to t The final feature map at the moment is calculated using the formula Sure t The time distribution information of the exogenous variables at the moment; for t Time distribution information of exogenous variables at the moment for t The final feature map at time t, for t exogenous variables at time, represents a full connection operation, Represents the shape of the reshaped tensor, Represents the addition of corresponding points in the tensor.

4. The device for analyzing the causes of sepsis based on matrix causal disentanglement according to claim 2, characterized in that: The prior network is determined by the following formula t The prior mean and prior variance of each latent variable at time: ； ； ； in, for t The prior mean of the latent variable associated with sepsis at time t, for t The prior variance of the latent variable associated with sepsis at each moment, for t The prior mean of the latent variables related to other diseases at each moment, for t Prior variance of latent variables related to other diseases at each moment, for t The prior mean of the latent variable that is not related to the disease at any time, for t The prior variance of the latent variable that is not related to the disease at any given moment, for t The posterior mean of the latent variable associated with sepsis at time -1, for t The posterior mean of the latent variables related to other diseases at time -1, for t The posterior mean of the latent variable that is not related to the disease at time -1, for t Time distribution information of exogenous variables at the moment, represents a gated recurrent unit, 、 、 、 、 、 Represents different fully connected operations.

5. The device for analyzing the causes of sepsis based on matrix causal disentanglement according to claim 2, characterized in that: The posterior network is determined using the following formula t The posterior mean and posterior variance of each latent variable at time: ； ； ； in, for t The posterior mean of the latent variable associated with sepsis at time t, for t The posterior variance of the latent variable associated with sepsis at time t, for t The posterior mean of the latent variables related to other diseases at each moment, for t The posterior variance of the latent variables related to other diseases at each moment, for t The posterior mean of the latent variable that is not related to the disease at any time, for t The posterior variance of the latent variable that is not related to the disease at any given moment, for t The posterior mean of the latent variable associated with sepsis at time -1, for t The posterior mean of the latent variables related to other diseases at time -1, for t The posterior mean of the latent variable that is not related to the disease at time -1, for t Time distribution information of exogenous variables at the moment, for t Time distribution information of laboratory test indicators at each moment, for t Time distribution information of personal attributes at each moment, represents the long short-term memory network, 、 、 、 、 、 Represents different fully connected operations.

6. The device for analyzing the causes of sepsis based on matrix causal disentanglement according to claim 2, characterized in that: The prior distribution is: ； in, is the prior distribution, are the parameters of the prior network, for t The set of hidden parameters at time t, ={ , }, for t The hidden parameter set at time -1, for t exogenous variables at time, for t Hidden variables associated with sepsis at all times, for t Hidden variables related to sepsis at time -1, for t Hidden variables that are constantly associated with other diseases, for t -1 time is associated with other diseases, for t Hidden variables that are not related to the disease at any time, for t -1 is a latent variable that is not related to the disease; The posterior distribution is: ； in, is the posterior distribution, are the parameters of the posterior network, for t The first intermediate parameter set at time, , is the second intermediate parameter set, , for t Laboratory test indicators at all times, for t The personal attributes of the moment, for T Clinical text of the moment, for T The disease label at all times, for T Hidden variables associated with sepsis at all times, for T Hidden variables that are constantly associated with other diseases, Represents a continuous multiplication operation.

7. The device for analyzing the causes of sepsis based on matrix causal disentanglement according to claim 6, characterized in that: The optimization goal of the serialized causal decoupling model is: ； ； ； ； in, is the optimization target value of the serialized causal decoupling model, for t Time distribution information of exogenous variables at the moment, for t -1 time distribution information of exogenous variables, is the first intermediate variable, is the second intermediate variable, describes the divergence between the prior and posterior distributions, is the third intermediate variable, Express expectations.

Images